{
    "paper_id": "2020",
    "header": {
        "generated_with": "S2ORC 1.0.0",
        "date_generated": "2023-01-19T12:52:45.048883Z"
    },
    "title": "\u57fa \u57fa \u57fa\u4e8e \u4e8e \u4e8e\u591a \u591a \u591a\u4efb \u4efb \u4efb\u52a1 \u52a1 \u52a1\u5b66 \u5b66 \u5b66\u4e60 \u4e60 \u4e60\u7684 \u7684 \u7684\u751f \u751f \u751f\u6210 \u6210 \u6210\u5f0f \u5f0f \u5f0f\u9605 \u9605 \u9605\u8bfb \u8bfb \u8bfb\u7406 \u7406 \u7406\u89e3 \u89e3 \u89e3",
    "authors": [
        {
            "first": "\u94b1",
            "middle": [
                "\u94b1"
            ],
            "last": "\u94b1\u9526",
            "suffix": "",
            "affiliation": {},
            "email": ""
        },
        {
            "first": "\uff0c",
            "middle": [
                "\uff0c"
            ],
            "last": "\uff0c\u9ec4",
            "suffix": "",
            "affiliation": {},
            "email": ""
        },
        {
            "first": "\u9ec4",
            "middle": [],
            "last": "\u9ec4\u8363 \u8363 \u8363\u6d9b",
            "suffix": "",
            "affiliation": {},
            "email": ""
        },
        {
            "first": "\uff0c",
            "middle": [
                "\uff0c"
            ],
            "last": "\uff0c\u90b9",
            "suffix": "",
            "affiliation": {},
            "email": ""
        },
        {
            "first": "\u90b9",
            "middle": [],
            "last": "\u90b9\u535a \u535a \u535a\u4f1f",
            "suffix": "",
            "affiliation": {},
            "email": ""
        },
        {
            "first": "\uff0c",
            "middle": [
                "\uff0c"
            ],
            "last": "\uff0c\u6d2a \u6d2a \u6d2a\u5b87",
            "suffix": "",
            "affiliation": {},
            "email": ""
        },
        {
            "first": "Jin",
            "middle": [],
            "last": "Qian",
            "suffix": "",
            "affiliation": {},
            "email": ""
        },
        {
            "first": "Rongtao",
            "middle": [],
            "last": "Huang",
            "suffix": "",
            "affiliation": {},
            "email": ""
        },
        {
            "first": "Bowei",
            "middle": [],
            "last": "Zou",
            "suffix": "",
            "affiliation": {},
            "email": "zoubowei@i2r.a-star.edu.sg"
        },
        {
            "first": "Yu",
            "middle": [],
            "last": "Hong",
            "suffix": "",
            "affiliation": {},
            "email": ""
        }
    ],
    "year": "",
    "venue": null,
    "identifiers": {},
    "abstract": "Generative reading comprehension is a novel and challenging research in the field of machine reading comprehension. Compared with the mainstream extractive reading comprehension, generative reading comprehension model is no longer limited to extract answers from paragraphs, but can combine questions and paragraphs to generate natural and complete statements as answers. However, the existing generative reading comprehension model lacks the understanding of the boundary information of answers in paragraphs and the question type information. To solve such issues, this paper proposes a generative reading comprehension model based on multi-task learning. In the training phase, the model takes the answer generation task as the main task, and the answer extraction and question classification tasks as auxiliary tasks for multi-task learning. The model simultaneously learns and optimizes the parameters of the model encoding layer. Then it loads the encoding layer in the test phase to decode and generate the answers. The experimental results show that the answer extraction model and the question classification model can effectively improve the performance of the generative reading comprehension model.",
    "pdf_parse": {
        "paper_id": "2020",
        "_pdf_hash": "",
        "abstract": [
            {
                "text": "Generative reading comprehension is a novel and challenging research in the field of machine reading comprehension. Compared with the mainstream extractive reading comprehension, generative reading comprehension model is no longer limited to extract answers from paragraphs, but can combine questions and paragraphs to generate natural and complete statements as answers. However, the existing generative reading comprehension model lacks the understanding of the boundary information of answers in paragraphs and the question type information. To solve such issues, this paper proposes a generative reading comprehension model based on multi-task learning. In the training phase, the model takes the answer generation task as the main task, and the answer extraction and question classification tasks as auxiliary tasks for multi-task learning. The model simultaneously learns and optimizes the parameters of the model encoding layer. Then it loads the encoding layer in the test phase to decode and generate the answers. The experimental results show that the answer extraction model and the question classification model can effectively improve the performance of the generative reading comprehension model.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Abstract",
                "sec_num": null
            }
        ],
        "body_text": [
            {
                "text": "1 \u5f15 \u5f15 \u5f15\u8a00 \u8a00 \u8a00 \u673a \u5668 \u9605 \u8bfb \u7406 \u89e3 \u5728 \u9605 \u8bfb \u548c \u7406 \u89e3 \u81ea \u7136 \u8bed \u8a00 \u7684 \u57fa \u7840 \u4e0a \uff0c \u6839 \u636e \u6587 \u672c \u5185 \u5bb9 \u56de \u7b54 \u7528 \u6237 \u63d0 \u51fa \u7684 \u95ee \u9898 \uff0c \u662f \u5f53 \u524d \u81ea \u52a8 \u95ee \u7b54 \u9886 \u57df \u7684 \u7814 \u7a76 \u70ed \u70b9 \u4e4b \u4e00 \u3002 \u8fd1 \u5e74 \u6765 \uff0c \u968f \u7740 \u5927 \u89c4 \u6a21 \u9605 \u8bfb \u7406 \u89e3 \u6570 \u636e \u96c6 \u7684 \u6784 \u5efa\uff0c\u5982SQuAD",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            },
            {
                "text": "\u3001HotpotQA \u3001CoQA (Reddy et al., 2018) \uff0c \u4ee5 \u53ca \u9884 \u8bad \u7ec3 \u6a21 \u578b \u7684 \u63d0 \u51fa \uff0c \u5982BERT (Devlin et al., 2018) \u3001UniLM (Dong et al., 2019 )\u3001ENRIE-GEN (Xiao et al., 2020) \uff0c\u673a\u5668\u9605\u8bfb\u7406\u89e3\u6280\u672f\u53d6\u5f97\u4e86\u5de8\u5927\u53d1\u5c55\u3002\u76ee\u524d\u4e3b\u6d41\u7684 \u673a\u5668\u9605\u8bfb\u7406\u89e3\u6a21\u578b\u901a\u5e38\u5c06\u7b54\u6848\u8bbe\u5b9a\u4e3a\u6bb5\u843d\u4e2d\u7684\u4e00\u4e2a\u8fde\u7eed\u7247\u6bb5\uff0c\u8fd9\u79cd\u62bd\u53d6\u5f0f\u9605\u8bfb\u7406\u89e3\u6a21\u578b\u5b58\u5728\u4e00 \u5b9a\u7684\u5c40\u9650\u6027\uff0c\u5176\u4ec5\u80fd\u76f4\u63a5\u4ee5\u6bb5\u843d\u4e2d\u7684\u7247\u6bb5\u4f5c\u4e3a\u7b54\u6848\uff0c\u5bfc\u81f4\u5728\u9488\u5bf9\u67d0\u4e9b\u95ee\u9898\u65f6\uff0c\u65e0\u6cd5\u7ed9\u51fa\u81ea\u7136 \u6d41\u7545\u7684\u7b54\u6848\uff0c\u4f8b\u5982\u8868 1(a)\u4e2d\u7684True/False\u95ee\u9898\u3002\u6b64\u5916\uff0c\u5982\u679c\u5c06\u95ee\u9898\u4e0e\u7b54\u6848\u5206\u79bb\uff0c\u4ec5\u6839\u636e\u7b54\u6848\u65e0 \u6cd5\u83b7\u5f97\u5b8c\u6574\u6e05\u6670\u7684\u4fe1\u606f\u3002\u8868 1(b)\u4e2d\u4f8b\u5b50\u6240\u793a\uff0c\u4e25\u683c\u610f\u4e49\u4e0a\u8bf4\uff0c\u62bd\u53d6\u5f0f\u6a21\u578b\u7ed9\u51fa\u7684\u7b54\u6848\"Season 5(\u7b2c5\u5b63)\"\u5e76\u4e0d\u901a\u987a\uff0c\u5728\u67d0\u4e9b\u5e94\u7528\u573a\u666f(\u5982\u804a\u5929\u673a\u5668\u4eba)\u4e2d\uff0c\u4f1a\u5bf9\u7528\u6237\u4f53\u9a8c\u9020\u6210\u5f71\u54cd\u3002",
                "cite_spans": [
                    {
                        "start": 16,
                        "end": 36,
                        "text": "(Reddy et al., 2018)",
                        "ref_id": "BIBREF16"
                    },
                    {
                        "start": 67,
                        "end": 88,
                        "text": "(Devlin et al., 2018)",
                        "ref_id": "BIBREF4"
                    },
                    {
                        "start": 96,
                        "end": 114,
                        "text": "(Dong et al., 2019",
                        "ref_id": "BIBREF5"
                    },
                    {
                        "start": 127,
                        "end": 146,
                        "text": "(Xiao et al., 2020)",
                        "ref_id": "BIBREF25"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            },
            {
                "text": "\u8868 1: \u62bd\u53d6\u5f0f\u4e0e\u751f\u6210\u5f0f\u673a\u5668\u9605\u8bfb\u7406\u89e3 (a) \u6bb5 \u6bb5 \u6bb5\u843d \u843d \u843d",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            },
            {
                "text": "That all stops now, thanks to the creators of Sarcastic Font, which is italics, but in reverse. Simply genius! Here is their manifesto to support the need for a sarcasm font: For too long e-mails, instant messages, web pages and documents have been unable to fully communicate the subtleties of sarcasm. \u8fd9\u4e00\u5207\u73b0\u5728\u90fd\u505c\u6b62\u4e86\uff0c\u611f\u8c22\u8bbd\u523a\u5b57\u4f53\u7684\u521b\u9020\u8005\uff0c\u8fd9\u662f\u659c\u4f53\uff0c\u4f46\u53cd\u8fc7\u6765\u3002\u7b80\u76f4 \u5c31\u662f\u5929\u624d\uff01\u4ee5\u4e0b\u662f\u4ed6\u4eec\u7684\u5ba3\u8a00\uff0c\u4ee5\u652f\u6301\u8bbd\u523a\u5b57\u4f53\u7684\u5fc5\u8981\u6027\uff1a\u592a\u957f\u65f6\u95f4\u4ee5\u6765\uff0c \u7535\u5b50\u90ae\u4ef6\u3001\u5373\u65f6\u6d88\u606f\u3001\u7f51\u9875\u548c\u6587\u6863\u90fd\u65e0\u6cd5\u5145\u5206\u4f20\u8fbe\u8bbd\u523a\u7684\u5fae\u5999\u4e4b\u5904\u3002 \u95ee \u95ee \u95ee\u9898 \u9898 \u9898 is there a font for sarcasm? \u6709\u8bbd\u523a\u5b57\u4f53\u5417\uff1f \u62bd \u62bd \u62bd\u53d6 \u53d6 \u53d6\u5f0f \u5f0f \u5f0f\u7b54 \u7b54 \u7b54\u6848 \u6848 \u6848 Sarcastic Font \u8bbd\u523a\u5b57\u4f53 \u751f \u751f \u751f\u6210 \u6210 \u6210\u5f0f \u5f0f \u5f0f\u7b54 \u7b54 \u7b54\u6848 \u6848 \u6848 Yes, there is sarcastic font for sarcasm. \u662f\u7684\uff0c\u8bbd\u523a\u6709\u8bbd\u523a\u5b57\u4f53\u3002 (b) \u6bb5 \u6bb5 \u6bb5\u843d \u843d \u843d Longmire will ride again. Netflix announced on Friday that they had renewed the Western for a 10-episode Season 5, just seven weeks after the pluckedfrom-the-ashes fourth season debuted on the streaming service...",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            },
            {
                "text": "\u897f \u9547 \u8b66 \u9b42 \u5c06 \u518d \u6b21 \u4e0a \u6620 \u3002Netflix\u5468 \u4e94 \u5ba3 \u5e03 \uff0c \u4ed6 \u4eec \u5df2 \u7ecf \u7eed \u8ba2 \u4e86 \u8fd9 \u90e8 \u897f \u90e8 \u7247 \u7b2c5\u5b63\uff0c\u517110\u96c6\u3002\u800c\u5c31\u57287\u5468\u524d\uff0c\u8131\u80ce\u6362\u9aa8\u7b2c\u56db\u5b63\u624d\u5728Netflix\u4e0a\u9996\u64ad... \u95ee \u95ee \u95ee\u9898 \u9898 \u9898 how many seasons does longmire have? \u897f\u9547\u8b66\u9b42\u6709\u591a\u5c11\u5b63\uff1f \u62bd \u62bd \u62bd\u53d6 \u53d6 \u53d6\u5f0f \u5f0f \u5f0f\u7b54 \u7b54 \u7b54\u6848 \u6848 \u6848 Season 5 \u7b2c5\u5b63 \u751f \u751f \u751f\u6210 \u6210 \u6210\u5f0f \u5f0f \u5f0f\u7b54 \u7b54 \u7b54\u6848 \u6848 \u6848",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            },
            {
                "text": "Longmire has 5 seasons. \u897f\u9547\u8b66\u9b42\u67095\u5b63\u3002 (Vaswani et al., 2017) \u7f16\u7801\u5668\uff0c\u5e76\u501f\u9274UniLMv2 (Bao et al., 2020) \u6a21\u578b\u4e2d\u7279\u6b8a\u8bbe\u8ba1\u7684\u81ea\u6ce8\u610f\u529b\u63a9\u7801\u673a\u5236\u63a7\u5236\u7b54\u6848\u751f\u6210\u8fc7\u7a0b\u4e2d\u7684\u53ef\u89c1\u4fe1\u606f\uff1b\u4efb\u52a1\u5c42\u5206\u4e3a\u7b54\u6848\u751f \u6210\u6a21\u578b\u3001\u7b54\u6848\u62bd\u53d6\u6a21\u578b\u548c\u95ee\u9898\u5206\u7c7b\u6a21\u578b\uff0c\u7b54\u6848\u751f\u6210\u6a21\u578b\u5728\u8bad\u7ec3\u9636\u6bb5\u901a\u8fc7\u9884\u6d4b\u88ab\u906e\u853d\u7b54\u6848\u5355\u8bcd\u7684 \u539f\u59cb\u4fe1\u606f\uff0c\u589e\u5f3a\u6a21\u578b\u7684\u751f\u6210\u80fd\u529b\uff0c\u5728\u6d4b\u8bd5\u9636\u6bb5\u76f4\u63a5\u91c7\u7528\u8bad\u7ec3\u597d\u7684\u7f16\u7801\u5c42\uff0c\u4ee5\u53ca\u675f\u641c\u7d22(Beam search) (Sutskever et al., 2014) (Rajpurkar et al., 2016) \u3001TriviaQA (Joshi et al., 2017) \u3001SearchQA (Dunn et al., 2017) \u3001HotpotQA (Bajaj et al., 2018) \u3001NarrativeQA (Ko\u010disk\u00fd et al., 2018) \u548cCoQA (Reddy et al., 2018) ",
                "cite_spans": [
                    {
                        "start": 33,
