{
    "paper_id": "2020",
    "header": {
        "generated_with": "S2ORC 1.0.0",
        "date_generated": "2023-01-19T12:52:45.743314Z"
    },
    "title": "A Human-machine Dialogue Intent Classification Method using Utterance Pseudo Label Attention",
    "authors": [
        {
            "first": "Jiande",
            "middle": [],
            "last": "Ding",
            "suffix": "",
            "affiliation": {
                "laboratory": "",
                "institution": "South China Agricultural University",
                "location": {
                    "country": "China"
                }
            },
            "email": ""
        },
        {
            "first": "Peijie",
            "middle": [],
            "last": "Huang",
            "suffix": "",
            "affiliation": {
                "laboratory": "",
                "institution": "South China Agricultural University",
                "location": {
                    "country": "China"
                }
            },
            "email": "pjhuang@scau.edu.cn"
        },
        {
            "first": "Jiabao",
            "middle": [],
            "last": "Xu",
            "suffix": "",
            "affiliation": {
                "laboratory": "",
                "institution": "South China Agricultural University",
                "location": {
                    "country": "China"
                }
            },
            "email": "xujiabao@stu.scau.edu.cn"
        },
        {
            "first": "Youming",
            "middle": [],
            "last": "Peng",
            "suffix": "",
            "affiliation": {
                "laboratory": "",
                "institution": "South China Agricultural University",
                "location": {
                    "country": "China"
                }
            },
            "email": ""
        }
    ],
    "year": "",
    "venue": null,
    "identifiers": {},
    "abstract": "In human-machine dialogue system, it needs to judge the user's intent through the intent classification, and then triggers the corresponding business type. Due to the characteristics of colloquialization, longer texts and sparse features of multi-turn dialogues, the existing classification methods still have great difficulties in the classification of human-machine dialogue intent. Based on hierarchical attention networks (HAN), we propose PLA-HAN model that combines the utterance pseudo-label attention. Through selecting utterance intent set, constructing utterance intent detection model and designing an utterance pseudo-label attention mechanism, PLA-HAN recognizes the pseudo-label of utterance intent and then computes utterance pseudolabel attention. Furthermore, the utterance pseudo-label attention is embedded into *\u901a\u8baf\u4f5c\u8005 c 2020 \u4e2d\u56fd\u8ba1\u7b97\u8bed\u8a00\u5b66\u5927\u4f1a \u6839\u636e\u300aCreative Commons Attribution 4.0 International License\u300b\u8bb8\u53ef\u51fa\u7248 \u8ba1\u7b97\u8bed\u8a00\u5b66 the hierarchical structure of HAN and is integrated with its sentence-level attention. Sentence-level attention that incorporates utterance intent information further improves the overall performance of the model. We conducted experiments on the shared task dataset of \"Customer Intent Classification Evaluation Competition for Customer Service Domain\" sponsored by the Chinese Information Processing Society of China. Experiment results show that the proposed model achieved better performance than HAN on dialogue intent classification.",
    "pdf_parse": {
        "paper_id": "2020",
        "_pdf_hash": "",
        "abstract": [
            {
                "text": "In human-machine dialogue system, it needs to judge the user's intent through the intent classification, and then triggers the corresponding business type. Due to the characteristics of colloquialization, longer texts and sparse features of multi-turn dialogues, the existing classification methods still have great difficulties in the classification of human-machine dialogue intent. Based on hierarchical attention networks (HAN), we propose PLA-HAN model that combines the utterance pseudo-label attention. Through selecting utterance intent set, constructing utterance intent detection model and designing an utterance pseudo-label attention mechanism, PLA-HAN recognizes the pseudo-label of utterance intent and then computes utterance pseudolabel attention. Furthermore, the utterance pseudo-label attention is embedded into *\u901a\u8baf\u4f5c\u8005 c 2020 \u4e2d\u56fd\u8ba1\u7b97\u8bed\u8a00\u5b66\u5927\u4f1a \u6839\u636e\u300aCreative Commons Attribution 4.0 International License\u300b\u8bb8\u53ef\u51fa\u7248 \u8ba1\u7b97\u8bed\u8a00\u5b66 the hierarchical structure of HAN and is integrated with its sentence-level attention. Sentence-level attention that incorporates utterance intent information further improves the overall performance of the model. We conducted experiments on the shared task dataset of \"Customer Intent Classification Evaluation Competition for Customer Service Domain\" sponsored by the Chinese Information Processing Society of China. Experiment results show that the proposed model achieved better performance than HAN on dialogue intent classification.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Abstract",
                "sec_num": null
            }
        ],
        "body_text": [
            {
                "text": "1 \u5f15 \u5f15 \u5f15\u8a00 \u8a00 \u8a00 \u8fd1\u5e74\u6765\uff0c\u4eba\u673a\u5bf9\u8bdd\u7531\u4e8e\u5176\u6f5c\u5728\u7684\u53d1\u5c55\u6f5c\u529b\u548c\u8bf1\u4eba\u7684\u5546\u4e1a\u4ef7\u503c\u800c\u6536\u5230\u8d8a\u6765\u8d8a\u591a\u7684\u5173\u6ce8 (\u4fde\u51ef \u7b49, 2015; Chen et al., 2017) (Haffner et al., 2003; Phan et al., 2008) \u3002\u968f\u540e\uff0c\u6df1 \u5ea6\u5b66\u4e60\u5728\u81ea\u7136\u8bed\u8a00\u5904\u7406(natural language processing, NLP)\u4e2d\u53d7\u5230\u5173\u6ce8\uff0c\u4e3b\u6d41\u7684\u5e94\u7528\u5305\u62ec\u6df1\u5ea6 \u4fe1\u5ff5\u7f51\u7edc(deep belief networks\uff0cDBN) (Sarikaya et al., 2011) \u3001CNN (Xu and Sarikaya, 2013; Kim, 2014) \u548cRNN (Xu and Sarikaya, 2014) \u7b49\uff0c\u5c24\u5176\u662fRNN\u4e2d\u6700\u5e38\u7528\u7684LSTM (Cheng et al., 2016; Ravuri and Stolcke, 2016; Vu et al., 2016 ; \u67ef\u5b50\u7b49, 2018)\u3002\u8fd1\u5e74\u6765\uff0c\u6ce8\u610f\u529b\u673a\u5236\u88ab\u5f15\u5165\u5230 \u4e86NLP\u4e2d\uff0c\u5b9e\u9a8c\u8bc1\u660e\u5176\u5584\u4e8e\u5728\u6587\u672c\u5206\u7c7b\u4efb\u52a1\u4e2d\u62bd\u53d6\u6587\u672c\u7684\u542b\u4e49\uff0c\u4f8b\u5982\u8bdd\u8bed\u610f\u56fe\u68c0\u6d4b (Liu and Lane, 2016) \u3001\u8bdd\u8bed\u9886\u57df\u5206\u7c7b (Kim et al., 2018) \u3001\u95ee\u7b54\u60c5\u611f\u5206\u7c7b (\u5b89\u660e\u6167\u7b49, 2019)\u548c\u9488\u5bf9\u8bed\u7bc7\u957f \u6587\u672c\u7684\u6587\u6863\u5206\u7c7b (Yang et al., 2016) \u7b49\u3002\u5c3d\u7ba1\u57fa\u4e8e\u6df1\u5ea6\u5b66\u4e60\u7684\u5206\u7c7b\u6a21\u578b\uff0c\u5c24\u5176\u662f\u5c42\u6b21\u6ce8\u610f\u529b\u7f51 \u7edc(hierarchical attention network, HAN) (Yang et al., 2016) ",
                "cite_spans": [
                    {
                        "start": 50,
                        "end": 62,
                        "text": "(\u4fde\u51ef \u7b49, 2015;",
                        "ref_id": "BIBREF2"
                    },
                    {
                        "start": 63,
                        "end": 81,
                        "text": "Chen et al., 2017)",
                        "ref_id": "BIBREF3"
                    },
                    {
                        "start": 82,
                        "end": 104,
                        "text": "(Haffner et al., 2003;",
                        "ref_id": "BIBREF7"
                    },
                    {
                        "start": 105,
                        "end": 123,
                        "text": "Phan et al., 2008)",
                        "ref_id": "BIBREF11"
                    },
                    {
                        "start": 221,
                        "end": 244,
                        "text": "(Sarikaya et al., 2011)",
                        "ref_id": "BIBREF13"
                    },
                    {
                        "start": 250,
                        "end": 273,
                        "text": "(Xu and Sarikaya, 2013;",
                        "ref_id": "BIBREF15"
                    },
                    {
                        "start": 274,
                        "end": 284,
                        "text": "Kim, 2014)",
                        "ref_id": "BIBREF9"
                    },
                    {
                        "start": 290,
                        "end": 313,
                        "text": "(Xu and Sarikaya, 2014)",
                        "ref_id": "BIBREF16"
                    },
                    {
                        "start": 332,
                        "end": 352,
                        "text": "(Cheng et al., 2016;",
                        "ref_id": "BIBREF4"
                    },
                    {
                        "start": 353,
                        "end": 378,
                        "text": "Ravuri and Stolcke, 2016;",
                        "ref_id": "BIBREF12"
                    },
                    {
                        "start": 379,