                        "end": 55,
                        "text": "(Vaswani et al., 2017)",
                        "ref_id": "BIBREF22"
                    },
                    {
                        "start": 71,
                        "end": 89,
                        "text": "(Bao et al., 2020)",
                        "ref_id": "BIBREF1"
                    },
                    {
                        "start": 223,
                        "end": 247,
                        "text": "(Sutskever et al., 2014)",
                        "ref_id": "BIBREF17"
                    },
                    {
                        "start": 248,
                        "end": 272,
                        "text": "(Rajpurkar et al., 2016)",
                        "ref_id": "BIBREF15"
                    },
                    {
                        "start": 283,
                        "end": 303,
                        "text": "(Joshi et al., 2017)",
                        "ref_id": "BIBREF7"
                    },
                    {
                        "start": 314,
                        "end": 333,
                        "text": "(Dunn et al., 2017)",
                        "ref_id": "BIBREF6"
                    },
                    {
                        "start": 344,
                        "end": 364,
                        "text": "(Bajaj et al., 2018)",
                        "ref_id": "BIBREF0"
                    },
                    {
                        "start": 378,
                        "end": 400,
                        "text": "(Ko\u010disk\u00fd et al., 2018)",
                        "ref_id": null
                    },
                    {
                        "start": 407,
                        "end": 427,
                        "text": "(Reddy et al., 2018)",
                        "ref_id": "BIBREF16"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            },
            {
                "text": "\u4e0e\u62bd\u53d6\u5f0f\u9605\u8bfb\u7406\u89e3\u76f8\u6bd4\uff0c\u751f\u6210\u5f0f\u9605\u8bfb\u7406\u89e3\u4e0d\u518d\u5c40\u9650\u4e8e\u76f4\u63a5\u4ece\u6bb5\u843d\u7247\u6bb5\u4e2d\u62bd\u53d6\u7b54\u6848\uff0c\u800c\u662f\u53c2 \u8003\u6bb5\u843d\u3001\u95ee\u9898\u3001\u751a\u81f3\u8bcd\u8868\uff0c\u751f\u6210\u66f4\u4e3a\u81ea\u7136\u548c\u5b8c\u6574\u7684\u8868\u8ff0\u4f5c\u4e3a\u7b54\u6848\u3002\u4f8b\u5982\uff0c\u8868 1(a)\u4e2d\uff0c\u751f\u6210\u5f0f\u9605 \u8bfb\u7406\u89e3\u6a21\u578b\u7ed9\u51fa\u7684\u7b54\u6848\u80fd\u591f\u4e0e\u95ee\u9898\u66f4\u81ea\u7136\u5730\u8854\u63a5\uff1b\u800c\u8868 1(b)\u4e2d\u7684\u751f\u6210\u5f0f\u7b54\u6848\u4e0e\u62bd\u53d6\u5f0f\u7b54\u6848\u76f8 \u6bd4\u66f4\u5b8c\u6574\uff0c\u786e\u4fdd\u4e86\u7b54\u6848\u5728\u72ec\u7acb\u4e8e\u95ee\u9898\u548c\u6bb5\u843d\u65f6\u4ecd\u80fd\u591f\u4fdd\u6301\u5b8c\u6574\u7684\u4fe1\u606f\u3002\u800c\u73b0\u6709\u7684\u751f\u6210\u5f0f\u9605\u8bfb\u7406 \u89e3\u6a21\u578b\u901a\u5e38\u57fa\u4e8e\u6574\u4e2a\u6bb5\u843d\u751f\u6210\u7b54\u6848\uff0c\u7f3a\u4e4f\u5bf9\u7b54\u6848\u8fb9\u754c\u548c\u95ee\u9898\u7c7b\u578b\u4fe1\u606f\u7684\u7406\u89e3\uff0c\u751f\u6210\u7b54\u6848\u6709\u65f6\u672a \u53c2\u8003\u6bb5\u843d\u4e2d\u7528\u4e8e\u751f\u6210\u7b54\u6848\u7684\u7247\u6bb5\u4ee5\u53ca\u95ee\u9898\u7684\u5177\u4f53\u7c7b\u578b\uff0c\u5bfc\u81f4\u751f\u6210\u7684\u7b54\u6848\u548c\u771f\u5b9e\u7b54\u6848\u4e4b\u95f4\u5b58\u5728\u5dee \u5f02\u3002 \u4e3a\u89e3\u51b3\u4e0a\u8ff0\u95ee\u9898\uff0c\u672c\u6587\u63d0\u51fa\u4e00\u79cd\u57fa\u4e8e\u591a\u4efb\u52a1\u5b66\u4e60\u7684\u751f\u6210\u5f0f\u9605\u8bfb\u7406\u89e3\u6846\u67b6\u3002\u591a\u4efb\u52a1\u5b66\u4e60\u80fd\u591f \u5b66\u5230\u591a\u4e2a\u5173\u8054\u4efb\u52a1\u7684\u5171\u4eab\u8868\u793a\uff0c\u5e76\u9002\u5e94\u8fd9\u4e9b\u4e0d\u540c\u4f46\u76f8\u5173\u7684\u4efb\u52a1\u76ee\u6807\uff0c\u4f7f\u4e3b\u4efb\u52a1\u83b7\u5f97\u66f4\u5f3a\u7684\u6cdb\u5316 \u6027\u80fd\u3002\u57fa\u4e8e\u6b64\uff0c\u672c\u6587\u5c06\u7b54\u6848\u751f\u6210\u4efb\u52a1\u4f5c\u4e3a\u4e3b\u4efb\u52a1\uff0c\u5c06\u7b54\u6848\u62bd\u53d6\u548c\u95ee\u9898\u5206\u7c7b\u4efb\u52a1\u4f5c\u4e3a\u8f85\u52a9\u4efb\u52a1\uff0c \u5728\u8bad\u7ec3\u9636\u6bb5\uff0c\u901a\u8fc7\u591a\u4efb\u52a1\u5b66\u4e60\u7684\u53c2\u6570\u5171\u4eab\u673a\u5236\uff0c\u8ba9\u6a21\u578b\u751f\u6210\u7b54\u6848\u7684\u540c\u65f6\u52a0\u5f3a\u5bf9\u7b54\u6848\u8fb9\u754c\u548c\u95ee\u9898 \u7c7b\u578b\u7684\u7406\u89e3\uff0c\u4ece\u800c\u8ba9\u7b54\u6848\u62bd\u53d6\u548c\u95ee\u9898\u5206\u7c7b\u4efb\u52a1\u8f85\u52a9\u7b54\u6848\u751f\u6210\u4efb\u52a1\uff0c\u6700\u7ec8\u63d0\u5347\u751f\u6210\u5f0f\u9605\u8bfb\u7406\u89e3\u6a21 \u578b\u7684\u6cdb\u5316\u6027\u80fd\u3002 \u9488\u5bf9\u7b54\u6848\u751f\u6210\u4efb\u52a1\uff0c\u672c\u6587\u63d0\u51fa\u7684\u751f\u6210\u5f0f\u9605\u8bfb\u7406\u89e3\u6a21\u578b\u7531\u7f16\u7801\u5c42\u548c\u4efb\u52a1\u5c42\u7ec4\u6210\u3002\u5176\u4e2d\uff0c \u7f16\u7801\u5c42\u57fa\u4e8e\u6df1\u5ea6\u53cc\u5411Transformer",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            },
            {
                "text": "\u2022 \u63d0\u51fa\u57fa\u4e8e\u591a\u4efb\u52a1\u5b66\u4e60\u7684\u751f\u6210\u5f0f\u9605\u8bfb\u7406\u89e3\u6a21\u578b\uff0c\u901a\u8fc7\u7b54\u6848\u62bd\u53d6\u6a21\u578b\u548c\u95ee\u9898\u5206\u7c7b\u6a21\u578b\u4f18\u5316\u751f\u6210 \u5f0f\u9605\u8bfb\u7406\u89e3\u6a21\u578b\u7684\u6027\u80fd\uff1b \u2022 \u672c\u6587\u5728\u4e09\u4e2a\u9605\u8bfb\u7406\u89e3\u6570\u636e\u96c6\u4e0a\u8fdb\u884c\u8be6\u7ec6\u5b9e\u9a8c\uff0c\u5747\u53d6\u5f97\u4e86\u76ee\u524d\u751f\u6210\u5f0f\u6a21\u578b\u7684\u6700\u4f73\u6027\u80fd\u3002 2 \u76f8 \u76f8 \u76f8\u5173 \u5173 \u5173\u5de5 \u5de5 \u5de5\u4f5c \u4f5c \u4f5c 2.1 \u751f \u751f \u751f\u6210 \u6210 \u6210\u5f0f \u5f0f \u5f0f\u673a \u673a \u673a\u5668 \u5668 \u5668\u9605 \u9605 \u9605\u8bfb \u8bfb \u8bfb\u7406 \u7406 \u7406\u89e3 \u89e3 \u89e3 \u8fd1 \u5e74 \u6765 \uff0c \u968f \u7740 \u5982SQuAD",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            },
            {
                "text": "\u548cQuAC (choi et al., 2018)\u7b49 \u5927 \u89c4 \u6a21 \u9605 \u8bfb \u7406 \u89e3 \u6570 \u636e \u96c6 \u7684 \u6784 \u5efa \uff0c \u4ee5 \u53ca \u5728 \u4ee5 \u795e \u7ecf \u7f51 \u7edc \u4e3a \u4ee3 \u8868 \u7684 \u6df1 \u5ea6 \u5b66 \u4e60 \u6280 \u672f \u548c \u8ba1 \u7b97 \u8d44 \u6e90 \u7684 \u63a8 \u52a8 \u4e0b \uff0c \u673a \u5668 \u9605 \u8bfb \u7406 \u89e3 \u9886 \u57df \u83b7 \u5f97 \u4e86 \u5de8 \u5927 \u53d1 \u5c55 \u3002 \u76ee \u524d \uff0cMS MARCO",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            },
            {
                "text": "\u7b49 \u6570 \u636e \u96c6 \u63d0 \u4f9b \u4eba \u5de5 \u7f16 \u8f91 \u751f \u6210 \u7684 \u7b54 \u6848 \uff0c \u8981 \u6c42 \u673a \u5668 \u80fd \u591f \u7406 \u89e3 \u95ee \u9898 \u548c \u6bb5 \u843d \u4e2d \u76f8 \u5173 \u53e5 \u5b50 \u7684 \u6f5c \u5728 \u8054 \u7cfb \uff0c \u4f9d \u8d56 \u4e00 \u5b9a \u7684 \u63a8 \u7406 \u80fd \u529b \u751f \u6210 \u6b63 \u786e \u7684 \u7b54 \u6848 \uff0c \u800c \u975e \u7b80 \u5355 \u7684 \u6587 \u672c \u5339 \u914d \u3002 \u968f \u7740 \u751f \u6210 \u5f0f \u9605 \u8bfb \u7406 \u89e3 \u6570 \u636e \u96c6 \u7684 \u53d1 \u5e03 \u4ee5 \u53ca \u81ea \u7136 \u8bed \u8a00 \u751f \u6210 \u6280 \u672f \u7684 \u53d1 \u5c55 \uff0c \u7814 \u7a76 \u8005 \u5f00 \u59cb \u5173 \u6ce8 \u4f7f \u7528 \u751f \u6210 \u6a21 \u578b \u6765 \u89e3 \u51b3 \u9605 \u8bfb \u7406 \u89e3 \u95ee \u9898\u3002McCann\u7b49 (2018)\u548cBauer\u7b49 (2018)\u91c7\u7528\u57fa\u4e8eRNN\u7684\u6307\u9488\u751f\u6210\u673a\u5236\u8fdb\u884c\u5355\u6587\u6863\u9605\u8bfb\u7406\u89e3 \u7b54\u6848\u7684\u751f\u6210\uff0cTan\u7b49 (2018)\u5728\u591a\u6587\u6863\u9605\u8bfb\u7406\u89e3\u4e2d\u91c7\u7528\u7ba1\u9053(Pipeline)\u7684\u65b9\u6cd5\uff0c\u5148\u4ece\u591a\u7bc7\u6587\u6863\u4e2d \u62bd\u53d6\u51fa\u6700\u6709\u53ef\u80fd\u6210\u4e3a\u7b54\u6848\u7684\u7247\u6bb5\uff0c\u7136\u540e\u5c06\u8be5\u7247\u6bb5\u4f5c\u4e3a\u7b54\u6848\u5408\u6210\u6a21\u5757(Seq2Seq\u751f\u6210\u6a21\u578b)\u7684\u4e00\u4e2a \u7279\u5f81,\u6700\u540e\u7efc\u5408\u95ee\u9898\u3001\u6587\u6863\u548c\u62bd\u53d6\u7279\u5f81\u5408\u6210\u7b54\u6848\u3002\u800c\u672c\u6587\u6240\u63d0\u51fa\u7684\u662f\u7aef\u5230\u7aef\u7684\u751f\u6210\u5f0f\u9605\u8bfb\u7406\u89e3\u6a21 \u578b\uff0c\u65e8\u5728\u8ba9\u7b54\u6848\u751f\u6210\u3001\u7b54\u6848\u62bd\u53d6\u4ee5\u53ca\u95ee\u9898\u5206\u7c7b\u5171\u4eab\u6a21\u578b\u7f16\u7801\u5c42\u53c2\u6570\u5e76\u8fdb\u884c\u4f18\u5316\uff0c\u6700\u7ec8\u8fbe\u5230\u63d0\u5347 \u751f\u6210\u6a21\u578b\u6027\u80fd\u7684\u76ee\u7684\u3002",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            },
            {
                "text": "\u76ee\u524d\uff0c\u9884\u8bad\u7ec3\u6a21\u578b\u5982Mass (Song et al., 2019) \u3001UniLM (Dong et al., 2019) \u3001BART (Lewis et al., 2019) ",
                "cite_spans": [
                    {
                        "start": 14,
                        "end": 33,
                        "text": "(Song et al., 2019)",
                        "ref_id": "BIBREF19"
                    },
                    {
                        "start": 41,
                        "end": 60,
                        "text": "(Dong et al., 2019)",
                        "ref_id": "BIBREF5"
                    },
                    {
                        "start": 67,
                        "end": 87,
                        "text": "(Lewis et al., 2019)",
                        "ref_id": "BIBREF9"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            },
            {
                "text": "EQUATION",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [
                    {
                        "start": 0,
                        "end": 8,
                        "text": "EQUATION",
                        "ref_id": "EQREF",
                        "raw_str": "\u4ee5\u53caERNIE-GEN (Xiao et al., 2020)\u7b49\u5728\u5404\u4e2a\u81ea\u7136\u8bed\u8a00\u751f\u6210\u4efb\u52a1\u4e2d\u76f8\u7ee7\u53d6\u5f97\u6700\u4f73 \u6027\u80fd\uff0c\u8fd9\u4e9b\u6a21\u578b\u53ea\u9700\u5728\u7279\u5b9a\u4efb\u52a1(\u5982\u9605\u8bfb\u7406\u89e3\u3001\u6587\u672c\u6458\u8981\u4ee5\u53ca\u673a\u5668\u7ffb\u8bd1\u7b49)\u8fdb\u884c\u5fae\u8c03\u5c31\u80fd\u53d6\u5f97 \u4ee4\u4eba\u6ee1\u610f\u7684\u6210\u7ee9\u3002\u5176\u4e2d\uff0cBao\u7b49 (2020)\u63d0\u51faUniLMv2\u6a21\u578b\uff0c\u5176\u4f7f\u7528\u4e00\u79cd\u65b0\u9896\u7684\u4f2a\u906e\u853d\u8bed\u8a00\u6a21 \u578b(pseudo-masked language model, PMLM)\u5c06\u81ea\u7f16\u7801\u6a21\u578b\u548c\u90e8\u5206\u81ea\u56de\u5f52\u6a21\u578b\u7edf\u4e00\u8d77\u6765\u8bad\u7ec3\uff0c\u5728 \u95ee\u9898\u751f\u6210\u3001\u81ea\u52a8\u6458\u8981\u7b49\u591a\u4e2a\u9886\u57df\u53d6\u5f97\u5f53\u524d\u7684\u6700\u4f73\u6027\u80fd\u3002\u672c\u6587\u5c06UniLMv2\u6a21\u578b\u4f5c\u4e3a\u57fa\u7ebf\u6a21\u578b\uff0c\u5e76 \u5728\u6b64\u57fa\u7840\u4e0a\u8fdb\u884c\u591a\u4efb\u52a1\u5b66\u4e60\u7684\u5b9e\u9a8c\u3002 2.2 \u591a \u591a \u591a\u4efb \u4efb \u4efb\u52a1 \u52a1 \u52a1\u5b66 \u5b66 \u5b66\u4e60 \u4e60 \u4e60 \u591a\u4efb\u52a1\u5b66\u4e60\u662f\u4e00\u79cd\u63d0\u9ad8\u6cdb\u5316\u6027\u80fd\u7684\u8fc1\u79fb\u673a\u5236\uff0c\u73b0\u6709\u7814\u7a76\u8868\u660e\u5b83\u5728\u63d0\u9ad8\u6a21\u578b\u6cdb\u5316\u80fd\u529b\u4e0a\u5341 \u5206\u6709\u6548\u3002\u8be5\u673a\u5236\u540c\u65f6\u5b66\u4e60\u591a\u4e2a\u76f8\u5173\u4efb\u52a1\uff0c\u8ba9\u8fd9\u4e9b\u4efb\u52a1\u540c\u65f6\u5171\u4eab\u77e5\u8bc6\uff0c\u5229\u7528\u4efb\u52a1\u4e4b\u95f4\u7684\u76f8\u5173 \u6027\uff0c\u63d0\u5347\u6bcf\u4e2a\u4efb\u52a1\u7684\u6cdb\u5316\u6027\u80fd\u3002\u591a\u4efb\u52a1\u5b66\u4e60\u7684\u4e00\u822c\u505a\u6cd5\u662f\uff0c\u5728\u6240\u6709\u4efb\u52a1\u4e0a\u5171\u4eab\u6a21\u578b\u7f16\u7801\u5c42\uff0c \u800c\u9488\u5bf9\u7279\u5b9a\u7684\u4efb\u52a1\u5c42\u6709\u6240\u533a\u522b\u3002\u4f8b\u5982\uff0cWang\u7b49 (2018)\u8bc1\u660e\u901a\u8fc7\u5171\u4eab\u6587\u6863\u6392\u5e8f\u4efb\u52a1\u548c\u591a\u6587\u6863 \u9605\u8bfb\u7406\u89e3\u4efb\u52a1\u7684\u7f16\u7801\u5c42\u80fd\u591f\u63d0\u5347\u6574\u4f53\u7684\u6027\u80fd\u3002Nishida\u7b49 (2019)\u5728\u9605\u8bfb\u7406\u89e3\u3001\u6587\u6863\u6392\u5e8f\u548c\u95ee\u9898 \u5206\u7c7b\u4e09\u79cd\u4efb\u52a1\u4e0a\u5171\u4eab\u95ee\u9898\u548c\u6587\u7ae0\u9605\u8bfb\u6a21\u5757\uff0c\u6709\u6548\u63d0\u5347\u4e86\u6a21\u578b\u7684\u6574\u4f53\u6027\u80fd\u3002Liu\u7b49 (2019)\u63d0\u51fa \u7684MT-DNN\u6a21\u578b\u5728BERT (Devlin et al., 2018)\u7684\u57fa\u7840\u4e0a\u5bf94\u79cd\u4e0b\u6e38\u4efb\u52a1\u5355\u53e5\u5206\u7c7b\u3001\u6210\u5bf9\u6587\u672c\u5206 \u7c7b\u3001\u6587\u672c\u76f8\u4f3c\u5ea6\u6253\u5206\u548c\u76f8\u5173\u6027\u6392\u5e8f\u8fdb\u884c\u8054\u5408\u5fae\u8c03\uff0c\u5728\u6027\u80fd\u4e0a\u8f83BERT\u6709\u4e86\u6781\u5927\u63d0\u5347\uff0c\u8bc1\u660e\u4e86\u591a \u4efb\u52a1\u5b66\u4e60\u80fd\u6709\u6548\u63d0\u5347\u6a21\u578b\u7684\u6cdb\u5316\u6027\u80fd\u3002\u6b64\u5916\uff0c\u4e0eMT-DNN\u6a21\u578b\u5728\u4e0b\u6e38\u4efb\u52a1\u4e0a\u8fdb\u884c\u591a\u4efb\u52a1\u5b66\u4e60\u4e0d \u540c\uff0cERNIE2.0 (Sun et al., 2020)\u5728\u6a21\u578b\u9884\u8bad\u7ec3\u9636\u6bb5\u5f15\u5165\u591a\u4efb\u52a1\u5b66\u4e60\uff0c\u901a\u8fc7\u548c\u591a\u4e2a\u5148\u9a8c\u77e5\u8bc6\u5e93\u8fdb \u884c\u4ea4\u4e92\u5e76\u91c7\u7528\u589e\u91cf\u5b66\u4e60\u7684\u65b9\u5f0f\uff0c\u4f7f\u5f97\u6a21\u578b\u80fd\u591f\u5b66\u4f1a\u591a\u6837\u5316\u7684\u8bed\u8a00\u77e5\u8bc6\uff0c\u6700\u7ec8\u5728\u5404\u79cd\u4e0b\u6e38\u4efb\u52a1\u4e0a \u6027\u80fd\u5f97\u5230\u63d0\u5347\u3002 \u53d7\u5230\u4e0a\u8ff0\u5de5\u4f5c\u7684\u542f\u53d1\uff0c\u4e3a\u4e86\u89e3\u51b3\u73b0\u6709\u7684\u751f\u6210\u5f0f\u9605\u8bfb\u7406\u89e3\u6a21\u578b\u7f3a\u4e4f\u5bf9\u7b54\u6848\u8fb9\u754c\u4fe1\u606f\u548c\u95ee\u9898\u7c7b \u522b\u4fe1\u606f\u7684\u7406\u89e3\u7684\u95ee\u9898\uff0c\u672c\u6587\u63d0\u51fa\u57fa\u4e8e\u591a\u4efb\u52a1\u5b66\u4e60\u7684\u751f\u6210\u5f0f\u9605\u8bfb\u7406\u89e3\u6a21\u578b\uff0c\u901a\u8fc7\u7b54\u6848\u62bd\u53d6\u6a21\u578b\u548c \u95ee\u9898\u5206\u7c7b\u6a21\u578b\u4f18\u5316\u751f\u6210\u5f0f\u9605\u8bfb\u7406\u89e3\u6a21\u578b\u6027\u80fd\u3002 3 \u57fa \u57fa \u57fa\u4e8e \u4e8e \u4e8e\u591a \u591a \u591a\u4efb \u4efb \u4efb\u52a1 \u52a1 \u52a1\u5b66 \u5b66 \u5b66\u4e60 \u4e60 \u4e60\u7684 \u7684 \u7684\u751f \u751f \u751f\u6210 \u6210 \u6210\u5f0f \u5f0f \u5f0f\u9605 \u9605 \u9605\u8bfb \u8bfb \u8bfb\u7406 \u7406 \u7406\u89e3 \u89e3 \u89e3\u6a21 \u6a21 \u6a21\u578b \u578b \u578b \u672c\u7ae0\u9996\u5148\u7ed9\u51fa\u751f\u6210\u5f0f\u9605\u8bfb\u7406\u89e3\u95ee\u9898\u7684\u5f62\u5f0f\u5316\u5b9a\u4e49\uff1b\u7136\u540e\u4ecb\u7ecd\u6a21\u578b\u7684\u7f16\u7801\u5c42\uff1b\u6700\u540e\u4ecb\u7ecd\u6a21\u578b \u7684\u4efb\u52a1\u5c42\uff0c\u5176\u5177\u4f53\u7531\u7b54\u6848\u751f\u6210\u6a21\u578b\u3001\u7b54\u6848\u62bd\u53d6\u6a21\u578b\u548c\u95ee\u9898\u5206\u7c7b\u6a21\u578b\u4e09\u90e8\u5206\u7ec4\u6210\u3002\u57fa\u4e8e\u591a\u4efb\u52a1\u5b66 \u4e60\u7684\u751f\u6210\u5f0f\u9605\u8bfb\u7406\u89e3\u6a21\u578b\u6846\u67b6\u5982\u56fe 1\u6240\u793a\u3002 ... ... ... ... ... ... ... ... \u56fe 1: \u57fa\u4e8e\u591a\u4efb\u52a1\u5b66\u4e60\u7684\u751f\u6210\u5f0f\u9605\u8bfb\u7406\u89e3\u6a21\u578b\u6846\u67b6 3.1 \u95ee \u95ee \u95ee\u9898 \u9898 \u9898\u5b9a \u5b9a \u5b9a\u4e49 \u4e49 \u4e49 \u7ed9\u5b9a\u95ee\u9898\u548c\u6bb5\u843d\u5206\u522b\u8868\u793a\u4e3aQ = {q i } m i=1 \u548cP = {p i } n i=1 \uff0c\u7b54\u6848\u8868\u793a\u4e3aA = {a i } t i=1 \uff0c\u5176 \u4e2dm\u3001n\u3001t\u5206\u522b\u8868\u793a\u95ee\u9898\u3001\u6bb5\u843d\u548c\u7b54\u6848\u7684\u957f\u5ea6\u3002\u5728\u751f\u6210\u5f0f\u9605\u8bfb\u7406\u89e3\u4e2d\uff0c\u672c\u6587\u5c06\u95ee\u9898\u548c\u6bb5\u843d\u7ec4\u6210 \u6e90\u5e8f\u5217\uff0c\u7b54\u6848\u4f5c\u4e3a\u76ee\u6807\u5e8f\u5217\uff0c\u76ee\u6807\u662f\u6839\u636eQ\u548cP \uff0c\u81ea\u52a8\u751f\u6210\u7b26\u5408\u8bed\u4e49\u7684\u76ee\u6807\u7b54\u6848A\u3002\u8be5\u4efb\u52a1\u7684\u76ee \u6807\u0101\u8868\u793a\u4e3a\u0101 = arg max a P (a|Q, P ) (1) \u5176\u4e2dP (a|Q, P )\u8868\u793a\u5728\u7ed9\u5b9aQ\u548cP \u7684\u6761\u4ef6\u4e0b\uff0c\u751f\u6210\u7b54\u6848\u7684\u5bf9\u6570\u6761\u4ef6\u6982\u7387\u3002 3.2 \u7f16 \u7f16 \u7f16\u7801 \u7801 \u7801\u5c42 \u5c42 \u5c42 \u672c\u6587\u57fa\u4e8e\u9884\u8bad\u7ec3\u6a21\u578bUniLMv2 1 (Bao et al., 2020)\u6784\u5efa\u7f16\u7801\u5c42\uff0c\u91c7\u7528\u9884\u8bad\u7ec3\u7684BERT\u8fdb\u884c\u95ee \u9898\u548c\u6bb5\u843d\u7684\u4ea4\u4e92\uff0c\u5f97\u5230\u5176\u8868\u793a\uff0c\u5e76\u5728BERT\u7684\u57fa\u7840\u4e0a\u6539\u8fdb\u4e86\u6ce8\u610f\u529b\u906e\u853d\u77e9\u9635\uff0c\u91c7\u7528\u4f2a\u906e\u853d\u8bed\u8a00 \u6a21\u578b\uff0c\u4f7f\u5f97\u6a21\u578b\u80fd\u5728\u9605\u8bfb\u7406\u89e3\u4efb\u52a1\u4e0a\u6839\u636e\u95ee\u9898\u548c\u6bb5\u843d\u9010\u5b57\u6216\u9010\u7247\u6bb5\u9884\u6d4b\u88ab\u906e\u853d\u7684\u7b54\u6848\u3002\u4ee5\u4e0b\u4ecb \u7ecd\u7f16\u7801\u5c42\u7684\u5177\u4f53\u5de5\u4f5c\u539f\u7406\u548c\u8fc7\u7a0b\u3002 \u9884 \u5904 \u7406 \u9636 \u6bb5 \uff0c \u91c7 \u7528WordPiece\u5206 \u8bcd \u5de5 \u5177 \uff0c \u5c06 \u95ee \u9898 \u3001 \u6bb5 \u843d \u548c \u7b54 \u6848 \u5206 \u8bcd \uff0c \u5f97 \u5230 \u5b50 \u8bcd (sub- word)\u7ea7\u522b\u7684\u82e5\u5e72\u8bcd\u9879\uff0c\u5176\u4e2d\u5bf9\u7b54\u6848\u4e2d\u7684\u90e8\u5206\u8bcd\u9879\u4ee5\u4e00\u5b9a\u6982\u7387\u8fdb\u884c\u906e\u853d\uff0c\u5e76\u5c06\u5176\u62fc\u63a5\u540e\u5c06 \u4f5c\u4e3a\u6a21\u578b\u8f93\u5165\u3002\u6bcf\u4e2a\u8bcd\u9879\u8868\u793a\u4e3a\u8bcd\u5411\u91cfWE(w i )\u3001\u6bb5\u5411\u91cfSE(w i )\u548c\u4f4d\u7f6e\u5411\u91cfPE(w i )\u7684\u548c\uff0c\u7ef4\u5ea6 \u5747\u4e3ad w \uff0c\u5176\u4e2d\u8bcd\u5411\u91cf\u7528\u4e8e\u8868\u793a\u4e0d\u540c\u8bcd\u9879\uff0c\u6bb5\u5411\u91cf\u7528\u4e8e\u533a\u5206\u8bcd\u6765\u81ea\u6e90\u5e8f\u5217\u8fd8\u662f\u76ee\u6807\u5e8f\u5217\uff0c\u4f4d\u7f6e\u5411 \u91cf\u7528\u4e8e\u8868\u793a\u8bcd\u5728\u8f93\u5165\u5e8f\u5217\u4e2d\u7684\u7edd\u5bf9\u4f4d\u7f6e\u3002\u8bcd\u5411\u91cfX i \u8868\u793a\u4e3a\uff1a X i = WE(w i ) + SE(w i ) + PE(w i ) \u5176\u4e2dw i \u4e3a\u7b2ci\u4e2a\u4f4d\u7f6e\u7684\u8bcd\u9879\u3002 \u672c\u6587\u5c06\u8bcd\u5411\u91cf\u96c6\u5408\u8868\u793a\u4e3a{X i } |x| i=1 \uff0c\u5219\u8f93\u5165\u5e8f\u5217\u8868\u793a\u4e3aH 0 = [X 1 , ..., X |x| ]\uff0c\u5176\u4e2d|x|\u4e3a\u8f93\u5165 \u5e8f\u5217\u7684\u957f\u5ea6\u3002UniLMv2\u7684\u7f16\u7801\u5c42\u4f7f\u752812\u5c42\u5806\u53e0\u7684Transformer\u7f51\u7edc\uff0c\u6bcf\u7ecf\u8fc7\u4e00\u5c42Transformer\u7f51 \u7edc\u90fd\u80fd\u5f97\u5230\u4e0d\u540c\u62bd\u8c61\u5c42\u6b21\u7684\u4e0a\u4e0b\u6587\u8868\u793a\uff1a H l = [h l 1 , ..., h l |x| ] = T ransf ormer l (H l\u22121 ), l \u2208 [1, 12]",
                        "eq_num": "(2)"
                    }
                ],
                "section": "",
                "sec_num": null
            },
            {
                "text": "EQUATION",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [
                    {
                        "start": 0,
                        "end": 8,
                        "text": "EQUATION",
                        "ref_id": "EQREF",
                        "raw_str": "\u5176\u4e2dl\u4e3a\u7b2cl\u5c42Transformer\u7f51\u7edc\uff0ch l i \u4e3a\u7b2ci\u4e2a\u8bcd\u9879\u7684l\u5c42\u9690\u5c42\u8868\u793a\u3002 Tranformer\u7f51\u7edc\u7531\u591a\u5934\u81ea\u6ce8\u610f\u529b\u673a\u5236\u548c\u524d\u5411\u795e\u7ecf\u7f51\u7edc\u4e24\u4e2a\u5b50\u5c42\u7ec4\u6210\uff0c\u6bcf\u4e2a\u5b50\u5c42\u5747\u4f7f\u7528\u6b8b\u5dee \u8fde\u63a5\u548c\u5c42\u6b63\u5219\u5316\uff0c\u56e0\u6b64\u6bcf\u4e2a\u5b50\u5c42\u7684\u8f93\u51fa\u53ef\u8868\u793a\u4e3a\uff1a LayerN orm(x + Sublayer(x)) \u7b2cl\u5c42Transformer\u7f51\u7edc\u7684\u81ea\u6ce8\u610f\u529b\u5934A l \u8ba1\u7b97\u5982\u4e0b\uff1a A l = sof tmax( Q l K T l \u221a d k + M)V l (3) Q l = H l\u22121 W Q l , K l = H l\u22121 W K l , V l = H l\u22121 W V l (4) M i,j = 0, \u5141\u8bb8\u88ab\u6ce8\u610f \u2212\u221e, \u4e0d\u5141\u8bb8\u88ab\u6ce8\u610f (5) \u5176 \u4e2dQ l \uff0cK l \u548cV l \u5206 \u522b \u4ee3 \u8868 \u7b2cl\u5c42 \u6ce8 \u610f \u529b \u673a \u5236 \u4e2d \u7684 \u67e5 \u8be2 (query) \u5411 \u91cf \u3001 \u952e (key) \u5411 \u91cf \u548c \u503c (value)\u5411\u91cf\uff0cd k \u4e3a\u5411\u91cfK l \u7684\u7ef4\u5ea6\uff0cW Q l , W K l , W V l \u2208 R d h \u00d7d k \u4e3a\u53ef\u5b66\u4e60\u53c2\u6570\u77e9\u9635\uff0cH l\u22121 \u2208 R |x|\u00d7d h \u4e3al \u2212 1\u5c42\u7684\u9690\u5c42\u8868\u793a\uff0cM\u4e3a\u751f\u6210\u5f0f\u9605\u8bfb\u7406\u89e3\u6a21\u578b\u6ce8\u610f\u529b\u906e\u853d\u77e9\u9635\uff0c\u5982\u56fe2\u6240\u793a\u3002 E E \u0549 E \u0549 E : \u00c7 \u202b\u0789\u202c \u0409 : \u00c7 \u202b\u0789\u202c \u0409 \u56fe 2: \u6ce8\u610f\u529b\u906e\u853d\u77e9\u9635 \u901a\u8fc7\u4e0a\u8ff0\u8bcd\u5d4c\u5165\u5c42\u548cTranformer\u7f51\u7edc\uff0c\u5f97\u5230\u8f93\u5165\u5e8f\u5217\u7684\u4e0a\u4e0b\u6587\u8868\u793aH 1 , H 2 , ..., H 12 \u3002\u672c\u6587\u4f7f \u7528\u6700\u540e\u4e00\u5c42\u8f93\u51faH 12 \u4f5c\u4e3a\u6574\u4e2a\u5e8f\u5217\u7684\u8868\u793a\u3002H 12 \u4e2d\u5305\u542b\u95ee\u9898\u3001\u6bb5\u843d\u548c\u7b54\u6848\u8868\u793a\uff0c\u5176\u4e2d\uff0c\u6bb5\u843d\u8868\u793a \u90e8\u5206\u8bb0\u4f5cH p \uff0c\u7b54\u6848\u8868\u793a\u90e8\u5206\u8bb0\u4f5cH a \uff0c\u95ee\u9898\u7c7b\u522b\u8868\u793a\u8bb0\u4f5cH cls \u3002\u6839\u636e\u56fe 2\u6240\u793a\u7684\u6ce8\u610f\u529b\u906e\u853d\u77e9\u9635 \u53ef\u77e5\uff0c\u95ee\u9898\u548c\u6bb5\u843d\u4e0d\u4f1a\u548c\u7b54\u6848\u8fdb\u884c\u4ea4\u4e92\uff0c\u4fdd\u8bc1\u4e86\u8bad\u7ec3\u548c\u6d4b\u8bd5\u9636\u6bb5H p \u548cH cls \u6240\u542b\u4fe1\u606f\u7684\u4e00\u81f4\u6027\u3002 3.3 \u4efb \u4efb \u4efb\u52a1 \u52a1 \u52a1\u5c42 \u5c42 \u5c42 \u4f5c\u4e3a\u57fa\u4e8e\u591a\u4efb\u52a1\u5b66\u4e60\u6846\u67b6\u7684\u6838\u5fc3\u90e8\u5206\uff0c\u4efb\u52a1\u5c42\u7531\u7b54\u6848\u751f\u6210\u6a21\u578b\u3001\u7b54\u6848\u62bd\u53d6\u6a21\u578b\u548c\u95ee\u9898\u5206\u7c7b \u6a21\u578b\u4e09\u90e8\u5206\u6784\u6210\u3002 \u7b54 \u7b54 \u7b54\u6848 \u6848 \u6848\u751f \u751f \u751f\u6210 \u6210 \u6210\u6a21 \u6a21 \u6a21\u578b \u578b \u578b \u8bad\u7ec3\u9636\u6bb5\uff0c\u771f\u5b9e\u7b54\u6848\u4f1a\u4ee5\u4e00\u5b9a\u6982\u7387\u88ab\u968f\u673a\u906e\u853d\uff0c\u5e76\u4e14\u540c\u65f6\u4fdd\u7559\u5176\u539f\u59cb\u4f4d\u7f6e\u4fe1\u606f \u6765\u5b9e\u73b0\u90e8\u5206\u81ea\u56de\u5f52(\u968f\u673a\u9884\u6d4b\u7b54\u6848\u88ab\u906e\u853d\u7684\u7247\u6bb5)\uff0c\u7b54\u6848\u4e2d\u88ab\u906e\u853d\u7684\u8bcd\u9879\u5728\u7ecf\u8fc7\u7f16\u7801\u540e\u5f97\u5230\u7b54 \u6848\u8868\u793aH a \u3002\u7b54\u6848\u751f\u6210\u6a21\u5757\u901a\u8fc7\u89e3\u7801\u5c42\u5bf9\u539f\u59cb\u7b54\u6848\u4e2d\u88ab\u906e\u853d\u7684\u8bcd\u9879\u8fdb\u884c\u9884\u6d4b\u6765\u751f\u6210\u7b54\u6848\u3002\u5177\u4f53\u6765 \u8bf4\uff0cH a \u9996\u5148\u7ecf\u8fc7\u7ebf\u6027\u5c42\u5e76\u7528Gelu\u51fd\u6570\u6fc0\u6d3b\u540e\u8fdb\u884c\u5c42\u6b63\u5219\u5316\uff1a H a = LayerN orm(Gelu(Linear(H a ))) (6) \u7136 \u540e \u901a \u8fc7 \u7ebf \u6027 \u5c42 \u5c06 \u6bcf \u4e2a \u88ab \u906e \u853d \u7684 \u8bcd \u9879 \u6620 \u5c04 \u5230 \u6a21 \u578b \u8bcd \u8868 \u7a7a \u95f4 \uff0c \u83b7 \u5f97 \u9884 \u6d4b \u5206 \u6570 \u3002 \u6700 \u540e \uff0c \u4f7f \u7528SoftMax\u51fd\u6570\u8ba1\u7b97\u8bcd\u7684\u6982\u7387\u5411\u91cfa\uff1a a = Sof tM ax(Linear(H a ))",
                        "eq_num": "(7)"
                    }
                ],
                "section": "",
                "sec_num": null
            },
            {
                "text": "\u672c\u6587\u91c7\u7528\u6709\u6807\u7b7e\u5e73\u6ed1\u4f18\u5316\u7684\u4ea4\u53c9\u71b5\u635f\u5931\u51fd\u6570\u8ba1\u7b97\u7b54\u6848\u751f\u6210\u6a21\u578b\u7684\u76ee\u6807\u51fd\u6570\uff1a",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            },
            {
                "text": "EQUATION",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [
                    {
                        "start": 0,
                        "end": 8,
                        "text": "EQUATION",
                        "ref_id": "EQREF",
                        "raw_str": "L generate = T t=1 K k=1 y a k t \u2022 log a k t",
                        "eq_num": "(8)"
                    }
                ],
                "section": "",
                "sec_num": null
            },
            {