                        "end": 394,
                        "text": "Vu et al., 2016",
                        "ref_id": "BIBREF14"
                    },
                    {
                        "start": 460,
                        "end": 480,
                        "text": "(Liu and Lane, 2016)",
                        "ref_id": "BIBREF8"
                    },
                    {
                        "start": 489,
                        "end": 507,
                        "text": "(Kim et al., 2018)",
                        "ref_id": "BIBREF10"
                    },
                    {
                        "start": 543,
                        "end": 562,
                        "text": "(Yang et al., 2016)",
                        "ref_id": "BIBREF18"
                    },
                    {
                        "start": 628,
                        "end": 647,
                        "text": "(Yang et al., 2016)",
                        "ref_id": "BIBREF18"
                    }
                ],
                "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": "2 \u57fa \u57fa \u57fa\u7840 \u7840 \u7840\u5de5 \u5de5 \u5de5\u4f5c \u4f5c \u4f5c 2.1 \u5c42 \u5c42 \u5c42\u6b21 \u6b21 \u6b21\u6ce8 \u6ce8 \u6ce8\u610f \u610f \u610f\u529b \u529b \u529b\u673a \u673a \u673a\u5236 \u5236 \u5236 \u66f4\u7b26\u5408\u957f\u6587\u672c\u5c42\u7ea7\u7ed3\u6784\u7684\u662fHAN(Yang et al., 2016)\u6240\u91c7\u7528\u7684\u5c42\u6b21\u6ce8\u610f\u529b\u673a\u5236\uff0c\u5b83\u5e26\u6709\u4e24\u4e2a \u5c42\u7ea7\u7684\u6ce8\u610f\u529b\u673a\u5236\uff0c\u5206\u522b\u662f\u8bcd\u7ea7\u522b\u548c\u53e5\u5b50\u7ea7\u522b\uff0c\u80fd\u591f\u66f4\u597d\u5730\u8868\u793a\u957f\u6587\u672c\u4e2d\u91cd\u8981\u4fe1\u606f\u7684\u4f4d\u7f6e\u3002 \u6ce8 \u6ce8 \u6ce8\u610f \u610f \u610f\u529b \u529b \u529b\u673a \u673a \u673a\u5236 \u5236 \u5236\u3002\u8fd1\u5e74\u6765\uff0c\u5728\u6587\u672c\u5206\u7c7b\u95ee\u9898\u4e0a\uff0c\u57fa\u4e8e\u6ce8\u610f\u529b\u673a\u5236\u7684\u6a21\u578b\u5728\u6548\u679c\u548c\u6548\u7387\u4e0a\u90fd\u5c55\u73b0 \u51fa\u4e86\u4e00\u5b9a\u7684\u4f18\u8d8a\u6027\u3002\u6211\u4eec\u8003\u8651\u975e\u5c42\u7ea7\u7ed3\u6784\u7684\u8f6f\u6ce8\u610f\u529b\u673a\u5236\u4ee5\u53ca\u5c42\u7ea7\u7ed3\u6784\u7684\u5206\u5c42\u6ce8\u610f\u529b\u673a\u5236\u3002 \u6ce8\u610f\u529b\u673a\u5236\u6700\u65e9\u5728\u673a\u5668\u7ffb\u8bd1\u4e2d\u5e94\u7528\uff0c\u73b0\u5728\u5df2\u7ecf\u6210\u4e3a\u795e\u7ecf\u7f51\u7edc\u76f8\u5173\u4e2d\u4e00\u4e2a\u5341\u5206\u5177\u6709\u5f71\u54cd\u529b \u7684\u6982\u5ff5\uff0c\u6ce8\u610f\u529b\u673a\u5236\u76f8\u5f53\u4e8e\u6a21\u4eff\u4eba\u7c7b\u5c06\u5927\u91cf\u89c6\u89c9\u4fe1\u606f\u538b\u7f29\u6210\u63cf\u8ff0\u6027\u8bed\u8a00\u7684\u975e\u51e1\u80fd\u529b(Xu et al., 2015)\uff0c\u5b83\u5c06\u7f16\u7801\u5668\u7684\u8f93\u51fa\u6620\u5c04\u4e3a\u6ce8\u610f\u529b\u6743\u91cd\uff0c\u5c06\u6743\u91cd\u4e0e\u7f16\u7801\u5668\u8f93\u51fa\u8fdb\u884c\u52a0\u6743\uff0c\u5982\u5f0f(1)-(3)\u6240 \u793a\u3002 u i = score(h i )",
                        "eq_num": "(1)"
                    }
                ],
                "section": "",
                "sec_num": null
            },
            {
                "text": "\u03b1 i = exp(u i ) L i (u i ) (2) h * = L i \u03b1 i h i (3)",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            },
            {
                "text": "\u5176\u4e2d\uff0c\u672c\u6587\u91c7\u7528\u524d\u9988\u795e\u7ecf\u7f51\u7edc\u4e3a\u5bf9\u9f50\u51fd\u6570(score)\u3001\u5728\u8bed\u97f3\u8bc6\u522b (Graves et al., 2013 )\u548c\u53e3\u8bed \u8bed\u8a00\u7406\u89e3 (Xu and Sarikaya, 2014) ",
                "cite_spans": [
                    {
                        "start": 32,
                        "end": 52,
                        "text": "(Graves et al., 2013",
                        "ref_id": "BIBREF6"
                    },
                    {
                        "start": 63,
                        "end": 86,
                        "text": "(Xu and Sarikaya, 2014)",
                        "ref_id": "BIBREF16"
                    }
                ],
                "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": "\u9886\u57df\u6709\u7740\u975e\u5e38\u6210\u529f\u5e94\u7528\u7684\u53cc\u5411LSTM\u4f5c\u4e3a\u7f16\u7801\u5668\u3002\u901a\u8fc7\u5bf9\u9f50\u51fd \u6570\u5c06\u7f16\u7801\u5668\u7684\u8f93\u51fa\u5bf9\u9f50\uff0c\u5f97\u5230\u6ce8\u610f\u529b\u5206\u6570\uff0c\u8fdb\u884c\u5f52\u4e00\u5316\u540e\u5f97\u5230\u6ce8\u610f\u529b\u6743\u91cd\u03b1\uff0c\u6700\u7ec8\u5c06\u7f16\u7801\u5668\u8f93 \u51fa\u4e0e\u6ce8\u610f\u529b\u6743\u91cd\u8fdb\u884c\u52a0\u6743\u3002 \u8bcd \u8bcd \u8bcd \u7ea7 \u7ea7 \u7ea7 \u522b \u522b \u522b \u3002 \u5bf9 \u4e8e \u4e00 \u4e2a \u53e5 \u5b50 \u7684 \u8bcdw * \uff0c \u5728 \u7ecf \u8fc7embedding\u4e4b \u540e \uff0c \u83b7 \u5f97 \u8bcd \u5411 \u91cfw * emb \uff0c \u7f16 \u7801 \u5668 \u5bf9 \u8bcd \u5411 \u91cf \u8fdb \u884c \u7f16 \u7801 \uff0c \u83b7 \u5f97 \u8bcd \u7ea7 \u522b \u7684 \u8868 \u8fbeh * word \uff0c \u901a \u8fc7 \u5355 \u5c42MLP\u5bf9h * word \u8fdb \u884c \u5bf9 \u9f50 \u8ba1 \u7b97 \uff0c \u83b7 \u5f97u * word \uff0c\u4f7f\u7528softmax\u51fd\u6570\u5f97\u5230\u5b57\u7ea7\u522b\u7684\u6ce8\u610f\u529b\u6743\u91cd\u03b1 * word \uff0c\u5728\u8fd9\u4e4b\u540e\uff0c\u901a\u8fc7\u5b57\u7ea7\u522b\u6ce8\u610f\u529b \u6743\u91cd\u03b1 * word \u4e0eh * word \u52a0\u6743\u6c42\u548c\u5f97\u5230\u53e5\u5b50\u5411\u91cfs * \u3002\u5982\u5f0f(4)-(8)\u6240\u793a\u3002 w * emb = embedding(w * ) (4) h * word = Encoder(w * emb ) (5) u * word = tanh(W w h * word + b w ) (6) \u03b1 * word = exp(u * word u w ) t exp(u * word u w )",
                        "eq_num": "(7)"
                    }
                ],
                "section": "",
                "sec_num": null
            },
            {
                "text": "EQUATION",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [
                    {
                        "start": 0,
                        "end": 8,
                        "text": "EQUATION",
                        "ref_id": "EQREF",
                        "raw_str": "s * = t (\u03b1 * word word * emb ) (8) \u53e5 \u53e5 \u53e5 \u5b50 \u5b50 \u5b50 \u7ea7 \u7ea7 \u7ea7 \u522b \u522b \u522b \u3002 \u4e0e \u8bcd \u7ea7 \u522b \u7684 \u505a \u6cd5 \u76f8 \u4f3c \uff0c \u7f16 \u7801 \u5668 \u5bf9 \u53e5 \u5b50 \u5411 \u91cfs * \u8fdb \u884c \u7f16 \u7801 \u83b7 \u5f97 \u53e5 \u5b50 \u5411 \u91cf \u7684 \u8868 \u8fbeh * sentence \uff0c\u540c\u6837\u901a\u8fc7\u5355\u5c42MLP\u5bf9\u53e5\u5b50\u5411\u91cf\u8fdb\u884c\u5bf9\u9f50\u8ba1\u7b97\uff0c\u83b7\u5f97u * sentence \uff0c\u4f7f\u7528softmax\u51fd\u6570 \u5f97\u5230\u53e5\u5b50\u7ea7\u522b\u7684\u6ce8\u610f\u529b\u6743\u91cd\u03b1 * sentence \uff0c\u6700\u7ec8\u901a\u8fc7\u53e5\u5b50\u7ea7\u522b\u6ce8\u610f\u529b\u6743\u91cd\u03b1 * sentence \u4e0eh * sentence \u52a0\u6743\u6c42 \u548c\u4f5c\u4e3a\u6587\u6863\u5411\u91cfv\u3002\u5982\u5f0f(9)-(12)\u6240\u793a\u3002 h * sentence = Encoder(s * ) (9) u * sentence = tanh(W s h * sentence + b s ) (10) \u03b1 * sentence = exp(u * sentence u s ) t exp(u * sentence u s )",
                        "eq_num": "(11)"
                    }
                ],
                "section": "",
                "sec_num": null
            },
            {
                "text": "EQUATION",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [
                    {
                        "start": 0,
                        "end": 8,
                        "text": "EQUATION",
                        "ref_id": "EQREF",
                        "raw_str": "s * = t (\u03b1 * sentence sentence * emb ) (12) 2.2 \u5355 \u5355 \u5355\u53e5 \u53e5 \u53e5\u8bdd \u8bdd \u8bdd\u8bed \u8bed \u8bed\u610f \u610f \u610f\u56fe \u56fe \u56fe\u8bc6 \u8bc6 \u8bc6\u522b \u522b \u522b\u6a21 \u6a21 \u6a21\u578b \u578b \u578b \u5bf9 \u4e8e \u5355 \u53e5 \u8bdd \u8bed \u7684 \u610f \u56fe \u8bc6 \u522b \uff0c \u6211 \u4eec \u91c7 \u7528 \u4e86 \u57fa \u4e8eBERT\u7684 \u53cc \u5411LSTM\u6a21 \u578b \uff0cBiLSTM\u7ed3 \u6784 \u5728 \u8bed \u97f3 \u8bc6 \u522b(Graves et al., 2013)\u548c \u53e3 \u8bed \u8bed \u8a00 \u7406 \u89e3(Xu and Sarikaya, 2014)\u9886 \u57df \u6709 \u7740 \u975e \u5e38 \u6210 \u529f \u7684 \u5e94 \u7528 \u3002 \u5728BERT\u5c42 \uff0c \u6211 \u4eec \u91c7 \u7528 \u9884 \u5148 \u8bad \u7ec3 \u597d \u7684 \u4e2d \u6587BERT\u6a21 \u578b \u5bf9 \u56fa \u5b9a \u957f \u5ea6 \u4e3aL\u7684 \u8bdd \u8bed \u5e8f \u5217w * = {w 1 , w 2 , ..., w L }\u8fdb \u884c \u7f16 \u7801 \uff0c \u6bcf \u4e2a \u5b57w i \u4f1a \u88ab \u7f16 \u7801 \u4e3a \u5b57 \u5411 \u91cfe i \uff1b \u8fd9 \u6837w * \u5c31 \u88ab \u7f16 \u7801 \u6210 \u4e86emb u = {e 1 , e 2 , ..., e L }\u3002\u5982\u5f0f(13)\u6240\u793a\u3002 emb u = BERT (w * )",
                        "eq_num": "(13)"
                    }
                ],
                "section": "",
                "sec_num": null
            },
            {
                "text": "EQUATION",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [
                    {
                        "start": 0,
                        "end": 8,
                        "text": "EQUATION",
                        "ref_id": "EQREF",