                "text": "\u5176\u4e2dT \u8868\u793a\u7b54\u6848\u7684\u957f\u5ea6\uff0cK\u8868\u793a\u8bcd\u8868\u7684\u5927\u5c0f\uff0cy a k t \u8868\u793a\u7b54\u6848\u4e2d\u7b2ct\u4e2a\u4f4d\u7f6e\u7ecf\u8fc7\u6807\u7b7e\u4f18\u5316\u7684\u771f\u5b9e\u6807 \u7b7e\uff0ca k t \u8868\u793a\u7b54\u6848\u4e2d\u7b2ct\u4e2a\u4f4d\u7f6e\u7684\u9884\u6d4b\u6807\u7b7e\u3002\u6ce8\u610f\uff0c\u672c\u6587\u53ea\u5bf9\u7b54\u6848\u4e2d\u88ab\u906e\u853d\u7684\u8bcd\u9879\u8ba1\u7b97\u635f\u5931\u3002 \u6d4b\u8bd5\u9636\u6bb5\uff0c\u6a21\u578b\u5bf9\u4e8e\u8f93\u5165\u7684\u95ee\u9898\u548c\u6bb5\u843d\uff0c\u6bcf\u4e2a\u65f6\u95f4\u6b65\u7ecf\u89e3\u7801\u5c42\u9884\u6d4b\u5f53\u524d\u8bcd\u7684\u751f\u6210\u6982\u7387\uff0c\u540c \u65f6\u4f7f\u7528\u675f\u641c\u7d22\u6bcf\u6b21\u4fdd\u7559\u751f\u6210\u6982\u7387\u6700\u5927\u7684\u524dk\u4e2a\u5e8f\u5217\uff0c\u76f4\u81f3\u6a21\u578b\u9884\u6d4b\u51fa[EOS]\u7ec8\u6b62\u7b26\u7ed3\u675f\u89e3\u7801\u3002\u6700 \u540e\uff0c\u6a21\u578b\u5c06\u675f\u641c\u7d22\u7ed3\u679c\u4e2d\u751f\u6210\u6982\u7387\u6700\u5927\u7684\u5e8f\u5217\u89e3\u7801\u8f93\u51fa\uff0c\u5176\u6982\u7387\u8ba1\u7b97\u4e3a\uff1a P (A|Q, P ) = P (a 1 |Q, P )P (a 2 |Q, P, a 1 )...P ([EOS]|Q, P, a 1 , a 2 , ...) ",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            },
            {
                "text": "\u5176\u4e2d\uff0cc\u4ee3\u8868\u95ee\u9898\u7c7b\u578b\u7684\u5206\u6570\u5411\u91cf\u3002 \u672c\u6587\u91c7\u7528\u4ea4\u53c9\u71b5\u635f\u5931\u51fd\u6570\u8ba1\u7b97\u95ee\u9898\u5206\u7c7b\u6a21\u578b\u7684\u76ee\u6807\u51fd\u6570\uff1a\uff1a jpurkar et al., 2016) \u3001HotpotQA \u672c \u6587 \u5728CoQA\u6570 \u636e \u96c6 \u4e0a \u4f7f \u7528F1\u503c (Rajpurkar et al., 2016) \u6765 \u8bc4 \u4ef7 \u6a21 \u578b \u7684 \u6027 \u80fd \uff0c \u5728MS MARCO\u548cNarrativeQA\u6570\u636e\u96c6\u4e0a\u4f7f\u7528BLEU (Papineni et al., 2002) \u548cROUGE-L (Lin, 2004) ",
                "cite_spans": [
                    {
                        "start": 44,
                        "end": 65,
                        "text": "jpurkar et al., 2016)",
                        "ref_id": null
                    },
                    {
                        "start": 100,
                        "end": 124,
                        "text": "(Rajpurkar et al., 2016)",
                        "ref_id": "BIBREF15"
                    },
                    {
                        "start": 175,
                        "end": 198,
                        "text": "(Papineni et al., 2002)",
                        "ref_id": "BIBREF14"
                    },
                    {
                        "start": 208,
                        "end": 219,
                        "text": "(Lin, 2004)",
                        "ref_id": "BIBREF10"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            },
            {
                "text": "EQUATION",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [
                    {
                        "start": 0,
                        "end": 8,
                        "text": "EQUATION",
                        "ref_id": "EQREF",
                        "raw_str": "L cls = K k=1 y c k \u2022 log c k",
                        "eq_num": "(13)"
                    }
                ],
                "section": "",
                "sec_num": null
            },
            {
                "text": "\u5176\u4e2dK = 4\u8868\u793a\u95ee\u9898\u7c7b\u522b\u6570\uff0cy c k \u8868\u793a\u771f\u5b9e\u7c7b\u522b\u6807\u7b7e\uff0cc k \u8868\u793a\u9884\u6d4b\u7c7b\u522b\u6807\u7b7e\u3002 \u591a \u591a \u591a\u4efb \u4efb \u4efb\u52a1 \u52a1 \u52a1\u5b66 \u5b66 \u5b66\u4e60 \u4e60 \u4e60 \u672c\u6587\u91c7\u7528\u591a\u4efb\u52a1\u5b66\u4e60\u7684\u65b9\u6cd5\uff0c\u5728\u8bad\u7ec3\u9636\u6bb5\u540c\u65f6\u5b66\u4e60\u548c\u66f4\u65b0\u7b54\u6848\u751f\u6210\u3001\u7b54\u6848\u62bd\u53d6\u548c\u95ee \u9898\u5206\u7c7b\u6a21\u5757\u5171\u4eab\u7684\u7f16\u7801\u5c42\u53c2\u6570\uff0c\u8ba9\u7b54\u6848\u62bd\u53d6\u548c\u95ee\u9898\u5206\u7c7b\u4efb\u52a1\u8f85\u52a9\u7b54\u6848\u751f\u6210\u4efb\u52a1\u63d0\u5347\u9605\u8bfb\u7406\u89e3\u6a21 \u578b\u7684\u6027\u80fd\u3002\u6a21\u578b\u7684\u635f\u5931\u7531\u751f\u6210\u635f\u5931\u3001\u62bd\u53d6\u635f\u5931\u548c\u5206\u7c7b\u635f\u5931\u4e09\u90e8\u5206\u5171\u540c\u7ec4\u6210\uff0c\u6574\u4e2a\u6a21\u578b\u7684\u76ee\u6807\u51fd \u6570\u4e3a\uff1a LOSS = L generate + \u03bb 1 L extract + \u03bb 2 L cls (14) \u5176\u4e2d\u03bb 1 \u548c\u03bb 2 \u4e3a\u8c03\u548c\u7cfb\u6570\uff0c\u7528\u4e8e\u8c03\u8282\u8f85\u52a9\u4efb\u52a1\u6743\u91cd\u3002 4 \u5b9e \u5b9e \u5b9e\u9a8c \u9a8c \u9a8c \u672c\u7ae0\u9996\u5148\u4ecb\u7ecd\u751f\u6210\u5f0f\u9605\u8bfb\u7406\u89e3\u4efb\u52a1\u6570\u636e\u96c6\u548c\u5b9e\u9a8c\u8bbe\u7f6e\uff0c\u7136\u540e\u62a5\u544a\u672c\u6587\u63d0\u51fa\u7684\u57fa\u4e8e\u591a\u4efb\u52a1\u7684 \u751f\u6210\u5f0f\u9605\u8bfb\u7406\u89e3\u6a21\u578b\u6027\u80fd\uff0c\u5e76\u9488\u5bf9\u5b9e\u9a8c\u7ed3\u679c\u8fdb\u884c\u5206\u6790\u3002 4.1 \u751f \u751f \u751f\u6210 \u6210 \u6210\u5f0f \u5f0f \u5f0f\u9605 \u9605 \u9605\u8bfb \u8bfb \u8bfb\u7406 \u7406 \u7406\u89e3 \u89e3 \u89e3\u4efb \u4efb \u4efb\u52a1 \u52a1 \u52a1\u6570 \u6570 \u6570\u636e \u636e \u636e\u96c6 \u96c6 \u96c6 \u73b0\u6709\u9605\u8bfb\u7406\u89e3\u6570\u636e\u96c6\u5927\u591a\u9488\u5bf9\u62bd\u53d6\u5f0f\u6a21\u578b\uff0c\u5373\u7b54\u6848\u4e3a\u7bc7\u7ae0\u4e2d\u7684\u4e00\u4e2a\u7247\u6bb5\uff0c\u5982SQuAD (Ra",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            },
            {
                "text": "\u8868 2\u5217 \u51fa \u4e86 \u672c \u6587 \u6240 \u91c7 \u7528 \u4e09 \u4e2a \u6570 \u636e \u96c6 \u7684 \u7edf \u8ba1 \u6570 \u636e \u3002CoQA\u4e2d \u5b58 \u572828.7%\u7684 \u547d \u540d \u5b9e \u4f53 \u7c7b \u95ee \u9898\u300119.6%\u7684\u540d\u8bcd\u77ed\u8bed\u7c7b\u95ee\u9898\u548c9.8%\u7684\u6570\u5b57\u7c7b\u95ee\u9898\uff1bNarrativeQA\u4e2d\u5b58\u572830.54%\u7684\u4eba\u540d\u7c7b\u95ee \u9898\u30019.73%\u7684\u5730\u70b9\u7c7b\u95ee\u9898\u548c\u7ea610%\u5de6\u53f3\u7684\u4e8b\u4ef6\u3001\u5b9e\u4f53\u3001\u6570\u5b57\u7c7b\u95ee\u9898\uff0c\u4e14CoQA\u548cNarrativeQA\u660e \u786e\u5141\u8bb8\u7b80\u77ed\u3001\u81ea\u7136\u7684\u7b54\u6848\uff0c\u56e0\u6b64CoQA\u548cNarrativeQA\u7684\u7b54\u6848\u666e\u904d\u8f83\u77ed\u3002MS MARCO(NLG)\u4e2d \u5b58\u572853.12%\u7684\u63cf\u8ff0\u578b\u95ee\u9898\uff0c\u4e14\u7b54\u6848\u4f1a\u878d\u5165\u95ee\u9898\u4fe1\u606f\u5f62\u6210\u5b8c\u6574\u7684\u8868\u8ff0\uff0c\u7b54\u6848\u666e\u904d\u8f83\u957f\u3002 4.2 \u5b9e \u5b9e \u5b9e\u9a8c \u9a8c \u9a8c\u8bbe \u8bbe \u8bbe\u7f6e \u7f6e \u7f6e \u672c\u6587\u4f7f\u7528\u7684\u6a21\u578b\u4e3a\u5fae\u8f6f\u5f00\u6e90\u7684unilm1.2-",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            },
            {
                "text": "\u6765 \u8bc4\u4ef7\u6a21\u578b\u7684\u6027\u80fd\u3002 4.3 \u5b9e \u5b9e \u5b9e\u9a8c \u9a8c \u9a8c\u7ed3 \u7ed3 \u7ed3\u679c \u679c \u679c\u4e0e \u4e0e \u4e0e\u5206 \u5206 \u5206\u6790 \u6790 \u6790 \u4e3a\u4e86\u9a8c\u8bc1\u672c\u6587\u57fa\u4e8e\u591a\u4efb\u52a1\u7684\u751f\u6210\u5f0f\u9605\u8bfb\u7406\u89e3\u65b9\u6cd5\u7684\u6709\u6548\u6027\uff0c\u672c\u6587\u4e0e\u4ee5\u4e0b\u9605\u8bfb\u7406\u89e3\u6a21\u578b\u8fdb\u884c \u4e86\u6bd4\u8f83\uff1a \u2022 UniLM\uff1a \uff1a \uff1a\u7531Dong\u7b49 (2019)\u63d0\u51fa\uff0c\u662f\u7b2c\u4e00\u4e2a\u5728CoQA\u6570\u636e\u96c6\u4e0a\u62a5\u544a\u5b9e\u9a8c\u6027\u80fd\u7684\u9884\u8bad\u7ec3\u751f\u6210 \u6a21\u578b\uff0c\u672c\u6587\u5728\u5b9e\u9a8c\u8bbe\u7f6e\u4e0a\u548c\u5b83\u4fdd\u6301\u4e00\u81f4\u3002 \u2022 ERNIE-GEN\uff1a \uff1a \uff1a\u7531Xiao\u7b49 (2020)\u63d0\u51fa\u7684\u57fa\u4e8e\u591a\u6d41(multi-flow)\u673a\u5236\u751f\u6210\u5b8c\u6574\u8bed\u4e49\u7247\u6bb5\u7684 \u9884\u8bad\u7ec3\u751f\u6210\u6a21\u578b\uff0c\u5728CoQA\u751f\u6210\u5f0f\u9605\u8bfb\u7406\u89e3\u4e2d\u8fbe\u5230\u4e86\u76ee\u524d\u6700\u597d\u7684\u6027\u80fd\u3002 \u2022 Masque\uff1a \uff1a \uff1a \u7531Nishida\u7b49 (2019)\u63d0 \u51fa \u7684 \u591a \u98ce \u683c \u751f \u6210 \u5f0f \u9605 \u8bfb \u7406 \u89e3 \u6a21 \u578b \uff0c \u5728MS MARCO(NLG)\u548cNarrativeQA\u6570\u636e\u96c6\u7684\u76f8\u5173\u6307\u6807\u4e0a\u8fbe\u5230\u4e86\u76ee\u524d\u7684\u6700\u597d\u6027\u80fd\u3002 \u2022 UniLMv2\uff1a \uff1a",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            },
            {
                "text": "\u8868 4\u4e3a\u672c\u6587\u63d0\u51fa\u7684\u6a21\u578b\u5728CoQA\u9a8c\u8bc1\u96c6\u4e0a\u7684\u6027\u80fd\uff0c\u6211\u4eec\u7684\u6a21\u578b\u5728F1\u6307\u6807\u4e0a\u6bd4\u5f53\u524d\u6027\u80fd \u6700 \u597d \u7684 \u751f \u6210 \u5f0f \u6a21 \u578bENRIE-GEN\u63d0 \u5347 \u4e862.2%\uff0c \u540c \u65f6 \u8f83 \u57fa \u7ebf \u6a21 \u578bUniLMv2\u63d0 \u5347 \u4e860.6%\u3002 \u672c \u6587 \u9488 \u5bf9 \u9884 \u8bad \u7ec3 \u751f \u6210 \u6a21 \u578b \u5728 \u7b54 \u6848 \u89e3 \u7801 \u65f6 \u51fa \u73b0 \u7684 \u5b50 \u8bcd \u7ed3 \u5408 \u4e0d \u51c6 \u786e \u95ee \u9898 \u52a0 \u4ee5 \u4fee \u590d \uff0c \u5b9e \u73b0 \u7684 \u57fa \u7ebf \u6a21\u578bUniLMv2\u9ad8\u4e8e\u539f\u59cb\u7248\u672c\u7684\u6027\u80fd\uff0c\u8f83ENRIE-GEN\u63d0\u53471.6%\u7684F1\u503c\u3002\u8868 5\u5217\u51fa\u4e86\u672c\u6587\u6a21\u578b \u5728CoQA\u4e0a\u7684\u6d88\u878d\u5b9e\u9a8c\u6027\u80fd\uff0c\u5728\u53bb\u9664\u7b54\u6848\u62bd\u53d6\u4efb\u52a1\u548c\u95ee\u9898\u5206\u7c7b\u4efb\u52a1\u4e4b\u540e\uff0c\u6027\u80fd\u8f83MTL-Model\u5206 \u522b\u4e0b\u964d0.5%\u548c0.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            },
            {
                "text": "\u8868 6\u4e3a\u672c\u6587\u63d0\u51fa\u7684\u6a21\u578b\u5728MS MARCO (NLG)\u9a8c\u8bc1\u96c6\u4e0a\u9009\u53d6\u6700\u4f73\u6587\u6863\u7684\u6027\u80fd\u8868\u73b0\u3002\u672c\u6587\u6a21 \u578b\u8f83\u57fa\u7ebf\u6a21\u578bUniLMv2\u5728BLEU-1\u6307\u6807\u4e0a\u63d0\u53470.77%\uff0cBLEU-4\u6307\u6807\u4e0a\u63d0\u53470.95%\uff0cROUGE-L\u6307 \u6807\u4e0a\u63d0\u53470.55%\u3002\u8fd9\u662f\u7531\u4e8eMS MARCO(NLG)\u6570\u636e\u96c6\u4e2d\u7b54\u6848\u548c\u9009\u5b9a\u6bb5\u843d\u4e2d\u7684\u90e8\u5206\u7247\u6bb5\u76f8\u4f3c\u5ea6\u8f83 \u9ad8\uff0c\u7b54\u6848\u62bd\u53d6\u4efb\u52a1\u80fd\u591f\u8f85\u52a9\u6a21\u578b\u5173\u6ce8\u7b54\u6848\u5728\u6bb5\u843d\u4e2d\u7684\u8fb9\u754c\u4fe1\u606f\uff0c\u5e76\u589e\u5f3a\u751f\u6210\u6a21\u578b\u5bf9\u95ee\u9898\u548c\u6bb5\u843d \u4e2d\u7b54\u6848\u7247\u6bb5\u4e4b\u95f4\u5173\u7cfb\u7684\u7406\u89e3\uff0c\u6700\u7ec8\u63d0\u5347\u751f\u6210\u6a21\u578b\u7684\u6027\u80fd\u3002\u6211\u4eec\u5728\u540c\u6837\u8bbe\u7f6e\u4e0b\u548cMasque\u6a21\u578b\u8fdb\u884c \u4e86\u5bf9\u6bd4\uff0c\u672c\u6587\u6240\u63d0\u6a21\u578b\u5728BLEU-1\u6307\u6807\u4e0a\u63d0\u5347\u4e862.39%\uff0cROUGE-L\u6307\u6807\u4e0a\u63d0\u5347\u4e861.84%\u3002\u8fd9\u4e3b\u8981 \u7531\u4e8eMasque\u6a21\u578b\u4ec5\u4f7f\u7528\u9759\u6001\u7684\u9884\u8bad\u7ec3\u8bcd\u5411\u91cf\u5e76\u57fa\u4e8eTransformer\u7f51\u7edc\u8fdb\u884c\u7b54\u6848\u751f\u6210\uff0c\u800c\u672c\u6587\u6a21 \u578b\u57fa\u4e8e\u7f51\u7edc\u66f4\u52a0\u590d\u6742\u7684\u9884\u8bad\u7ec3\u6a21\u578bUniLMv2\u751f\u6210\u7b54\u6848\uff0c\u56e0\u6b64\u5728\u5b9e\u9a8c\u6027\u80fd\u4e0a\u53d6\u5f97\u8f83\u5927\u63d0\u5347\u3002 \u8868 7\u4e3a \u672c \u6587 \u6a21 \u578b \u5728NarrativeQA(summary)\u6d4b \u8bd5 \u96c6 \u4e0a \u7684 \u6027 \u80fd \u8868 \u73b0 \u3002 \u672c \u6587 \u6a21 \u578b \u8f83 \u57fa \u7ebf \u6a21 \u578bUniLMv2\u5728BLEU-1\u6307 \u6807 \u4e0a \u63d0 \u53470.39%\uff0cBLEU-4\u6307 \u6807 \u4e0a \u63d0 \u53470.61%\uff0cROUGE-L\u6307 \u6807 \u4e0a \u63d0 \u53470.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            },
            {
                "text": "https://github.com/microsoft/unilm",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            },
            {
                "text": "https://stanfordnlp.github.io/coqa/ 3 https://microsoft.github.io/msmarco/ 4 https://github.com/deepmind/narrativeqa 5 https://unilm.blob.core.windows.net/ckpt/unilm1.2-base-uncased.bin",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            }
        ],
        "back_matter": [],
        "bib_entries": {
            "BIBREF0": {
                "ref_id": "b0",
                "title": "MS MARCO: a human-generated machine reading comprehension dataset",
                "authors": [
                    {
                        "first": "Payal",
                        "middle": [],
                        "last": "Bajaj",
                        "suffix": ""
                    },
                    {
                        "first": "Daniel",
                        "middle": [],
                        "last": "Campos",
                        "suffix": ""
                    },
                    {
                        "first": "Nick",
                        "middle": [],
                        "last": "Craswell",
                        "suffix": ""
                    },
                    {
                        "first": "Li",
                        "middle": [],
                        "last": "Deng",
                        "suffix": ""
                    },
                    {