                        "raw_str": "\u5728BiLSTM\u7f51\u7edc\u5c42\uff0c\u6211\u4eec\u5c06emb u = {e 1 , e 2 , ..., e L }\u8f93\u5165\u5230\u53cc\u5411\u7684\u957f\u77ed\u65f6\u8bb0\u5fc6\u7f51\u7edc\uff0c\u7136\u540e\u5206 \u522b\u4ece\u7f51\u7edc\u4e2d\u5f97\u5230\u6b63\u5411\u7684\u8f93\u51fah f w = {h f w 1 , h f w 2 , ...h f w L }\u548c\u53cd\u5411\u7684\u8f93\u51fah bw = {h bw 1 , h bw 2 , ...h bw L }\u3002\u5982 \u5f0f(14)\u6240\u793a\u3002 h f w , hbw = BiLST M (emb u )",
                        "eq_num": "(14)"
                    }
                ],
                "section": "",
                "sec_num": null
            },
            {
                "text": "\u6211\u4eec\u5c06\u6b63\u5411\u548c\u53cd\u5411\u7684\u7ed3\u679c\u62fc\u63a5\u8d77\u6765\u5f97\u5230\u53cc\u5411LSTM\u7684\u8f93\u51fah\u3002\u5982\u5f0f(15)\u6240\u793a\u3002",
                "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": "h = [h f w , f bw ]",
                        "eq_num": "(15)"
                    }
                ],
                "section": "",
                "sec_num": null
            },
            {
                "text": "EQUATION",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [
                    {
                        "start": 0,
                        "end": 8,
                        "text": "EQUATION",
                        "ref_id": "EQREF",
                        "raw_str": "\u5bf9 \u4e8e \u8f93 \u5165 \u5e8f \u5217 \u4e2d \u7684 \u6bcf \u4e2a \u5143 \u7d20 \uff0c \u6bcf \u4e2aLSTM\u7ed3 \u6784 \u8ba1 \u7b97 \u4ee5 \u4e0b \u51fd \u6570 \uff0c \u5176 \u4e2dh t \u662ft\u65f6 \u523b \u7684 \u9690 \u85cf \u5c42\uff0cc t \u662ft\u65f6\u523b\u7684\u8bb0\u5fc6\u5355\u5143\uff0cx t \u662ft\u65f6\u523b\u7684\u8f93\u5165\uff0ch t\u22121 \u662ft-1\u65f6\u523b\u7684\u9690\u85cf\u5c42\uff0ci t \u3001f t \u3001c t \u548co t \u5206\u522b \u662f\u8f93\u5165\u95e8\u3001\u9057\u5fd8\u95e8\u3001\u8bb0\u5fc6\u5355\u5143\u95e8\u548c\u8f93\u51fa\u95e8\uff0c\u03c3\u662fsigmoid\u51fd\u6570\uff0c \u662f\u54c8\u8fbe\u739b\u79ef\u3002\u5982\u5f0f(16)-(21)\u6240 \u793a\u3002 i t = \u03c3(W ii x t + b ii + W hi h t\u22121 + b hi ) (16) f t = \u03c3(W if x t + b if + W hf h t\u22121 + b hf )",
                        "eq_num": "(17)"
                    }
                ],
                "section": "",
                "sec_num": null
            },
            {
                "text": "EQUATION",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [
                    {
                        "start": 0,
                        "end": 8,
                        "text": "EQUATION",
                        "ref_id": "EQREF",
                        "raw_str": "c t = tanh(W ig x t + b ig + W hg h t\u22121 + b hg )",
                        "eq_num": "(18)"
                    }
                ],
                "section": "",
                "sec_num": null
            },
            {
                "text": "EQUATION",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [
                    {
                        "start": 0,
                        "end": 8,
                        "text": "EQUATION",
                        "ref_id": "EQREF",
                        "raw_str": "o t = \u03c3(W io x t + b io + W ho h t\u22121 + b ho )",
                        "eq_num": "(19)"
                    }
                ],
                "section": "",
                "sec_num": null
            },
            {
                "text": "EQUATION",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [
                    {
                        "start": 0,
                        "end": 8,
                        "text": "EQUATION",
                        "ref_id": "EQREF",
                        "raw_str": "c t = f t c t\u22121 + i t c t (20) h t = o t tanh(c t ) (21) \u63a5\u7740\u5c06h\u7d2f\u52a0\u5f97\u5230\u8bdd\u8bed\u5e8f\u5217\u7684\u53e5\u5b50\u5411\u91cfs\uff0c\u6700\u540e\u63a5\u4e0a\u5168\u8fde\u63a5-softmax\u5c42\u5f97\u5230\u5404\u4e2a\u610f\u56fe\u7c7b\u522b\u7684 \u6982\u7387\u8f93\u51fa\u3002\u5982\u5f0f(22)\u6240\u793a\u3002 y = sof tmax(W h + b) (22) 3 \u672c \u672c \u672c\u6587 \u6587 \u6587\u7684 \u7684 \u7684\u65b9 \u65b9 \u65b9\u6cd5 \u6cd5 \u6cd5 \u5728\u672c\u8282\u4e2d\uff0c\u6211\u4eec\u5c06\u4ecb\u7ecd\u5982\u4f55\u5c06\u610f\u56fe\u4f2a\u6807\u7b7e\u6a21\u578b\u5b66\u5230\u7684\u4f2a\u6807\u7b7e\u6ce8\u610f\u529b\u6743\u91cd\u548c\u5c42\u7ea7\u6ce8\u610f\u529b\u6743\u91cd \u76f8\u878d\u5408\uff0c\u4ece\u800c\u6784\u6210\u5c42\u7ea7\u7ed3\u6784\u4e2d\u5d4c\u5165\u5355\u53e5\u8bdd\u8bed\u610f\u56fe\u4f2a\u6807\u7b7e\u6ce8\u610f\u529b\u7684\u673a\u5236\u3002 3.1 \u4f2a \u4f2a \u4f2a\u6807 \u6807 \u6807\u7b7e \u7b7e \u7b7e\u6ce8 \u6ce8 \u6ce8\u610f \u610f \u610f\u529b \u529b \u529b \u5728\u9762\u5411\u957f\u6587\u672c\u7684\u4efb\u52a1\u578b\u610f\u56fe\u5206\u7c7b\u573a\u666f\u4e2d\uff0c\u80fd\u591f\u63d0\u53d6\u5230\u957f\u6587\u4e2d\u7684\u5173\u952e\u5355\u53e5\u8bdd\u8bed\u5c06\u4f1a\u5bf9\u6574\u4f53\u7684 \u5206\u7c7b\u6548\u679c\u5e26\u6765\u63d0\u5347\u3002\u9664\u4e86\u5728\u6a21\u578b\u5185\u90e8\u91c7\u7528\u6ce8\u610f\u529b\u673a\u5236\u6765\u5b9e\u73b0\uff0c\u4e5f\u53ef\u4ee5\u901a\u8fc7\u5916\u90e8\u7ed9\u6a21\u578b\u5e26\u6765\u989d\u5916 \u7684\u4fe1\u606f\u4ee5\u66f4\u597d\u5730\u63d0\u53d6\u5230\u5173\u952e\u5355\u53e5\u8bdd\u8bed\u3002\u5728\u6b64\uff0c\u6211\u4eec\u901a\u8fc7\u5355\u53e5\u8bdd\u8bed\u7684\u610f\u56fe\u4f2a\u6807\u7b7e\u6765\u5b9e\u73b0\u5c06\u5916\u90e8\u4fe1 \u606f\u6ce8\u5165\u6a21\u578b\u7ed3\u6784\u4e2d\u3002\u957f\u6587\u672c\u7684\u6bcf\u4e2a\u5355\u53e5\u8bdd\u8bed\u610f\u56fe\u4f2a\u6807\u7b7e\u7684\u52a0\u5165\uff0c\u53ef\u4ee5\u89c6\u4e3a\u6784\u5efa\u4e86\u4e00\u4e2a\u5185\u90e8\u7684\"\u5bf9 \u8bdd\u610f\u56fe-\u4f2a\u610f\u56fe\u6807\u7b7e\"\u7684\u5206\u5e03\uff0c\u53cd\u6620\u603b\u4f53\u5bf9\u8bdd\u6240\u8868\u8fbe\u7684\u610f\u56fe\u503e\u5411\u3002\u6211\u4eec\u901a\u8fc7\u4e00\u4e2a\u6ce8\u610f\u529b\u6a21\u578b\u5b66\u4e60 \u5230\u4e86\u6bcf\u4e2a\u4f2a\u6807\u7b7e\u6240\u5bf9\u5e94\u7684\u6ce8\u610f\u529b\uff0c\u8fd9\u79cd\u6ce8\u610f\u529b\u4f5c\u4e3a\u4e00\u79cd\u989d\u5916\u7684\u4fe1\u606f\uff0c\u53cd\u6620\u4e86\u5176\u5bf9\u5e94\u7684\u5355\u53e5\u8bdd\u8bed \u5bf9\u4e8e\u603b\u4f53\u5bf9\u8bdd\u610f\u56fe\u7684\u91cd\u8981\u7a0b\u5ea6\uff0c\u5229\u7528\u8be5\u4fe1\u606f\u53ef\u4ee5\u8ba9\u6a21\u578b\u66f4\u597d\u5730\u9009\u62e9\u51fa\u91cd\u8981\u7684\u5355\u53e5\u8bdd\u8bed\u3002 \u901a \u8fc7 \u76f8 \u5173 \u8054 \u7684 \u5355 \u53e5 \u8bdd \u8bed \u610f \u56fe \u4efb \u52a1 \uff0c \u6211 \u4eec \u8bad \u7ec3 \u4e86 \u4e00 \u4e2a \u9488 \u5bf9 \u5355 \u53e5 \u8bdd \u8bed \u7684 \u610f \u56fe \u8bc6 \u522b \u6a21 \u578b(\u8be6 \u89c12.2\u8282)\uff0c \u8bb0 \u4e3amodel intent \uff0c \u901a \u8fc7 \u6b64 \u6a21 \u578b \uff0c \u7ed9 \u4e3b \u4efb \u52a1 \u4e2d \u7684 \u6bcf \u4e2a \u5b50 \u53e5 \u6807 \u6ce8 \u4e0a \u610f \u56fe \u4f2a \u6807 \u7b7eintent pseudo \uff0c \u4f7f \u7528one-hot\u5bf9 \u610f \u56fe \u4f2a \u6807 \u7b7e \u8fdb \u884c \u72ec \u70ed \u7f16 \u7801 \uff0c \u83b7 \u5f97 \u610f \u56fe \u4f2a \u6807 \u7b7e \u7684 \u6587 \u672c \u8868 \u793aintent pseudo emb * \uff0c \u7ecf \u8fc7 \u7f16 \u7801 \u5668 \u5f97 \u5230intent pseudo h * \uff0c \u5728 \u6b64 \u5904 \u91c7 \u7528 \u8f6f \u6ce8 \u610f \u529b \u6765 \u83b7 \u53d6 \u6ce8 \u610f \u529b \u6743 \u91cd\u03b2 pseudo (\u8be6\u89c12.1\u8282\u7684\u6ce8\u610f\u529b\u673a\u5236\u90e8\u5206)\uff0c\u8bb0\u4e3a\u4f2a\u6807\u7b7e\u610f\u56fe\u6ce8\u610f\u529b\uff0c\u901a\u8fc7\u4f2a\u6807\u7b7e\u610f\u56fe\u6ce8\u610f\u529b\u53ef \u4ee5\u53cd\u6620\u51fa\"\u5bf9\u8bdd\u610f\u56fe-\u4f2a\u610f\u56fe\u6807\u7b7e\"\u7684\u5206\u5e03\uff0c\u4ece\u800c\u8868\u8fbe\u5355\u53e5\u8bdd\u8bed\u610f\u56fe\u7684\u91cd\u8981\u7a0b\u5ea6\u3002\u5982\u5f0f(23)-(25)\u6240 \u793a\u3002 intent pseudo emb * = OneHot(intent pseudo )",
                        "eq_num": "(23)"
                    }
                ],
                "section": "",
                "sec_num": null
            },
            {
                "text": "intent pseudo (1) HAN\u7684\u5b57\u7ea7\u6ce8\u610f\u529b\uff1a\u5728HAN\u4e2d\uff0c\u5bf9\u4e8e\u7ecf\u8fc7BERT\u7f16\u7801\u7684\u7bc7\u7ae0\u7ed3\u6784\u4f5c\u4e3a\u8f93\u5165\u8ba1\u7b97\u5f97\u5230\u8bcd \u7ea7\u522b\u7684\u6ce8\u610f\u529b\u03b1 word (\u8be6\u89c12.1\u8282)\u3002",
                "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": "h * = Encoder(intent pesudo emb * )",
                        "eq_num": "(24)"
                    }
                ],
                "section": "",
                "sec_num": null
            },
            {
                "text": "EQUATION",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [
                    {
                        "start": 0,
                        "end": 8,
                        "text": "EQUATION",
                        "ref_id": "EQREF",
                        "raw_str": "\u03b2 pseudo = Attention(intent pseudo h * )",
                        "eq_num": "(25)"
                    }
                ],
                "section": "",
                "sec_num": null
            },
            {