                        "first": "Jianfeng",
                        "middle": [],
                        "last": "Gao",
                        "suffix": ""
                    },
                    {
                        "first": "Xiaodong",
                        "middle": [],
                        "last": "Liu",
                        "suffix": ""
                    },
                    {
                        "first": "Rangan",
                        "middle": [],
                        "last": "Majumder",
                        "suffix": ""
                    },
                    {
                        "first": "Andrew",
                        "middle": [],
                        "last": "Mcnamara",
                        "suffix": ""
                    },
                    {
                        "first": "Bhaskar",
                        "middle": [],
                        "last": "Mitra",
                        "suffix": ""
                    },
                    {
                        "first": "Tri",
                        "middle": [],
                        "last": "Nguyen",
                        "suffix": ""
                    },
                    {
                        "first": "Mir",
                        "middle": [],
                        "last": "Rosenberg",
                        "suffix": ""
                    },
                    {
                        "first": "Xia",
                        "middle": [],
                        "last": "Song",
                        "suffix": ""
                    },
                    {
                        "first": "Alina",
                        "middle": [],
                        "last": "Stoica",
                        "suffix": ""
                    },
                    {
                        "first": "Saurabh",
                        "middle": [],
                        "last": "Tiwary",
                        "suffix": ""
                    },
                    {
                        "first": "Tong",
                        "middle": [],
                        "last": "Wang",
                        "suffix": ""
                    }
                ],
                "year": 2018,
                "venue": "Computing Research Repository (CoRR)",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {
                    "arXiv": [
                        "arXiv:1611.09268"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Payal Bajaj, Daniel Campos, Nick Craswell, Li Deng, Jianfeng Gao, Xiaodong Liu, Rangan Majumder, Andrew McNamara, Bhaskar Mitra, Tri Nguyen, Mir Rosenberg, Xia Song, Alina Stoica, Saurabh Tiwary, and Tong Wang. 2018. MS MARCO: a human-generated machine reading comprehension dataset. Computing Research Repository (CoRR), arXiv:1611.09268. Version 3.",
                "links": null
            },
            "BIBREF1": {
                "ref_id": "b1",
                "title": "Unilmv2: Pseudo-masked language models for unified language model pre-training",
                "authors": [
                    {
                        "first": "Hangbo",
                        "middle": [],
                        "last": "Bao",
                        "suffix": ""
                    },
                    {
                        "first": "Li",
                        "middle": [],
                        "last": "Dong",
                        "suffix": ""
                    },
                    {
                        "first": "Furu",
                        "middle": [],
                        "last": "Wei",
                        "suffix": ""
                    },
                    {
                        "first": "Wenhui",
                        "middle": [],
                        "last": "Wang",
                        "suffix": ""
                    },
                    {
                        "first": "Xiaodong",
                        "middle": [],
                        "last": "Nanyang",
                        "suffix": ""
                    },
                    {
                        "first": "Yu",
                        "middle": [],
                        "last": "Liu",
                        "suffix": ""
                    },
                    {
                        "first": "Songhao",
                        "middle": [],
                        "last": "Wang",
                        "suffix": ""
                    },
                    {
                        "first": "Jian-Feng",
                        "middle": [],
                        "last": "Piao",
                        "suffix": ""
                    },
                    {
                        "first": "Ming",
                        "middle": [],
                        "last": "Gao",
                        "suffix": ""
                    },
                    {
                        "first": "Hsiao-Wuen",
                        "middle": [],
                        "last": "Zhou",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Hon",
                        "suffix": ""
                    }
                ],
                "year": 2020,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {
                    "arXiv": [
                        "arXiv:2002.12804"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Hangbo Bao, Li Dong, Furu Wei, Wenhui Wang, NanYang, Xiaodong Liu, Yu Wang, Songhao Piao, Jian-feng Gao, Ming Zhou, and Hsiao-Wuen Hon. 2020. Unilmv2: Pseudo-masked language models for unified language model pre-training. arXiv preprint arXiv:2002.12804.",
                "links": null
            },
            "BIBREF2": {
                "ref_id": "b2",
                "title": "Commonsense for generative multi-hop question answering tasks",
                "authors": [
                    {
                        "first": "Lisa",
                        "middle": [],
                        "last": "Bauer",
                        "suffix": ""
                    },
                    {
                        "first": "Yicheng",
                        "middle": [],
                        "last": "Wang",
                        "suffix": ""
                    },
                    {
                        "first": "Mohit",
                        "middle": [],
                        "last": "Bansal",
                        "suffix": ""
                    }
                ],
                "year": 2018,
                "venue": "Empirical Methods in NaturalLanguage Processing (EMNLP)",
                "volume": "",
                "issue": "",
                "pages": "4220--4230",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Lisa Bauer, Yicheng Wang, and Mohit Bansal. 2018. Commonsense for generative multi-hop question answering tasks. In Empirical Methods in NaturalLanguage Processing (EMNLP), pages 4220-4230.",
                "links": null
            },
            "BIBREF3": {
                "ref_id": "b3",
                "title": "QuAC: Question answering in context",
                "authors": [
                    {
                        "first": "Eunsol",
                        "middle": [],
                        "last": "Choi",
                        "suffix": ""
                    },
                    {
                        "first": "He",
                        "middle": [],
                        "last": "He",
                        "suffix": ""
                    },
                    {
                        "first": "Mohit",
                        "middle": [],
                        "last": "Iyyer",
                        "suffix": ""
                    },
                    {
                        "first": "Mark",
                        "middle": [],
                        "last": "Yatskar",
                        "suffix": ""
                    },
                    {
                        "first": "Wentau",
                        "middle": [],
                        "last": "Yih",
                        "suffix": ""
                    },
                    {
                        "first": "Yejin",
                        "middle": [],
                        "last": "Choi",
                        "suffix": ""
                    },
                    {
                        "first": "Percy",
                        "middle": [],
                        "last": "Liang",
                        "suffix": ""
                    },
                    {
                        "first": "Luke",
                        "middle": [],
                        "last": "Zettlemoyer",
                        "suffix": ""
                    }
                ],
                "year": 2018,
                "venue": "Proceedings of the EMNLP",
                "volume": "",
                "issue": "",
                "pages": "2174--2184",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Eunsol Choi, He He, Mohit Iyyer, Mark Yatskar, Wentau Yih, Yejin Choi, Percy Liang, and Luke Zettlemoyer. 2018. QuAC: Question answering in context. In Proceedings of the EMNLP, pages 2174-2184, Brussels, Belgium.",
                "links": null
            },
            "BIBREF4": {
                "ref_id": "b4",
                "title": "Bert: Pre-training of deep bidirectional transformers for language understanding",
                "authors": [
                    {
                        "first": "Jacob",
                        "middle": [],
                        "last": "Devlin",
                        "suffix": ""
                    },
                    {
                        "first": "Mingwei",
                        "middle": [],
                        "last": "Chang",
                        "suffix": ""
                    },
                    {
                        "first": "Kenton",
                        "middle": [],
                        "last": "Lee",
                        "suffix": ""
                    },
                    {
                        "first": "Kristina",
                        "middle": [],
                        "last": "Toutanova",
                        "suffix": ""
                    }
                ],
                "year": 2018,
                "venue": "Computing Research Repository (CoRR)",
                "volume": "1",
                "issue": "",
                "pages": "",
                "other_ids": {
                    "arXiv": [
                        "arXiv:1810.04805"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Jacob Devlin, MingWei Chang, Kenton Lee, and Kristina Toutanova. 2018. Bert: Pre-training of deep bidirectional transformers for language understanding. Computing Research Repository (CoRR), arXiv:1810.04805. Version 1.",
                "links": null
            },
            "BIBREF5": {
                "ref_id": "b5",
                "title": "Unified language model pre-training for natural language understanding and generation",
                "authors": [
                    {
                        "first": "Li",
                        "middle": [],
                        "last": "Dong",
                        "suffix": ""
                    },
                    {
                        "first": "Nan",
                        "middle": [],
                        "last": "Yang",
                        "suffix": ""
                    },
                    {
                        "first": "Wenhui",
                        "middle": [],
                        "last": "Wang",
                        "suffix": ""
                    },
                    {
                        "first": "Furu",
                        "middle": [],
                        "last": "Wei",
                        "suffix": ""
                    },
                    {
                        "first": "Xiaodong",
                        "middle": [],
                        "last": "Liu",
                        "suffix": ""
                    },
                    {
                        "first": "Yu",
                        "middle": [],
                        "last": "Wang",
                        "suffix": ""
                    },
                    {
                        "first": "Jianfeng",
                        "middle": [],
                        "last": "Gao",
                        "suffix": ""
                    },
                    {
                        "first": "Ming",