                "text": "(2) \u610f\u56fe\u4f2a\u6807\u7b7e\u8bc6\u522b\uff1a\u901a\u8fc7\u5355\u53e5\u8bdd\u8bed\u610f\u56fe\u8bc6\u522b\u6a21\u578b\uff0c\u7ed9\u5bf9\u8bdd\u6bb5\u4e2d\u7684\u6bcf\u4e2a\u8bdd\u8bed\u6253\u4e0a\u4f2a\u610f\u56fe\u6807 \u7b7e\uff0c\u6b64\u65f6\u5f97\u5230\u5bf9\u8bdd\u6bb5\u7684\"\u5bf9\u8bdd\u610f\u56fe-\u4f2a\u6807\u7b7e\u610f\u56fe\"\u5206\u5e03\u3002 PLA-HAN\u53e5 \u53e5 \u53e5\u5b50 \u5b50 \u5b50\u7ea7 \u7ea7 \u7ea7\u5c42 \u5c42 \u5c42\u3002\u5305\u62ec\u4e86HAN\u6a21\u578b\u4e2d\u7684\u53e5\u5b50\u7ea7\u6ce8\u610f\u529b\u7f51\u7edc(\u56fe1\u7684\u5de6\u4e2d\u90e8\u5206)\u548c\u5355\u53e5\u8bdd \u8bed\u7684\u610f\u56fe\u4f2a\u6807\u7b7e\u6ce8\u610f\u529b\u673a\u5236(\u56fe1\u7684\u53f3\u4e2d\u90e8\u5206)\u3002",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            },
            {
                "text": "(1) HAN\u7684\u53e5\u5b50\u7ea7\u6ce8\u610f\u529b\uff1a\u5728HAN\u4e2d\uff0c\u5728\u53e5\u5b50\u5411\u91cf\u7684\u57fa\u7840\u4e0a\u8fdb\u4e00\u6b65\u8ba1\u7b97\u53e5\u5b50\u7ea7\u522b\u7684\u6ce8\u610f \u529b\u03b1 sentence (\u8be6\u89c12.1\u8282)\u3002",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            },
            {
                "text": "(2) \u8bdd\u8bed\u4f2a\u6807\u7b7e\u6ce8\u610f\u529b\uff1a\u540c\u65f6\uff0c\u6211\u4eec\u91c7\u75283.1\u8282\u4e2d\u4ecb\u7ecd\u7684\u65b9\u6cd5\uff0c\u901a\u8fc7\u989d\u5916\u6784\u5efa\u7684\u6ce8\u610f\u529b\u6a21 \u578b\uff0c\u5c06\"\u5bf9\u8bdd\u610f\u56fe-\u4f2a\u6807\u7b7e\u610f\u56fe\"\u5206\u5e03\u8f6c\u5316\u4e3a\u4f2a\u6807\u7b7e\u6ce8\u610f\u529b\u03b2 pseudo \u3002 \u534f \u534f \u534f\u540c \u540c \u540c\u6ce8 \u6ce8 \u6ce8\u610f \u610f \u610f\u529b \u529b \u529b\u5c42 \u5c42 \u5c42\u3002\u6211\u4eec\u91c7\u7528\u52a0\u6cd5\u5c06\u5206\u5c42\u6ce8\u610f\u529b\u673a\u5236\u7684\u53e5\u5b50\u7ea7\u522b\u6ce8\u610f\u529b\u548c\u4f2a\u6807\u7b7e\u6ce8\u610f\u529b\u76f8\u878d \u5408\uff0c\u8bb0\u4e3a\u03c9 i \u3002\u5982\u5f0f(26)\u6240\u793a\u3002 ",
                "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": "\u03c9 i = \u03b1 sentence + \u03b2 pseudo (26) \u6700\u540e\u6211\u4eec\u5229\u7528\u6ce8\u610f\u529b\u6743\u91cd\u5bf9\u5206\u5c42\u6ce8\u610f\u529b\u7ed3\u6784\u4e2d\u7684\u53e5\u5b50\u5411\u91cfs * \u8fdb\u884c\u52a0\u6743\u7f29\u653e\uff0c\u5e76\u5c06\u5176\u7d2f\u52a0\u5f97 \u5230\u7684h * \u4f5c\u4e3a\u6ce8\u610f\u529b\u5c42\u7684\u8f93\u51fa\u3002\u5982\u5f0f(27)\u6240\u793a\u3002 h * = t (\u03c9 i * s i ) (27) \u5168 \u5168 \u5168 \u8fde \u8fde \u8fde \u63a5 \u63a5 \u63a5 \u5c42 \u5c42 \u5c42 \u3002 \u3002 \u3002 \u6211 \u4eec \u5229 \u7528 \u5168 \u8fde \u63a5 \u5c42 \u5c06 \u6a21 \u578b \u7684 \u8f93 \u51fa \u6620 \u5c04 \u4e3a \u76f8 \u5e94 \u610f \u56fe \u7c7b \u522b \u6570 \u91cf \uff0c \u63a5 \u7740 \u4f7f \u7528softmax\u8f93\u51fa\u5404\u4e2a\u610f\u56fe\u7c7b\u522b\u7684\u6982\u7387\uff0c\u6700\u7ec8\u91c7\u7528\u6982\u7387\u6700\u9ad8\u7684\u610f\u56fe\u7c7b\u522b\u4f5c\u4e3a\u8f93\u51fa\u3002\u516c\u5f0f\u5982\u5f0f(28)- (29)\u6240\u793a\u3002 y * = sof tmax(W * h * + b) (28) predict = argmax(y * )",
                        "eq_num": "(29"
                    }
                ],
                "section": "",
                "sec_num": null
            },
            {
                "text": "\u5728\u8986\u76d6\u7387\u5206\u6790\u7684\u57fa\u7840\u4e0a\uff0c\u6211\u4eec\u8fdb\u4e00\u6b65\u4ece\u8bed\u4e49\u89d2\u5ea6\u6311\u9009\u51fa\u9ad8\u5173\u8054\u5ea6\u7684\u4f2a\u6807\u7b7e\u5d4c\u5165\u5230\u6a21\u578b\u4e2d\uff0c \u53bb\u6389\u4e86\u90e8\u5206\u5173\u8054\u5ea6\u4e0d\u9ad8\u7684\u6807\u7b7e\u3002\u5728\u672c\u6587\u7684\u5b9e\u9a8c\u4e2d\uff0c\u7528\u4ee5\u8bad\u7ec3\u5355\u53e5\u8bdd\u8bed\u610f\u56fe\u8bc6\u522b\u6a21\u578b\u7684\u6570\u636e\u96c6\u5171 \u670948\u79cd\u5355\u53e5\u8bdd\u8bed\u610f\u56fe\u7c7b\u522b\u3002\u7ecf\u8fc7\u4f2a\u6807\u7b7e\u6311\u9009\uff0c\u6211\u4eec\u7684\u6700\u7ec8\u65b9\u6848\u4e2d\uff0c\u9009\u53d6\u4e8625\u79cd\u4e0e\u4eba\u673a\u5bf9\u8bdd\u610f\u56fe \u5206\u7c7b\u4efb\u52a1\u9ad8\u5173\u8054\u5ea6\u7684\u5355\u53e5\u8bdd\u8bed\u610f\u56fe\uff0c\u540c\u65f6\u5c06\u5176\u5b83\u7684\u5355\u53e5\u8bdd\u8bed\u610f\u56fe\u6807\u7b7e\u6807\u8bb0\u4e3a\"\u5176\u5b83\"\u3002\u4e0d\u540c\u4f2a\u6807 \u7b7e\u96c6\u7684\u5b9e\u9a8c\u6548\u679c\u89c14.5\u5c0f\u8282\u3002 4 \u5b9e \u5b9e \u5b9e\u9a8c \u9a8c \u9a8c 4.1 \u6570 \u6570 \u6570\u636e \u636e \u636e\u96c6 \u96c6 \u96c6 \u672c\u6587\u7684\u4eba\u673a\u5bf9\u8bdd\u5b9e\u9a8c\u6570\u636e\u6765\u81ea\u4e8e\u4e2d\u56fd\u4e2d\u6587\u4fe1\u606f\u5b66\u4f1a\u4e3b\u529e\u7684\"\u5ba2\u670d\u9886\u57df\u7528\u6237\u610f\u56fe\u5206\u7c7b\u8bc4\u6d4b\u6bd4 \u8d5b\" 1 \uff0c\u5c5e\u4e8e\u5ba2\u670d\u9886\u57df\u5bf9\u8bdd\u6587\u672c\uff0c\u89c6\u4e3a\u591a\u8f6e\u8bdd\u7684\u957f\u6587\u672c\u3002\u540c\u65f6\u8fd9\u9879\u6bd4\u8d5b\u8fd8\u6709\u4e00\u4e2a\u5b50\u4efb\u52a1\u662f\u5355\u53e5\u7528 \u6237\u8bdd\u8bed\u7684\u81ea\u7136\u8bed\u8a00\u7406\u89e3(\u5305\u62ec\u610f\u56fe\u8bc6\u522b\u548c\u69fd\u586b\u5145)\u3002\u672c\u6587\u5c06\u5176\u4e2d\u7684\u8bdd\u8bed\u610f\u56fe\u8bc6\u522b\u4efb\u52a1\u89c6\u4e3a\u672c\u6587\u4eba \u673a\u5bf9\u8bdd\u4efb\u52a1\u7684\u8f85\u52a9\u4efb\u52a1\uff0c\u7528\u4e8e\u4ea7\u751f\u4eba\u673a\u5bf9\u8bdd\u6bb5\u4e2d\u6bcf\u4e00\u53e5\u5355\u53e5\u8bdd\u8bed\u7684\u610f\u56fe\u4f2a\u6807\u7b7e\u3002 \u4eba \u4eba \u4eba\u673a \u673a \u673a\u5bf9 \u5bf9 \u5bf9\u8bdd \u8bdd \u8bdd\u6570 \u6570 \u6570\u636e \u636e \u636e\u96c6 \u96c6 \u96c6\u4eba\u673a\u5bf9\u8bdd\u6570\u636e\u96c6\u4e3a2\u4e07\u6761\u771f\u5b9e\u5ba2\u670d\u5bf9\u8bdd\u6bb5\u6807\u6ce8\u6570\u636e\uff0c\u6b64\u6570\u636e\u96c6\u4e2d\u5171\u670935\u79cd \u4eba\u673a\u5bf9\u8bdd\u610f\u56fe\u7c7b\u522b\uff0c\u88682\u7ed9\u51fa\u4e86\u4e1a\u52a1\u7c7b\u578b\u4e0e\u7528\u6237\u610f\u56fe\u7684\u79cd\u7c7b\u3002\u6211\u4eec\u6309\u71678\uff1a2\u7684\u6bd4\u4f8b\u5206\u522b\u5212\u5206\u8bad\u7ec3 \u96c6\u548c\u6d4b\u8bd5\u96c6\u3002\u8bad\u7ec3\u65f6\uff0c\u518d\u4ece\u8bad\u7ec3\u96c6\u4e2d\u5212\u5206\u51fa20%\u4f5c\u4e3a\u9a8c\u8bc1\u96c6\u3002 \u5355 \u5355 \u5355\u53e5 \u53e5 \u53e5\u8bdd \u8bdd \u8bdd\u8bed \u8bed \u8bed\u610f \u610f \u610f\u56fe \u56fe \u56fe\u8bc6 \u8bc6 \u8bc6\u522b \u522b \u522b\u6570 \u6570 \u6570\u636e \u636e \u636e\u96c6 \u96c6 \u96c6 \u5355\u53e5\u8bdd\u8bed\u7684\u6570\u636e\u96c6\u4e3a2\u4e07\u6761\u771f\u5b9e\u7684\u5355\u53e5\u8bdd\u8bed\u6570\u636e\uff0c\u6b64\u6570\u636e\u96c6\u4e2d\u5171 \u670948\u79cd\u5355\u53e5\u610f\u56fe\u7c7b\u522b\uff0c\u5982\u88683\u6240\u793a\u3002\u7ecf\u8fc7\u4f2a\u6807\u7b7e\u4f18\u9009\uff0c\u9009\u53d6\u4e86\u4e0e\u5bf9\u8bdd\u6bb5\u610f\u56fe\u5206\u7c7b\u4efb\u52a1\u76f8\u5173\u5ea6\u9ad8 \u768425\u79cd\u5355\u53e5\u610f\u56fe\u7c7b\u522b\uff0c\u5176\u5b83\u7c7b\u522b\u7684\u4f2a\u6807\u7b7e\u6807\u8bb0\u4e3a\"\u5176\u5b83\"\u3002\u6309\u71678\uff1a2\u5207\u5206\u8bad\u7ec3\u96c6\u548c\u9a8c\u8bc1\u96c6\uff0c\u4e0d\u8bbe \u7f6e\u6d4b\u8bd5\u96c6\u3002 \u4e1a \u4e1a \u4e1a\u52a1 \u52a1 \u52a1\u7c7b \u7c7b \u7c7b\u578b \u578b \u578b \u5bf9 \u5bf9 \u5bf9\u8bdd \u8bdd \u8bdd\u610f \u610f \u610f\u56fe \u56fe \u56fe \u54a8\u8be2(\u542b\u67e5\u8be2) \u4e1a\u52a1\u8ba2\u8d2d\u4fe1\u606f\u67e5\u8be2\u3001\u4e1a\u52a1\u89c4\u5b9a\u3001\u4e1a\u52a1\u8ba2\u8d2d\u4fe1\u606f\u67e5\u8be2\u3001\u4e1a\u52a1\u8d44\u8d39\u3001 \u4ea7\u54c1/\u4e1a\u52a1\u529f\u80fd\u3001\u4f7f\u7528\u65b9\u5f0f\u3001\u529e\u7406\u65b9\u5f0f\u3001\u53f7\u7801\u72b6\u6001\u3001\u5bbd\u5e26\u8986\u76d6\u8303 \u56f4\u3001\u5de5\u5355\u5904\u7406\u7ed3\u679c\u3001\u5de5\u5355\u5904\u7406\u7ed3\u679c\u3001\u670d\u52a1\u6e20\u9053\u4fe1\u606f\u3001\u7528\u6237\u8d44\u6599\u3001 \u7535\u5546\u8d27\u54c1\u4fe1\u606f\u3001\u8425\u9500\u6d3b\u52a8\u4fe1\u606f\u3001\u8d26\u6237\u4fe1\u606f \u529e\u7406 \u4e0b\u8f7d/\u8bbe\u7f6e\u3001\u505c\u590d\u673a\u3001\u53d6\u6d88\u3001\u53d8\u66f4\u3001\u5f00\u901a\u3001\u6253\u5370/\u90ae\u5bc4\u3001\u79fb\u673a/\u88c5 \u673a/\u62c6\u673a\u3001\u7f34\u8d39\u3001\u8865\u6362\u5361\u3001\u91cd\u7f6e/\u4fee\u6539/\u8865\u53d1\u3001\u9500\u6237/\u91cd\u5f00 \u6295\u8bc9(\u542b\u62b1\u6028) \u4e0d\u77e5\u60c5\u5b9a\u5236\u95ee\u9898\u3001\u4e1a\u52a1\u4f7f\u7528\u95ee\u9898\u3001\u4e1a\u52a1\u529e\u7406\u95ee\u9898\u3001\u4e1a\u52a1\u89c4\u5b9a\u4e0d\u6ee1\u3001 \u4fe1\u606f\u5b89\u5168\u95ee\u9898\u3001\u670d\u52a1\u95ee\u9898\u3001\u7f51\u7edc\u95ee\u9898\u3001\u8425\u9500\u95ee\u9898\u3001\u8d39\u7528\u95ee\u9898",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            },
            {
                "text": "4.2 \u5b9e \u5b9e \u5b9e\u9a8c \u9a8c \u9a8c\u8bbe \u8bbe \u8bbe\u7f6e \u7f6e \u7f6e \u4eba \u4eba \u4eba \u673a \u673a \u673a \u5bf9 \u5bf9 \u5bf9 \u8bdd \u8bdd \u8bdd \u610f \u610f \u610f \u56fe \u56fe \u56fe \u5206 \u5206 \u5206 \u7c7b \u7c7b \u7c7b \u5b9e \u5b9e \u5b9e \u9a8c \u9a8c \u9a8c batch size\u8bbe \u7f6e \u4e3a16\uff0cepoch\u8bbe \u7f6e \u4e3a20\uff0cDropout\u8bbe \u7f6e \u4e3a0.1\uff0c \u91c7 \u7528Adam\u4f18\u5316\u5668\uff0c\u5b66\u4e60\u7387\u4e3a1e-5\uff0c\u975e\u5c42\u7ea7\u6a21\u578b\u8bbe\u7f6e\u6700\u5927\u53e5\u5b50\u957f\u5ea6\u4e3a512\uff0c\u5c42\u7ea7\u6a21\u578b\u8bbe\u7f6e\u6700\u5927\u53e5 \u5b50\u6570\u91cf\u4e3a25\uff0c\u6700\u5927\u53e5\u957f\u4e3a25\u3002 \u8bdd \u8bdd \u8bdd \u8bed \u8bed \u8bed \u4f2a \u4f2a \u4f2a \u6807 \u6807 \u6807 \u7b7e \u7b7e \u7b7e \u9884 \u9884 \u9884 \u6d4b \u6d4b \u6d4b \u5b9e \u5b9e \u5b9e \u9a8c \u9a8c \u9a8c batch size\u8bbe \u7f6e \u4e3a32\uff0cepoch\u8bbe \u7f6e \u4e3a100\uff0cDropout\u8bbe \u7f6e \u4e3a0.1\uff0c \u91c7 \u7528Adam\u4f18\u5316\u5668\uff0c\u5b66\u4e60\u7387\u4e3a1e-3\uff0c\u975e\u5c42\u7ea7\u6a21\u578b\u8bbe\u7f6e\u6700\u5927\u53e5\u5b50\u957f\u5ea6\u4e3a30\u3002 4.3 \u5bf9 \u5bf9 \u5bf9\u6bd4 \u6bd4 \u6bd4\u65b9 \u65b9 \u65b9\u6cd5 \u6cd5 \u6cd5",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            },
            {
                "text": "\u672c\u6587\u63d0\u51fa\u7684PLA-HAN\u6a21\u578b\u5c06\u4e0e\u4ee5\u4e0b\u4ee3\u8868\u6027\u7684\u57fa\u7ebf\u65b9\u6848\u8fdb\u884c\u6bd4\u8f83\uff0c\u4e3a\u4e86\u516c\u5e73\u6bd4\u8f83\uff0c\u5168\u90e8\u6a21 \u578b\u90fd\u91c7\u7528\u4e86BERT\u7f16\u7801\uff1a -BERT FineTune\uff1a\u8be5\u65b9\u6cd5\u5c06BERT fine-tuning (Devlin et al., 2019) \u5e94\u7528\u5230\u5206\u7c7b\u4efb\u52a1\uff0c \u5728BERT\u5206\u7c7b\u5c42\u589e\u52a0\u4e86\u4e00\u4e2a\u65b0\u7684\u8f93\u51fa\u5c42\u3002 -BERT BiLSTM\uff1a\u8be5\u65b9\u6cd5\u662f\u6587\u672c\u5206\u7c7b\u7684\u7ecf\u5178\u57fa\u7ebf (Vu et al., 2016 )\uff0c\u9002\u5408\u4e8e\u5e8f\u5217\u95ee\u9898\u3002 -BERT SoftAtt\uff1a\u8be5\u65b9\u6cd5\u91c7\u7528\u4e86Liu\u7b49 (Liu and Lane, 2016) \u5728ATIS\u6570\u636e\u96c6\u7684\u8bdd\u8bed\u610f\u56fe\u8bc6\u522b \u7684BILSTM\u6a21\u578b\u4e2d\u91c7\u7528\u7684\u8f6f\u6ce8\u610f\u529b\u3002 -BERT HAN\uff1aYang\u7b49 (Yang et al., 2016) ",
                "cite_spans": [
                    {
                        "start": 91,
                        "end": 112,
                        "text": "(Devlin et al., 2019)",
                        "ref_id": "BIBREF5"
                    },
                    {
                        "start": 169,
                        "end": 185,
                        "text": "(Vu et al., 2016",
                        "ref_id": "BIBREF14"
                    },
                    {
                        "start": 222,
                        "end": 242,
                        "text": "(Liu and Lane, 2016)",
                        "ref_id": "BIBREF8"
                    },
                    {
                        "start": 294,
                        "end": 313,
                        "text": "(Yang et al., 2016)",
                        "ref_id": "BIBREF18"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            },
            {
                "text": "\u63d0\u51fa\u7684\u66f4\u52a0\u9002\u5408\u7bc7\u7ae0\u7ed3\u6784\u6587\u672c\u7684\u5c42\u6b21\u6ce8\u610f\u529b\u6a21\u578b \u7ed3\u6784\u3002 4.4 \u6574 \u6574 \u6574\u4f53 \u4f53 \u4f53\u6027 \u6027 \u6027\u80fd \u80fd \u80fd \u6211\u4eec\u63d0\u51fa\u7684PLA-HAN\u6a21\u578b\u4e0e\u51e0\u79cd\u5bf9\u6bd4\u65b9\u6cd5\u8fdb\u884c\u6bd4\u8f83\uff0c\u5305\u62ecBERT FineTune\u3001BiLSTM\u3001 \u5e26\u8f6f\u6ce8\u610f\u529b\u7684BiLSTM\u4ee5\u53ca\u5c42\u6b21\u6ce8\u610f\u529b\u6a21\u578bHAN\u3002\u5b9e\u9a8c\u7ed3\u679c\u5982\u88684\u6240\u793a\u3002 \u6a21 \u6a21 \u6a21\u578b \u578b \u578b\u7c7b \u7c7b \u7c7b\u522b \u522b \u522b \u6a21 \u6a21 \u6a21\u578b \u578b \u578b \u610f \u610f \u610f\u56fe \u56fe \u56fe\u5206 \u5206 \u5206\u7c7b \u7c7b \u7c7b\u6b63 \u6b63 \u6b63\u786e \u786e \u786e\u7387 \u7387 \u7387(%) \u603b \u603b \u603b\u4f53 \u4f53 \u4f53 \u54a8 \u54a8 \u54a8\u8be2 \u8be2 \u8be2(\u542b \u542b \u542b\u67e5 \u67e5 \u67e5\u8be2 \u8be2 \u8be2) \u529e \u529e \u529e\u7406 \u7406 \u7406 \u6295 \u6295 \u6295\u8bc9 \u8bc9 \u8bc9 \u957f\u6587\u672c\u7ed3\u6784 BERT",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            }
        ],
        "back_matter": [],
        "bib_entries": {
            "BIBREF2": {
                "ref_id": "b2",
                "title": "\u4efb\u52a1\u578b\u4eba\u673a\u5bf9\u8bdd\u7cfb\u7edf\u4e2d\u7684\u8ba4\u77e5\u6280\u672f--\u6982\u5ff5\u3001\u8fdb\u5c55\u53ca\u5176\u672a\u6765. \u8ba1\u7b97\u673a\u5b66\u62a5",
                "authors": [
                    {
                        "first": "\u9648\u9732",
                        "middle": [],
                        "last": "\u4fde\u51ef",
                        "suffix": ""
                    },
                    {
                        "first": "\u9648\u535a",
                        "middle": [],
                        "last": "",
                        "suffix": ""
                    },
                    {
                        "first": "\u7b49",
                        "middle": [],
                        "last": "",
                        "suffix": ""
                    }
                ],
                "year": 2015,
                "venue": "",
                "volume": "38",
                "issue": "",
                "pages": "2333--2348",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "\u4fde\u51ef, \u9648\u9732, \u9648\u535a, \u7b49. 2015. \u4efb\u52a1\u578b\u4eba\u673a\u5bf9\u8bdd\u7cfb\u7edf\u4e2d\u7684\u8ba4\u77e5\u6280\u672f--\u6982\u5ff5\u3001\u8fdb\u5c55\u53ca\u5176\u672a\u6765. \u8ba1\u7b97\u673a\u5b66\u62a5, 38 (12):2333-2348.",
                "links": null
            },
            "BIBREF3": {
                "ref_id": "b3",
                "title": "A survey on dialogue systems: Recent advances and new frontiers",
                "authors": [
                    {
                        "first": "Hongshen",
                        "middle": [],
                        "last": "Chen",
                        "suffix": ""
                    },
                    {
                        "first": "Xiaorui",
                        "middle": [],
                        "last": "Liu",
                        "suffix": ""
                    },
                    {
                        "first": "Dawei",
                        "middle": [],
                        "last": "Yin",
                        "suffix": ""
                    },
                    {
                        "first": "Jiliang",
                        "middle": [],
                        "last": "Tang",
                        "suffix": ""
                    }
                ],
                "year": 2017,
                "venue": "",
                "volume": "19",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Chen, Hongshen and Liu, Xiaorui and Yin, Dawei and Tang, Jiliang. 2017. A survey on dialogue systems: Recent advances and new frontiers, volume 19. ACM New York, NY, USA.",
                "links": null
            },
            "BIBREF4": {
                "ref_id": "b4",
                "title": "Long short-term memory-networks for machine reading",
                "authors": [
                    {
                        "first": "J",
                        "middle": [
                            "P"
                        ],
                        "last": "Cheng",
                        "suffix": ""
                    },
                    {
                        "first": "L",
                        "middle": [],
                        "last": "Dong",
                        "suffix": ""
                    },
                    {
                        "first": "M",
                        "middle": [],
                        "last": "Lapata",
                        "suffix": ""
                    }
                ],
                "year": 2016,
                "venue": "Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing",
                "volume": "",
                "issue": "",
                "pages": "551--561",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "J. P. Cheng, L. Dong, and M. Lapata. 2016. Long short-term memory-networks for machine reading. Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing (EMNLP 2016), pp. 551-561.",
                "links": null
            },
            "BIBREF5": {
                "ref_id": "b5",
                "title": "BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding",
                "authors": [
                    {
                        "first": "J",
                        "middle": [],
                        "last": "Devlin",
                        "suffix": ""
                    },
                    {
                        "first": "M",
                        "middle": [],
                        "last": "Chang",
                        "suffix": ""
                    },
                    {
                        "first": "K",
                        "middle": [],
                        "last": "Lee",
                        "suffix": ""
                    }
                ],
                "year": 2019,
                "venue": "Proceedings of the 17th Annual Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (NAACL-HLT 2019)",
                "volume": "",
                "issue": "",
                "pages": "4171--4186",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "J. Devlin, M. Chang, K. Lee, et al. 2019. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. Proceedings of the 17th Annual Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (NAACL-HLT 2019), pp. 4171-4186.",
                "links": null
            },
            "BIBREF6": {
                "ref_id": "b6",
                "title": "Hybrid speech recognition with deep bidirectional LSTM",
                "authors": [
                    {
                        "first": "A",
                        "middle": [],
                        "last": "Graves",
                        "suffix": ""
                    },
                    {
                        "first": "N",
                        "middle": [],