                        "middle": [],
                        "last": "Zhou",
                        "suffix": ""
                    },
                    {
                        "first": "Hsiao-Wuen",
                        "middle": [],
                        "last": "Hon",
                        "suffix": ""
                    }
                ],
                "year": 2019,
                "venue": "33rd Conference on Neural Information Processing Systems",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Li Dong, Nan Yang, Wenhui Wang, Furu Wei, Xiaodong Liu, Yu Wang, Jianfeng Gao, Ming Zhou, and Hsiao-Wuen Hon. 2019. Unified language model pre-training for natural language understanding and generation. In 33rd Conference on Neural Information Processing Systems (NeurIPS 2019).",
                "links": null
            },
            "BIBREF6": {
                "ref_id": "b6",
                "title": "SearchQA: A new q&a dataset augmented with context from a search engine",
                "authors": [
                    {
                        "first": "Matthew",
                        "middle": [],
                        "last": "Dunn",
                        "suffix": ""
                    },
                    {
                        "first": "Levent",
                        "middle": [],
                        "last": "Sagun",
                        "suffix": ""
                    },
                    {
                        "first": "Mike",
                        "middle": [],
                        "last": "Higgins",
                        "suffix": ""
                    },
                    {
                        "first": "Ugur",
                        "middle": [],
                        "last": "Guney",
                        "suffix": ""
                    },
                    {
                        "first": "Volkan",
                        "middle": [],
                        "last": "Cirik",
                        "suffix": ""
                    },
                    {
                        "first": "Kyunghyun",
                        "middle": [],
                        "last": "Cho",
                        "suffix": ""
                    }
                ],
                "year": 2017,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {
                    "arXiv": [
                        "arXiv:1704.05179"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Matthew Dunn, Levent Sagun, Mike Higgins, Ugur Guney, Volkan Cirik, and Kyunghyun Cho. 2017. SearchQA: A new q&a dataset augmented with context from a search engine. arXiv preprint arXiv:1704.05179.",
                "links": null
            },
            "BIBREF7": {
                "ref_id": "b7",
                "title": "TriviaQA: A large scale distantly supervised challenge dataset for reading comprehension",
                "authors": [
                    {
                        "first": "Mandar",
                        "middle": [],
                        "last": "Joshi",
                        "suffix": ""
                    },
                    {
                        "first": "Eunsol",
                        "middle": [],
                        "last": "Choi",
                        "suffix": ""
                    },
                    {
                        "first": "Daniel",
                        "middle": [
                            "S"
                        ],
                        "last": "Weld",
                        "suffix": ""
                    },
                    {
                        "first": "Luke",
                        "middle": [],
                        "last": "Zettlemoyer",
                        "suffix": ""
                    }
                ],
                "year": 2017,
                "venue": "Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Mandar Joshi, Eunsol Choi, Daniel S. Weld, and Luke Zettlemoyer. 2017. TriviaQA: A large scale distantly supervised challenge dataset for reading comprehension. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics.",
                "links": null
            },
            "BIBREF8": {
                "ref_id": "b8",
                "title": "The narrativeqa reading comprehension challenge",
                "authors": [
                    {
                        "first": "Tom\u00e1\u0161",
                        "middle": [],
                        "last": "Ko\u010disk\u1ef3",
                        "suffix": ""
                    },
                    {
                        "first": "Jonathan",
                        "middle": [],
                        "last": "Schwarz",
                        "suffix": ""
                    },
                    {
                        "first": "Phil",
                        "middle": [],
                        "last": "Blunsom",
                        "suffix": ""
                    },
                    {
                        "first": "Chris",
                        "middle": [],
                        "last": "Dyer",
                        "suffix": ""
                    },
                    {
                        "first": "Karl",
                        "middle": [
                            "Moritz"
                        ],
                        "last": "Hermann",
                        "suffix": ""
                    },
                    {
                        "first": "G\u00e1bor",
                        "middle": [],
                        "last": "Melis",
                        "suffix": ""
                    },
                    {
                        "first": "Edward",
                        "middle": [],
                        "last": "Grefenstette",
                        "suffix": ""
                    }
                ],
                "year": 2018,
                "venue": "Transactions of the Association for Computational Linguistic (TACL)",
                "volume": "6",
                "issue": "",
                "pages": "317--328",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Tom\u00e1\u0161 Ko\u010disk\u1ef3, Jonathan Schwarz, Phil Blunsom, Chris Dyer, Karl Moritz Hermann, and G\u00e1bor Melis, and Edward Grefenstette. 2018. The narrativeqa reading comprehension challenge. Transactions of the Association for Computational Linguistic (TACL), 6:317-328",
                "links": null
            },
            "BIBREF9": {
                "ref_id": "b9",
                "title": "Bart: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension",
                "authors": [
                    {
                        "first": "Mike",
                        "middle": [],
                        "last": "Lewis",
                        "suffix": ""
                    },
                    {
                        "first": "Yinhan",
                        "middle": [],
                        "last": "Liu",
                        "suffix": ""
                    },
                    {
                        "first": "Naman",
                        "middle": [],
                        "last": "Goyal",
                        "suffix": ""
                    },
                    {
                        "first": "Marjan",
                        "middle": [],
                        "last": "Ghazvininejad",
                        "suffix": ""
                    },
                    {
                        "first": "Abdelrahman",
                        "middle": [],
                        "last": "Mohamed",
                        "suffix": ""
                    },
                    {
                        "first": "Omer",
                        "middle": [],
                        "last": "Levy",
                        "suffix": ""
                    },
                    {
                        "first": "Ves",
                        "middle": [],
                        "last": "Stoyanov",
                        "suffix": ""
                    },
                    {
                        "first": "Luke",
                        "middle": [],
                        "last": "Zettlemoyer",
                        "suffix": ""
                    }
                ],
                "year": 2019,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {
                    "arXiv": [
                        "arXiv:1910.13461"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Ves Stoyanov, and Luke Zettlemoyer. 2019. Bart: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension. arXiv preprint arXiv:1910.13461.",
                "links": null
            },
            "BIBREF10": {
                "ref_id": "b10",
                "title": "ROUGE: A package for automatic evaluation of summaries",
                "authors": [
                    {
                        "first": "Chin-Yew",
                        "middle": [],
                        "last": "Lin",
                        "suffix": ""
                    }
                ],
                "year": 2004,
                "venue": "Proc. ACL Workshop on Text Summarization Branches Out",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Chin-Yew Lin. 2004. ROUGE: A package for automatic evaluation of summaries. In Proc. ACL Workshop on Text Summarization Branches Out.",
                "links": null
            },
            "BIBREF11": {
                "ref_id": "b11",
                "title": "Multi-task deep neural networks for natural language understanding",
                "authors": [
                    {
                        "first": "Xiaodong",
                        "middle": [],
                        "last": "Liu",
                        "suffix": ""
                    },
                    {
                        "first": "Pengcheng",
                        "middle": [],
                        "last": "He",
                        "suffix": ""
                    },
                    {
                        "first": "Weizhu",
                        "middle": [],
                        "last": "Chen",
                        "suffix": ""
                    },
                    {
                        "first": "Jianfeng",
                        "middle": [],
                        "last": "Gao",
                        "suffix": ""
                    }
                ],
                "year": 2019,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {
                    "arXiv": [
                        "arXiv:1901.11504"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Xiaodong Liu, Pengcheng He, Weizhu Chen, and Jianfeng Gao. 2019. Multi-task deep neural networks for natural language understanding. arXiv preprint arXiv:1901.11504.",
                "links": null
            },
            "BIBREF12": {
                "ref_id": "b12",
                "title": "The natural language decathlon: Multitask learning as question answering",
                "authors": [
                    {
                        "first": "Bryan",
                        "middle": [],
                        "last": "Mccann",
                        "suffix": ""
                    },
                    {
                        "first": "Nitish",
                        "middle": [],
                        "last": "Shirish Keskar",
                        "suffix": ""
                    },
                    {
                        "first": "Caiming",
                        "middle": [],
                        "last": "Xiong",
                        "suffix": ""
                    },
                    {
                        "first": "Richard",
                        "middle": [],
                        "last": "Socher",
                        "suffix": ""
                    }
                ],
                "year": 2018,
                "venue": "Computing Research Repository (CoRR)",
                "volume": "1",
                "issue": "",
                "pages": "",
                "other_ids": {
                    "arXiv": [
                        "arXiv:1806.08730"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Bryan McCann, Nitish Shirish Keskar, Caiming Xiong, and Richard Socher. 2018. The natural language decathlon: Multitask learning as question answering. Computing Research Repository (CoRR), arXiv:1806.08730. Version 1.",
                "links": null
            },
            "BIBREF13": {