                        "last": "Jaitly",
                        "suffix": ""
                    },
                    {
                        "first": "A",
                        "middle": [],
                        "last": "Mohamed",
                        "suffix": ""
                    }
                ],
                "year": 2013,
                "venue": "Proceedings of the 2013 IEEE workshop on automatic speech recognition and understanding",
                "volume": "",
                "issue": "",
                "pages": "273--278",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "A. Graves, N.Jaitly and A. Mohamed. 2013. Hybrid speech recognition with deep bidirectional LSTM. Proceedings of the 2013 IEEE workshop on automatic speech recognition and understanding (ASRU 2013), 273-278.",
                "links": null
            },
            "BIBREF7": {
                "ref_id": "b7",
                "title": "Optimizing SVMs for complex call classification",
                "authors": [
                    {
                        "first": "P",
                        "middle": [],
                        "last": "Haffner",
                        "suffix": ""
                    },
                    {
                        "first": "G",
                        "middle": [],
                        "last": "Tur",
                        "suffix": ""
                    },
                    {
                        "first": "J",
                        "middle": [
                            "H"
                        ],
                        "last": "Wright",
                        "suffix": ""
                    }
                ],
                "year": 2003,
                "venue": "Proceedings of the 28th International Conference on Acoustics, Speech, and Signal Processing",
                "volume": "",
                "issue": "",
                "pages": "632--635",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "P. Haffner, G. Tur, and J. H. Wright. 2003. Optimizing SVMs for complex call classification. Proceedings of the 28th International Conference on Acoustics, Speech, and Signal Processing (ICASSP 2003), pp. 632-635.",
                "links": null
            },
            "BIBREF8": {
                "ref_id": "b8",
                "title": "Attention-based recurrent neural network models for joint intent detection and slot filling",
                "authors": [
                    {
                        "first": "B",
                        "middle": [],
                        "last": "Liu",
                        "suffix": ""
                    },
                    {
                        "first": "T",
                        "middle": [],
                        "last": "Lane",
                        "suffix": ""
                    }
                ],
                "year": 2016,
                "venue": "Proceedings of the 17th Annual Conference of the Inter-national Speech Communication Association (INTERSPEECH 2016)",
                "volume": "",
                "issue": "",
                "pages": "685--689",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "B. Liu and T. Lane. 2016. Attention-based recurrent neural network models for joint intent detection and slot filling. Proceedings of the 17th Annual Conference of the Inter-national Speech Communication Association (INTERSPEECH 2016), pp. 685-689.",
                "links": null
            },
            "BIBREF9": {
                "ref_id": "b9",
                "title": "Convolutional Neural Networks for Sentence Classification",
                "authors": [
                    {
                        "first": "Y",
                        "middle": [],
                        "last": "Kim",
                        "suffix": ""
                    }
                ],
                "year": 2014,
                "venue": "Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing",
                "volume": "",
                "issue": "",
                "pages": "1746--1751",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Y. Kim. 2014. Convolutional Neural Networks for Sentence Classification. Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP 2014), pp. 1746-1751.",
                "links": null
            },
            "BIBREF10": {
                "ref_id": "b10",
                "title": "Efficient large-scale neural domain classification with personalized attention",
                "authors": [
                    {
                        "first": "Y",
                        "middle": [],
                        "last": "Kim",
                        "suffix": ""
                    },
                    {
                        "first": "D",
                        "middle": [],
                        "last": "Kim",
                        "suffix": ""
                    },
                    {
                        "first": "A",
                        "middle": [],
                        "last": "Kumar",
                        "suffix": ""
                    }
                ],
                "year": 2018,
                "venue": "Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (ACL 2018)",
                "volume": "",
                "issue": "",
                "pages": "2214--2224",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Y. Kim, D. Kim, A. Kumar. 2018. Efficient large-scale neural domain classification with personalized attention. Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (ACL 2018), pp. 2214-2224.",
                "links": null
            },
            "BIBREF11": {
                "ref_id": "b11",
                "title": "Learning to classify short and sparse text & web with hidden topics from largescale data collections",
                "authors": [
                    {
                        "first": "X",
                        "middle": [
                            "H"
                        ],
                        "last": "Phan",
                        "suffix": ""
                    },
                    {
                        "first": "L",
                        "middle": [
                            "M"
                        ],
                        "last": "Nguyen",
                        "suffix": ""
                    },
                    {
                        "first": "S",
                        "middle": [],
                        "last": "Horiguchi",
                        "suffix": ""
                    }
                ],
                "year": 2008,
                "venue": "Proceedings of the 17th International Conference on World Wide Web",
                "volume": "",
                "issue": "",
                "pages": "91--100",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "X .H. Phan, L. M. Nguyen, S. Horiguchi. 2008. Learning to classify short and sparse text & web with hidden topics from largescale data collections. Proceedings of the 17th International Conference on World Wide Web (WWW 2008), pp. 91-100.",
                "links": null
            },
            "BIBREF12": {
                "ref_id": "b12",
                "title": "A comparative study of recurrent neural network models for lexical domain classification",
                "authors": [
                    {
                        "first": "S",
                        "middle": [],
                        "last": "Ravuri",
                        "suffix": ""
                    },
                    {
                        "first": "A",
                        "middle": [],
                        "last": "Stolcke",
                        "suffix": ""
                    }
                ],
                "year": 2016,
                "venue": "Proceedings of the 41th IEEE International Conference on Acoustics, Speech, and Signal Processing",
                "volume": "",
                "issue": "",
                "pages": "6075--6079",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "S. Ravuri and A. Stolcke. 2016. A comparative study of recurrent neural network models for lexical domain classification. Proceedings of the 41th IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP 2016), pp. 6075-6079.",
                "links": null
            },
            "BIBREF13": {
                "ref_id": "b13",
                "title": "Deep belief nets for natural language call-routing",
                "authors": [
                    {
                        "first": "R",
                        "middle": [],
                        "last": "Sarikaya",
                        "suffix": ""
                    },
                    {
                        "first": "G",
                        "middle": [
                            "E"
                        ],
                        "last": "Hinton",
                        "suffix": ""
                    },
                    {
                        "first": "B",
                        "middle": [],
                        "last": "Ramabhadran",
                        "suffix": ""
                    }
                ],
                "year": 2011,
                "venue": "Proceedings of the 36th IEEE International Conference on Acoustics, Speech, and Signal Processing",
                "volume": "",
                "issue": "",
                "pages": "5680--5683",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "R. Sarikaya, G. E. Hinton, and B. Ramabhadran. 2011. Deep belief nets for natural language call-routing. Proceedings of the 36th IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP 2011), pp. 5680-5683.",
                "links": null
            },
            "BIBREF14": {
                "ref_id": "b14",
                "title": "Bi-directional recurrent neural network with ranking loss for spoken language understanding",
                "authors": [
                    {
                        "first": "N",
                        "middle": [
                            "T"
                        ],
                        "last": "Vu",
                        "suffix": ""
                    },
                    {
                        "first": "P",
                        "middle": [],
                        "last": "Gupta",
                        "suffix": ""
                    },
                    {
                        "first": "H",
                        "middle": [],
                        "last": "Adel",
                        "suffix": ""
                    }
                ],
                "year": 2016,
                "venue": "Proceedings of the 41th IEEE International Conference on Acoustics, Speech and Signal Processing",
                "volume": "",
                "issue": "",
                "pages": "6060--6064",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "N. T. Vu, P. Gupta, H. Adel, et al. 2016. Bi-directional recurrent neural network with ranking loss for spoken language understanding. Proceedings of the 41th IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP 2016), pp. 6060-6064.",
                "links": null
            },
            "BIBREF15": {
                "ref_id": "b15",
                "title": "Convolutional neural network based triangular CRF for joint intent detection and slot filling",
                "authors": [
                    {
                        "first": "P",
                        "middle": [
                            "Y"
                        ],
                        "last": "Xu",
                        "suffix": ""
                    },
                    {
                        "first": "R",
                        "middle": [],
                        "last": "Sarikaya",
                        "suffix": ""
                    }
                ],
                "year": 2013,
                "venue": "Proceedings of the 2013 IEEE Workshop on Automatic Speech Recognition and Understanding (ASRU 2013)",
                "volume": "",
                "issue": "",
                "pages": "78--83",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "P. Y. Xu and R. Sarikaya. 2013. Convolutional neural network based triangular CRF for joint intent detection and slot filling. Proceedings of the 2013 IEEE Workshop on Automatic Speech Recognition and Understanding (ASRU 2013), pp. 78-83.",
                "links": null
            },
            "BIBREF16": {
                "ref_id": "b16",
                "title": "Contextual domain classification in spoken language understanding systems using recurrent neural network",
                "authors": [
                    {
                        "first": "P",
                        "middle": [
                            "Y"
                        ],
                        "last": "Xu",
                        "suffix": ""
                    },
                    {
                        "first": "R",
                        "middle": [],
                        "last": "Sarikaya",
                        "suffix": ""
                    }
                ],
                "year": 2014,
                "venue": "Proceedings of the 39th International Conference on Acoustics, Speech and Signal Processing",
                "volume": "",
                "issue": "",
                "pages": "136--140",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "P. Y. Xu and R. Sarikaya. 2014. Contextual domain classification in spoken language understanding sys- tems using recurrent neural network. Proceedings of the 39th International Conference on Acoustics, Speech and Signal Processing (ICASSP 2014), pp. 136-140.",
                "links": null
            },
            "BIBREF17": {
                "ref_id": "b17",
                "title": "Show, attend and tell: Neural image caption generation with visual attention",
                "authors": [
                    {
                        "first": "K",
                        "middle": [],
                        "last": "Xu",
                        "suffix": ""
                    },
                    {
                        "first": "J",
                        "middle": [],
                        "last": "Ba",
                        "suffix": ""
                    },
                    {
                        "first": "R",
                        "middle": [],
                        "last": "Kiros",
                        "suffix": ""
                    }
                ],
                "year": 2015,
                "venue": "Proceedings of the 31st International Conference on Machine Learning",
                "volume": "",
                "issue": "",
                "pages": "2048--2057",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "K. Xu, J. Ba and R. Kiros,et al. 2015. Show, attend and tell: Neural image caption generation with visual attention. Proceedings of the 31st International Conference on Machine Learning (ICML 2015), pp. 2048-2057.",
                "links": null
            },
            "BIBREF18": {
                "ref_id": "b18",
                "title": "Hierarchical attention networks for document classification",
                "authors": [
                    {
                        "first": "Z",
                        "middle": [
                            "C"
                        ],
                        "last": "Yang",
                        "suffix": ""
                    },
                    {
                        "first": "D",
                        "middle": [
                            "Y"
                        ],
                        "last": "Yang",
                        "suffix": ""
                    },
                    {
                        "first": "C",
                        "middle": [],
                        "last": "Dyer",
                        "suffix": ""
                    }
                ],
                "year": 2016,
                "venue": "Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics (NAACL 2016)",
                "volume": "",
                "issue": "",
                "pages": "1480--1489",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Z. C. Yang, D. Y. Yang, C. Dyer, et al. 2016. Hierarchical attention networks for document classifi- cation. Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics (NAACL 2016), pp. 1480-1489.",
                "links": null
            }
        },
        "ref_entries": {
            "FIGREF0": {
                "num": null,
                "text": "\u4e0d\u540c\u957f\u5ea6\u5bf9\u8bdd\u6bb5\u7684\u6a21\u578b\u6027\u80fd\u5bf9\u6bd4\u53c2 \u53c2 \u53c2\u8003 \u8003 \u8003\u6587 \u6587 \u6587\u732e \u732e \u732e",
                "uris": null,
                "type_str": "figure"
            },
            "TABREF1": {
                "html": null,
                "content": "<table/>",
                "type_str": "table",
                "num": null,
                "text": "\u4e00\u4e2a\u610f\u56fe\u5206\u7c7b\u4f8b\u5b50\u4eba\u673a\u5bf9\u8bdd\u610f\u56fe\u5206\u7c7b\u5c5e\u4e8e\u6587\u672c\u5206\u7c7b\u4efb\u52a1\u3002\u5df2\u7ecf\u6709\u5f88\u591a\u6587\u672c\u5206\u7c7b\u7684\u7814\u7a76\u81f4\u529b\u4e8e\u63d0\u9ad8\u5206\u7c7b\u5668\u6027 \u80fd\uff0c\u65e9\u671f\u7684\u5178\u578b\u4ee3\u8868\u662fSVM\u548c\u6700\u5927\u71b5\u7b49"
            },
            "TABREF5": {
                "html": null,
                "content": "<table><tr><td>\u67e5\u8be2\u3001\u67e5\u8be2\u5957\u9910\u4f59\u91cf\u3001\u67e5\u8be2\u5145\u503c\u7f34\u8d39\u8bb0\u5f55\u3001\u67e5\u8be2\u6d41\u91cf\u3001\u67e5\u8be2\u672c\u673a\u53f7\u7801\u3001\u67e5\u8be2\u672c\u673a\u4e1a\u52a1\u3001</td></tr><tr><td>\u67e5\u8be2\u4f59\u989d\u3001\u67e5\u8be2\u77ed\u4fe1\u3001\u67e5\u8be2\u79ef\u5206\u3001\u67e5\u8be2\u8bed\u97f3\u3001\u67e5\u8be2\u6708\u521d\u6263\u8d39\u3001\u67e5\u8be2\u8d26\u5355\u3001\u67e5\u8be2\u5bbd\u5e26\u3001\u67e5</td></tr><tr><td>\u8be2\u624b\u673a\u3001\u54a8\u8be2\u5bbd\u5e26\u3001\u5177\u5b9e\u5e2e\u52a9\u3001\u5177\u5b9e\u8fd4\u56de\u3001\u5177\u5b9e\u8f6c\u4eba\u5de5\u3001\u5177\u5b9e\u9000\u51fa\u3001\u5177\u5b9e\u91cd\u542c\u3001\u5177\u5b9e\u4e1a</td></tr><tr><td>\u52a1\u5217\u8868\u3001\u5177\u5b9e\u8f6civr\u3001\u529e\u7406\u5957\u9910\u3001\u529e\u7406\u624b\u673a\u5145\u503c\u3001\u53d6\u6d88\u6d41\u91cf\u3001\u9884\u7ea6\u5bbd\u5e26\u3001\u4fee\u6539\u5bbd\u5e26\u3001\u5f00\u901a</td></tr><tr><td>\u6d41\u91cf\u3001\u91cd\u7f6e\u3001\u91cd\u7f6e\u624b\u673a\u3001\u91cd\u7f6e\u5bbd\u5e26\u3001GPRS\u3001\u62d2\u8bc6\u3001\u4fee\u6539\u3001\u786e\u8ba4\u3001\u624b\u673a\u3001\u4fee\u6539\u624b\u673a\u3001\u77ed</td></tr><tr><td>\u4fe1\u3001\u5bbd\u5e26\u3001\u6d41\u91cf\u3001empty\u3001\u96c6\u5916\u8bf4\u6cd5\u6279\u8bc4,\u62b1\u6028,\u810f\u8bdd\u3001\u96c6\u5916\u8bf4\u6cd5\u673a\u5668\u4eba\u3001\u96c6\u5916\u8bf4\u6cd5\u96c6\u5916\u4e1a</td></tr><tr><td>\u52a1\u3001\u54a8\u8be2\u3001\u96c6\u5916\u8bf4\u6cd5\u7ed3\u675f\u3001\u96c6\u5916\u8bf4\u6cd5\u77ed\u62d2\u8bc6\u3001\u96c6\u5916\u8bf4\u6cd5\u53cb\u597d\u95ee\u5019\u3001\u96c6\u5916\u8bf4\u6cd5\u611f\u8c22</td></tr></table>",