                "ref_id": "b13",
                "title": "Multi-style generative reading comprehension",
                "authors": [
                    {
                        "first": "Kyosuke",
                        "middle": [],
                        "last": "Nishida",
                        "suffix": ""
                    },
                    {
                        "first": "Itsumi",
                        "middle": [],
                        "last": "Saito",
                        "suffix": ""
                    },
                    {
                        "first": "Kosuke",
                        "middle": [],
                        "last": "Nishida",
                        "suffix": ""
                    },
                    {
                        "first": "Kazutoshi",
                        "middle": [],
                        "last": "Shinoda",
                        "suffix": ""
                    },
                    {
                        "first": "Atsushi",
                        "middle": [],
                        "last": "Otsuka",
                        "suffix": ""
                    },
                    {
                        "first": "Hisako",
                        "middle": [],
                        "last": "Asano",
                        "suffix": ""
                    },
                    {
                        "first": "Junji",
                        "middle": [],
                        "last": "Tomita",
                        "suffix": ""
                    }
                ],
                "year": 2019,
                "venue": "ACL",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Kyosuke Nishida, Itsumi Saito, Kosuke Nishida, Kazutoshi Shinoda, Atsushi Otsuka, Hisako Asano, and Junji Tomita. 2019. Multi-style generative reading comprehension. In ACL.",
                "links": null
            },
            "BIBREF14": {
                "ref_id": "b14",
                "title": "BLEU: a method for automatic evaluation of machine translation",
                "authors": [
                    {
                        "first": "Kishore",
                        "middle": [],
                        "last": "Papineni",
                        "suffix": ""
                    },
                    {
                        "first": "Salim",
                        "middle": [],
                        "last": "Roukos",
                        "suffix": ""
                    },
                    {
                        "first": "Todd",
                        "middle": [],
                        "last": "Ward",
                        "suffix": ""
                    },
                    {
                        "first": "Weijing",
                        "middle": [],
                        "last": "Zhu",
                        "suffix": ""
                    }
                ],
                "year": 2002,
                "venue": "Association for Computational Linguistics (ACL)",
                "volume": "",
                "issue": "",
                "pages": "311--318",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Kishore Papineni, Salim Roukos, Todd Ward, and WeiJing Zhu. 2002. BLEU: a method for auto- matic evaluation of machine translation. In Association for Computational Linguistics (ACL), pages 311-318.",
                "links": null
            },
            "BIBREF15": {
                "ref_id": "b15",
                "title": "SQuAD: 100,000+ questions for machine comprehension of text",
                "authors": [
                    {
                        "first": "Pranav",
                        "middle": [],
                        "last": "Rajpurkar",
                        "suffix": ""
                    },
                    {
                        "first": "Jian",
                        "middle": [],
                        "last": "Zhang",
                        "suffix": ""
                    },
                    {
                        "first": "Konstantin",
                        "middle": [],
                        "last": "Lopyrev",
                        "suffix": ""
                    },
                    {
                        "first": "Percy",
                        "middle": [],
                        "last": "Liang",
                        "suffix": ""
                    }
                ],
                "year": 2016,
                "venue": "Empirical Methods in Natural Language Processing(EMNLP)",
                "volume": "",
                "issue": "",
                "pages": "2383--2392",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Pranav Rajpurkar, Jian Zhang, Konstantin Lopyrev, and Percy Liang. 2016. SQuAD: 100,000+ questions for machine comprehension of text. In Empirical Methods in Natural Language Processing(EMNLP), pages 2383-2392.",
                "links": null
            },
            "BIBREF16": {
                "ref_id": "b16",
                "title": "CoQA: A conversational question answering challenge",
                "authors": [
                    {
                        "first": "Siva",
                        "middle": [],
                        "last": "Reddy",
                        "suffix": ""
                    },
                    {
                        "first": "Danqi",
                        "middle": [],
                        "last": "Chen",
                        "suffix": ""
                    },
                    {
                        "first": "Christopher",
                        "middle": [
                            "D"
                        ],
                        "last": "Manning",
                        "suffix": ""
                    }
                ],
                "year": 2018,
                "venue": "Computing Research Repository (CoRR)",
                "volume": "1",
                "issue": "",
                "pages": "",
                "other_ids": {
                    "arXiv": [
                        "arXiv:1808.07042"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Siva Reddy, Danqi Chen, and Christopher D. Manning. 2018. CoQA: A conversational question answering challenge. Computing Research Repository (CoRR), arXiv:1808.07042. Version 1.",
                "links": null
            },
            "BIBREF17": {
                "ref_id": "b17",
                "title": "Sequence to Sequence Learning with Neural Networks",
                "authors": [
                    {
                        "first": "Ilya",
                        "middle": [],
                        "last": "Sutskever",
                        "suffix": ""
                    },
                    {
                        "first": "Oriol",
                        "middle": [],
                        "last": "Vinyals",
                        "suffix": ""
                    },
                    {
                        "first": "V",
                        "middle": [],
                        "last": "Quoc",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Le",
                        "suffix": ""
                    }
                ],
                "year": 2014,
                "venue": "Advances in neural information processing systems",
                "volume": "",
                "issue": "",
                "pages": "3104--3112",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Ilya Sutskever, Oriol Vinyals, and Quoc V.Le. 2014. Sequence to Sequence Learning with Neural Net- works. In Advances in neural information processing systems, pages 3104-3112.",
                "links": null
            },
            "BIBREF18": {
                "ref_id": "b18",
                "title": "Bidirectional attention flow for machine comprehension",
                "authors": [
                    {
                        "first": "Min Joon",
                        "middle": [],
                        "last": "Seo",
                        "suffix": ""
                    },
                    {
                        "first": "Aniruddha",
                        "middle": [],
                        "last": "Kembhavi",
                        "suffix": ""
                    },
                    {
                        "first": "Ali",
                        "middle": [],
                        "last": "Farhadi",
                        "suffix": ""
                    },
                    {
                        "first": "Hannaneh",
                        "middle": [],
                        "last": "Hajishirzi",
                        "suffix": ""
                    }
                ],
                "year": 2017,
                "venue": "ICLR",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Min Joon Seo, Aniruddha Kembhavi, Ali Farhadi, and Hannaneh Hajishirzi. 2017. Bidirectional attention flow for machine comprehension. In ICLR.",
                "links": null
            },
            "BIBREF19": {
                "ref_id": "b19",
                "title": "MASS: Masked Sequence to Sequence Pre-training for Language Generation",
                "authors": [
                    {
                        "first": "Kaitao",
                        "middle": [],
                        "last": "Song",
                        "suffix": ""
                    },
                    {
                        "first": "Xu",
                        "middle": [],
                        "last": "Tan",
                        "suffix": ""
                    },
                    {
                        "first": "Tao",
                        "middle": [],
                        "last": "Qin",
                        "suffix": ""
                    },
                    {
                        "first": "Jianfeng",
                        "middle": [],
                        "last": "Lu",
                        "suffix": ""
                    },
                    {
                        "first": "Tie-Yan",
                        "middle": [],
                        "last": "Liu",
                        "suffix": ""
                    }
                ],
                "year": 2019,
                "venue": "Proceedings of the 36th InternationalConference on Machine Learning",
                "volume": "97",
                "issue": "",
                "pages": "5926--5936",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Kaitao Song, Xu Tan, Tao Qin, Jianfeng Lu, and Tie-Yan Liu. 2019. MASS: Masked Sequence to Sequence Pre-training for Language Generation. In Proceedings of the 36th InternationalConference on Machine Learning, volume 97, pages 5926-5936. PMLR.",
                "links": null
            },
            "BIBREF20": {
                "ref_id": "b20",
                "title": "Ernie 2.0: A continual pre-training framework for language understanding",
                "authors": [
                    {
                        "first": "Yu",
                        "middle": [],
                        "last": "Sun",
                        "suffix": ""
                    },
                    {
                        "first": "Shuohuan",
                        "middle": [],
                        "last": "Wang",
                        "suffix": ""
                    },
                    {
                        "first": "Yukun",
                        "middle": [],
                        "last": "Li",
                        "suffix": ""
                    },
                    {
                        "first": "Shikun",
                        "middle": [],
                        "last": "Feng",
                        "suffix": ""
                    },
                    {
                        "first": "Hua",
                        "middle": [],
                        "last": "Hao Tian",
                        "suffix": ""
                    },
                    {
                        "first": "Haifeng",
                        "middle": [],
                        "last": "Wu",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Wang",
                        "suffix": ""
                    }
                ],
                "year": 2020,
                "venue": "AAAI",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Yu Sun, Shuohuan Wang, Yukun Li, Shikun Feng, Hao Tian, Hua Wu, and Haifeng Wang. 2020. Ernie 2.0: A continual pre-training framework for language understanding. In AAAI.",
                "links": null