                "type_str": "table",
                "num": null,
                "text": "\u4e1a\u52a1\u7c7b\u578b\u79cd\u7c7b\u4e0e35\u79cd\u5bf9\u8bdd\u610f\u56fe\u7c7b\u522b\u5355 \u5355 \u5355\u53e5 \u53e5 \u53e5\u8bdd \u8bdd \u8bdd\u8bed \u8bed \u8bed\u610f \u610f \u610f\u56fe \u56fe \u56fe\u6807 \u6807 \u6807\u7b7e \u7b7e \u7b7e"
            },
            "TABREF6": {
                "html": null,
                "content": "<table/>",
                "type_str": "table",
                "num": null,
                "text": "48\u79cd\u5355\u53e5\u8bdd\u8bed\u610f\u56fe\u7c7b\u522b"
            },
            "TABREF8": {
                "html": null,
                "content": "<table><tr><td>: \u4e0d\u540c\u6a21\u578b\u6027\u80fd\u5bf9\u6bd4</td></tr><tr><td>\u4ece\u88684\u7684\u7ed3\u679c\u53ef\u4ee5\u770b\u5230\uff1a</td></tr><tr><td>(1) \u5728\u4ee5\u6574\u7bc7\u957f\u6587\u672c\u76f4\u63a5\u4f5c\u4e3a\u8f93\u5165\u7684\u6a21\u578b\u4e2d\uff0c\u901a\u8fc7\u52a0\u5165\u7f16\u7801\u5668BiLSTM\u548c\u6ce8\u610f\u529b\u673a\u5236\u7684\u5e94</td></tr><tr><td>\u7528\uff0c\u53ef\u4ee5\u5728BERT\u7684\u57fa\u7840\u4e0a\u8fdb\u4e00\u6b65\u63d0\u5347\u5bf9\u8bdd\u6bb5\u610f\u56fe\u5206\u7c7b\u7684\u6027\u80fd\u3002</td></tr><tr><td>(2) \u76f8\u6bd4\u4e8e\u6574\u7bc7\u957f\u6587\u672c\u7ed3\u6784\u7684\u8f93\u5165\uff0c\u5c42\u6b21\u7ed3\u6784\u6a21\u578b\u53d6\u5f97\u4e86\u66f4\u597d\u7684\u5206\u7c7b\u6027\u80fd\u3002\u8fd9\u4e00\u65b9\u9762\u5f97\u76ca</td></tr><tr><td>\u4e8e\u4e0d\u53d7BERT\u8f93\u5165\u957f\u5ea6\u7684\u9650\u5236\uff0c\u53e6\u4e00\u65b9\u9762\u4e5f\u7531\u4e8e\u5206\u5c42\u6b21\u7684\u6ce8\u610f\u529b\u80fd\u66f4\u597d\u5730\u5efa\u6a21\u5b57\u5230\u53e5\u5b50\u518d\u5230\u5bf9</td></tr><tr><td>\u8bdd\u6bb5\u7684\u8bed\u4e49\u7ed3\u6784\u3002</td></tr><tr><td>(3) \u878d\u5408\u4e86\u5355\u53e5\u8bdd\u8bed\u4f2a\u6807\u7b7e\u6ce8\u610f\u529b\u7684PLA-HAN\u53d6\u5f97\u4e86\u6700\u597d\u7684\u6027\u80fd\uff0c\u4f18\u4e8eHAN\u6a21\u578b\u3002</td></tr></table>",
                "type_str": "table",
                "num": null,
                "text": "\u8fdb \u8fdb \u8fdb\u4e00 \u4e00 \u4e00\u6b65 \u6b65 \u6b65\u5206 \u5206 \u5206\u6790 \u6790 \u6790 \u901a\u8fc7\u6bd4\u8f83\u4ee5\u4e0a\u7ed3\u679c\u53ef\u4ee5\u770b\u51faPLA-HAN\u6a21\u578b\u53d6\u5f97\u4e86\u826f\u597d\u7684\u6027\u80fd\uff0c\u6211\u4eec\u4e5f\u60f3\u8fdb\u4e00\u6b65\u63a2\u7a76\u6a21\u578b \u80fd\u6709\u6240\u63d0\u5347\u7684\u539f\u56e0\u3002\u6211\u4eec\u9996\u5148\u5206\u6790\u4e86\u4e0d\u540c\u7684\u4f2a\u6807\u7b7e\u96c6\u5bf9\u6a21\u578b\u6027\u80fd\u7684\u5f71\u54cd\u3002\u7136\u540e\uff0c\u6211\u4eec\u7ed9\u51fa\u4e86\u4e00 \u4e2a\u4e0d\u540c\u957f\u5ea6\u5bf9\u8bdd\u6bb5\u60c5\u51b5\u4e0b\u7684PLA-HAN\u6a21\u578b\u4e0e\u57fa\u7840\u7684HAN\u6a21\u578b\u7684\u6027\u80fd\u5bf9\u6bd4\u7684\u5b9a\u91cf\u5206\u6790\u3002 \u4e0d \u4e0d \u4e0d\u540c \u540c \u540c\u4f2a \u4f2a \u4f2a\u6807 \u6807 \u6807\u7b7e \u7b7e \u7b7e\u96c6 \u96c6 \u96c6\u5bf9 \u5bf9 \u5bf9\u6a21 \u6a21 \u6a21\u578b \u578b \u578b\u6027 \u6027 \u6027\u80fd \u80fd \u80fd\u7684 \u7684 \u7684\u5f71 \u5f71 \u5f71\u54cd \u54cd \u54cd\u3002\u4e3a\u4e86\u7814\u7a76\u4e0d\u540c\u4f2a\u6807\u7b7e\u96c6\u5728PLA-HAN\u6a21\u578b\u4e2d\u6548\u679c\uff0c\u6211\u4eec"
            },
            "TABREF9": {
                "html": null,
                "content": "<table><tr><td>\u7ed3\u679c\u8868\u660e\uff0c\u6211\u4eec\u7684\u4f2a\u6807\u7b7e\u96c6\u9009\u62e9\u7b56\u7565\u662f\u5fc5\u8981\u7684\uff0c\u4e0d\u540c\u7684\u4f2a\u6807\u7b7e\u96c6\u5bf9\u6a21\u578b\u6574\u4f53\u6027\u80fd\u5b58\u5728\u660e\u663e</td></tr><tr><td>\u7684\u5f71\u54cd\u3002\u8be6\u7ec6\u7684\u5206\u6790\u5982\u4e0b\uff1a</td></tr><tr><td>\u2022 PLA-HAN(All)\uff1a\u5728\u6b64\u65b9\u6848\u4e2d\uff0c\u6211\u4eec\u91c7\u7528\u4e86\u6240\u670948\u79cd\u4f2a\u6807\u7b7e\u3002\u5176\u4e2d\u90e8\u5206\u4f2a\u6807\u7b7e\u5b9e\u9645\u4e0a\u548c\u4eba</td></tr><tr><td>\u673a\u5bf9\u8bdd\u610f\u56fe\u5206\u7c7b\u4efb\u52a1\u4e2d\u768435\u79cd\u610f\u56fe\u7684\u76f8\u5173\u6027\u5e76\u4e0d\u5f3a\u3002\u4ece\u5b9e\u9a8c\u7ed3\u679c\u53ef\u4ee5\u770b\u5230\uff0c\u5206\u7c7b\u6027\u80fd\u90fd\u4e0d</td></tr><tr><td>\u591f\u597d\uff0c\u751a\u81f3\u90fd\u7565\u5fae\u4f4e\u4e8eHAN\u6a21\u578b\u3002</td></tr><tr><td>\u2022 PLA-HAN(Fit)\uff1a\u5728\u6b64\u65b9\u6848\u4e2d\uff0c\u6211\u4eec\u9009\u53d6\u4e0e\u4eba\u673a\u5bf9\u8bdd\u4efb\u52a1\u7684\u610f\u56fe\u6709\u8f83\u9ad8\u8986\u76d6\u7387\u7684\u4f2a\u6807\u7b7e</td></tr><tr><td>\u5d4c\u5165\u5230\u6a21\u578b\u4e2d\uff0c\u5c06\u57fa\u672c\u4e0d\u76f8\u5173\u7684\u6807\u7b7e\u6807\u8bb0\u4e3a\"\u5176\u5b83\"\u3002\u5b9e\u9a8c\u4e2d\u901a\u8fc7\u9a8c\u8bc1\u96c6\u9009\u53d6\u4e8632\u4e2a\u4f2a\u6807</td></tr><tr><td>\u7b7e\u3002\u4ece\u5b9e\u9a8c\u7ed3\u679c\u53ef\u4ee5\u770b\u5230\uff0c\u5206\u7c7b\u6027\u80fd\u76f8\u6bd4\u4e8ePLA-HAN(All)\u6709\u4e86\u660e\u663e\u63d0\u9ad8\uff0c\u76f8\u540c\u65f6\u4e5f\u4f18</td></tr><tr><td>\u4e8eHAN\u6a21\u578b\u3002</td></tr><tr><td>\u2022 PLA-HAN(Select)\uff1a\u5728\u6b64\u65b9\u6848\u4e2d\uff0c\u6211\u4eec\u5728PLA-HAN(Fit)\u7684\u57fa\u7840\u4e0a\uff0c\u8fdb\u4e00\u6b65\u4ece\u8bed\u4e49\u89d2\u5ea6\u6311</td></tr><tr><td>\u9009\u51fa\u9ad8\u5173\u8054\u5ea6\u7684\u4f2a\u6807\u7b7e\u5d4c\u5165\u5230\u6a21\u578b\u4e2d\uff0c\u53bb\u6389\u4e86\u90e8\u5206\u5173\u8054\u5ea6\u4e0d\u9ad8\u7684\u6807\u7b7e\uff0c\u4e00\u5171\u9009\u53d6\u4e8625\u79cd\u4f2a</td></tr><tr><td>\u6807\u7b7e\u3002\u4ece\u5b9e\u9a8c\u7ed3\u679c\u770b\uff0cPLA-HAN(Select)\u53d6\u5f97\u4e86\u6700\u597d\u7684\u6027\u80fd\u3002</td></tr><tr><td>\u4e0d \u4e0d \u578b\u4e0e\u57fa\u7840\u7684HAN\u6a21\u578b\u7684\u6027\u80fd\u3002\u6211\u4eec\u6309\u7167\u5bf9\u8bdd\u6bb5\u7684\u957f\u5ea6\u5206\u4e3a\u957f(600\u5b57\u4ee5\u4e0a)\u3001\u4e2d(301-600\u5b57)\u548c</td></tr><tr><td>\u77ed(300\u5b57\u4ee5\u4e0b)\u4e09\u7c7b\u8fdb\u884c\u89c2\u5bdf\uff0c\u7ed3\u679c\u5982\u56fe2\u6240\u793a\u3002</td></tr><tr><td>\u4ece\u56fe2\u53ef\u4ee5\u770b\u5230\uff1a</td></tr><tr><td>(1) \u957f\u7684\u5bf9\u8bdd\u6bb5\u7684\u610f\u56fe\u5206\u7c7b\u5b58\u5728\u8f83\u5927\u7684\u6311\u6218\uff0c\u6b63\u786e\u7387\u660e\u663e\u4f4e\u4e8e\u77ed\u7684\u5bf9\u8bdd\u6bb5\u3002</td></tr><tr><td>(2) \u6211\u4eec\u7684PLA-HAN\uff0c\u5728\u4f2a\u6807\u7b7e\u6ce8\u610f\u529b\u7684\u5e2e\u52a9\u4e0b\uff0c\u5728\u4e0d\u540c\u957f\u5ea6\u7684\u5bf9\u8bdd\u6bb5\u6027\u80fd\u5747\u4f18\u4e8eHAN\u3002</td></tr><tr><td>\u5c24\u5176\u662f\u957f\u7684\u5bf9\u8bdd\u6bb5(\u8d85\u8fc7600\u5b57)\uff0c\u610f\u56fe\u5206\u7c7b\u6b63\u786e\u7387\u63d0\u5347\u8f83\u4e3a\u663e\u8457\uff0c\u8fbe\u52301.56%\u3002</td></tr><tr><td>5 \u7ed3 \u7ed3 \u7ed3\u675f \u675f \u675f\u8bed \u8bed \u8bed</td></tr><tr><td>\u9488\u5bf9\u73b0\u6709\u6587\u672c\u5206\u7c7b\u65b9\u6cd5\u5728\u4eba\u673a\u5bf9\u8bdd\u610f\u56fe\u5206\u7c7b\u4e0a\u5b58\u5728\u7684\u6311\u6218\uff0c\u672c\u6587\u63d0\u51fa\u4e86\u4e00\u79cd\u7ed3\u5408\u8bdd\u8bed\u4f2a\u6807</td></tr><tr><td>\u7b7e\u6ce8\u610f\u529b\u7684\u5c42\u6b21\u6ce8\u610f\u529b\u7f51\u7edc\u6a21\u578bPLA-HAN\u3002PLA-HAN\u901a\u8fc7\u4f18\u9009\u4f2a\u6807\u7b7e\u96c6\uff0c\u8bbe\u8ba1\u548c\u8ba1\u7b97\u5355\u53e5\u8bdd</td></tr><tr><td>\u8bed\u610f\u56fe\u4f2a\u6807\u7b7e\u6ce8\u610f\u529b\uff0c\u5e76\u5c06\u5176\u5d4c\u5165\u5230HAN\u7684\u5c42\u7ea7\u7ed3\u6784\u4e2d\uff0c\u4e0eHAN\u4e2d\u7684\u53e5\u5b50\u7ea7\u522b\u6ce8\u610f\u529b\u76f8\u878d\u5408\uff0c</td></tr><tr><td>\u63d0\u5347\u4e86\u4eba\u673a\u5bf9\u8bdd\u610f\u56fe\u5206\u7c7b\u6027\u80fd\u3002\u6211\u4eec\u5728\u4e2d\u56fd\u4e2d\u6587\u4fe1\u606f\u5b66\u4f1a\u4e3b\u529e\u7684\"\u5ba2\u670d\u9886\u57df\u7528\u6237\u610f\u56fe\u5206\u7c7b\u8bc4\u6d4b\u6bd4</td></tr><tr><td>\u8d5b\"\u7684\u8bc4\u6d4b\u8bed\u6599\u4e0a\u8fdb\u884c\u5b9e\u9a8c\uff0c\u5b9e\u9a8c\u7ed3\u679c\u8bc1\u660ePLA-HAN\u6a21\u578b\u53d6\u5f97\u4e86\u4f18\u4e8eHAN\u7b49\u7814\u7a76\u8fdb\u5c55\u6587\u672c\u5206\u7c7b</td></tr><tr><td>\u65b9\u6cd5\u7684\u610f\u56fe\u5206\u7c7b\u6b63\u786e\u7387\u3002</td></tr><tr><td>\u81f4 \u81f4 \u81f4\u8c22 \u8c22 \u8c22</td></tr><tr><td>\u672c\u6587\u53d7\u5230\u56fd\u5bb6\u81ea\u7136\u79d1\u5b66\u57fa\u91d1(\u9879\u76ee\u7f16\u53f7:71472068)\u7684\u8d44\u52a9\u3002</td></tr></table>",
                "type_str": "table",
                "num": null,
                "text": "\u4e0d\u540c\u4f2a\u6807\u7b7e\u96c6\u7684PLA-HAN\u6a21\u578b\u6027\u80fd\u5bf9\u6bd4 \u4e0d\u540c \u540c \u540c\u957f \u957f \u957f\u5ea6 \u5ea6 \u5ea6\u5bf9 \u5bf9 \u5bf9\u8bdd \u8bdd \u8bdd\u6bb5 \u6bb5 \u6bb5\u7684 \u7684 \u7684\u6a21 \u6a21 \u6a21\u578b \u578b \u578b\u6027 \u6027 \u6027\u80fd \u80fd \u80fd\u5bf9 \u5bf9 \u5bf9\u6bd4 \u6bd4 \u6bd4\u3002\u6211\u4eec\u8fdb\u4e00\u6b65\u5bf9\u6bd4\u4e0d\u540c\u5bf9\u8bdd\u6bb5\u957f\u5ea6\u60c5\u51b5\u4e0b\u7684PLA-HAN\u6a21"
            }
        }
    }
}