            },
            "BIBREF21": {
                "ref_id": "b21",
                "title": "S-net: From answer extraction to answer synthesis for machine reading comprehension",
                "authors": [
                    {
                        "first": "Chuanqi",
                        "middle": [],
                        "last": "Tan",
                        "suffix": ""
                    },
                    {
                        "first": "Furu",
                        "middle": [],
                        "last": "Wei",
                        "suffix": ""
                    },
                    {
                        "first": "Nan",
                        "middle": [],
                        "last": "Yang",
                        "suffix": ""
                    },
                    {
                        "first": "Bowen",
                        "middle": [],
                        "last": "Du",
                        "suffix": ""
                    },
                    {
                        "first": "Weifeng",
                        "middle": [],
                        "last": "Lv",
                        "suffix": ""
                    },
                    {
                        "first": "Ming",
                        "middle": [],
                        "last": "Zhou",
                        "suffix": ""
                    }
                ],
                "year": 2018,
                "venue": "Association for the Ad-vancement of Artificial Intelligence (AAAI)",
                "volume": "",
                "issue": "",
                "pages": "5940--5947",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Chuanqi Tan, Furu Wei, Nan Yang, Bowen Du, Weifeng Lv, and Ming Zhou. 2018. S-net: From answer extraction to answer synthesis for machine reading comprehension. In Association for the Ad-vancement of Artificial Intelligence (AAAI), pages 5940-5947.",
                "links": null
            },
            "BIBREF22": {
                "ref_id": "b22",
                "title": "Attention is all you need",
                "authors": [
                    {
                        "first": "Ashish",
                        "middle": [],
                        "last": "Vaswani",
                        "suffix": ""
                    },
                    {
                        "first": "Noam",
                        "middle": [],
                        "last": "Shazeer",
                        "suffix": ""
                    },
                    {
                        "first": "Niki",
                        "middle": [],
                        "last": "Parmar",
                        "suffix": ""
                    },
                    {
                        "first": "Jakob",
                        "middle": [],
                        "last": "Uszkoreit",
                        "suffix": ""
                    },
                    {
                        "first": "Llion",
                        "middle": [],
                        "last": "Jones",
                        "suffix": ""
                    },
                    {
                        "first": "Aidan",
                        "middle": [
                            "N"
                        ],
                        "last": "Gomez",
                        "suffix": ""
                    },
                    {
                        "first": "Illia",
                        "middle": [],
                        "last": "Polosukhin",
                        "suffix": ""
                    }
                ],
                "year": 2017,
                "venue": "Advances in Neural Information Processing Systems",
                "volume": "30",
                "issue": "",
                "pages": "5998--6008",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, L ukaszKaiser, and Illia Polosukhin. 2017. Attention is all you need. In I. Guyon, U. V. Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, editors,Advances in Neural Information Processing Systems 30, pages 5998-6008. Curran Associates, Inc.",
                "links": null
            },
            "BIBREF23": {
                "ref_id": "b23",
                "title": "Pointer networks",
                "authors": [
                    {
                        "first": "Oriol",
                        "middle": [],
                        "last": "Vinyals",
                        "suffix": ""
                    },
                    {
                        "first": "Meire",
                        "middle": [],
                        "last": "Fortunato",
                        "suffix": ""
                    },
                    {
                        "first": "Navdeep",
                        "middle": [],
                        "last": "Jaitly",
                        "suffix": ""
                    }
                ],
                "year": 2015,
                "venue": "Advances in Neural Information Processing Systems 28: Annual Conference on Neural Information Processing Systems",
                "volume": "",
                "issue": "",
                "pages": "2692--2700",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Oriol Vinyals, Meire Fortunato, and Navdeep Jaitly. 2015. Pointer networks. In Advances in Neural Information Processing Systems 28: Annual Conference on Neural Information Processing Systems 2015, December 7-12, 2015, Montreal, Quebec, Canada, pages 2692-2700.",
                "links": null
            },
            "BIBREF24": {
                "ref_id": "b24",
                "title": "Reinforced reader-ranker for open-domain question answering",
                "authors": [
                    {
                        "first": "Shuohang",
                        "middle": [],
                        "last": "Wang",
                        "suffix": ""
                    },
                    {
                        "first": "Mo",
                        "middle": [],
                        "last": "Yu",
                        "suffix": ""
                    },
                    {
                        "first": "Xiaoxiao",
                        "middle": [],
                        "last": "Guo",
                        "suffix": ""
                    },
                    {
                        "first": "Zhiguo",
                        "middle": [],
                        "last": "Wang",
                        "suffix": ""
                    },
                    {
                        "first": "Tim",
                        "middle": [],
                        "last": "Klinger",
                        "suffix": ""
                    },
                    {
                        "first": "Wei",
                        "middle": [],
                        "last": "Zhang",
                        "suffix": ""
                    },
                    {
                        "first": "Shiyu",
                        "middle": [],
                        "last": "Chang",
                        "suffix": ""
                    },
                    {
                        "first": "Gerald",
                        "middle": [],
                        "last": "Tesauro",
                        "suffix": ""
                    },
                    {
                        "first": "Bowen",
                        "middle": [],
                        "last": "Zhou",
                        "suffix": ""
                    },
                    {
                        "first": "Jing",
                        "middle": [],
                        "last": "Jiang",
                        "suffix": ""
                    }
                ],
                "year": 2018,
                "venue": "Association for the Advancement of Artificial Intelligence (AAAI)",
                "volume": "",
                "issue": "",
                "pages": "5981--5988",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Shuohang Wang, Mo Yu, Xiaoxiao Guo, Zhiguo Wang, Tim Klinger, Wei Zhang, Shiyu Chang, Gerald Tesauro, Bowen Zhou, and Jing Jiang. 2018. Reinforced reader-ranker for open-domain question answering. In Association for the Advancement of Artificial Intelligence (AAAI), pages 5981-5988.",
                "links": null
            },
            "BIBREF25": {
                "ref_id": "b25",
                "title": "ERNIE-GEN: An Enhanced Multi-Flow Pre-training and Fine-tuning Framework for Natural Language Generation",
                "authors": [
                    {
                        "first": "Dongling",
                        "middle": [],
                        "last": "Xiao",
                        "suffix": ""
                    },
                    {
                        "first": "Han",
                        "middle": [],
                        "last": "Zhang",
                        "suffix": ""
                    },
                    {
                        "first": "Yukun",
                        "middle": [],
                        "last": "Li",
                        "suffix": ""
                    },
                    {
                        "first": "Yu",
                        "middle": [],
                        "last": "Sun",
                        "suffix": ""
                    },
                    {
                        "first": "Hua",
                        "middle": [],
                        "last": "Hao Tian",
                        "suffix": ""
                    },
                    {
                        "first": "Haifeng",
                        "middle": [],
                        "last": "Wu",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Wang",
                        "suffix": ""
                    }
                ],
                "year": 2020,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {
                    "arXiv": [
                        "arXiv:2001.11314"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Dongling Xiao, Han Zhang, Yukun Li, Yu Sun, Hao Tian, Hua Wu, and Haifeng Wang. 2020. ERNIE- GEN: An Enhanced Multi-Flow Pre-training and Fine-tuning Framework for Natural Language Generation. arXiv preprint arXiv:2001.11314.",
                "links": null
            },
            "BIBREF26": {
                "ref_id": "b26",
                "title": "Hotpotqa: A dataset for diverse, explainable multi-hop question answering",
                "authors": [
                    {
                        "first": "Zhilin",
                        "middle": [],
                        "last": "Yang",
                        "suffix": ""
                    },
                    {
                        "first": "Peng",
                        "middle": [],
                        "last": "Qi",
                        "suffix": ""
                    },
                    {
                        "first": "Saizheng",
                        "middle": [],
                        "last": "Zhang",
                        "suffix": ""
                    },
                    {
                        "first": "Yoshua",
                        "middle": [],
                        "last": "Bengio",
                        "suffix": ""
                    },
                    {
                        "first": "William",
                        "middle": [
                            "W"
                        ],
                        "last": "Cohen",
                        "suffix": ""
                    },
                    {
                        "first": "Ruslan",
                        "middle": [],
                        "last": "Salakhutdinov",
                        "suffix": ""
                    },
                    {
                        "first": "Christopher",
                        "middle": [
                            "D"
                        ],
                        "last": "Manning",
                        "suffix": ""
                    }
                ],
                "year": 2018,
                "venue": "EMNLP",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Zhilin Yang, Peng Qi, Saizheng Zhang, Yoshua Bengio, William W.Cohen, Ruslan Salakhutdinov, and Christopher D.Manning. 2018. Hotpotqa: A dataset for diverse, explainable multi-hop question answering. In EMNLP.",
                "links": null
            }
        },
        "ref_entries": {}
    }
}