{
    "paper_id": "2020",
    "header": {
        "generated_with": "S2ORC 1.0.0",
        "date_generated": "2023-01-19T12:53:52.697899Z"
    },
    "title": "Chinese question answering method based on multi-head attention and BiLSTM improved DAM model",
    "authors": [
        {
            "first": "Hanzhong",
            "middle": [],
            "last": "Qin",
            "suffix": "",
            "affiliation": {},
            "email": ""
        },
        {
            "first": "Chongchong",
            "middle": [],
            "last": "Yu",
            "suffix": "",
            "affiliation": {},
            "email": ""
        },
        {
            "first": "Weijie",
            "middle": [],
            "last": "Jiang",
            "suffix": "",
            "affiliation": {},
            "email": ""
        },
        {
            "first": "Xia",
            "middle": [],
            "last": "Zhao",
            "suffix": "",
            "affiliation": {},
            "email": "zhaox@btbu.edu.cn"
        }
    ],
    "year": "",
    "venue": null,
    "identifiers": {},
    "abstract": "Aiming at the current problem of Deep Attention Matching Network(DAM) can not effectively match response details, and will cause semantic confusion, a Chinese question answering method based on multi-head attention and Bi-directional Long Short-Term Memory (BiLSTM) improved DAM model was proposed. This method can model longer multiple rounds of dialogue and handle the matching relationship between the response selection and the context. In addition, this paper uses the BiLSTM Network in the feature fusion process to improve the accuracy of multi-turn response selection tasks by capturing the time-dependent relation. In this paper, we test the improved DAM on two public multi-turn response selection datasets, the Douban Conversion Corpus and the E-commerce Dialogue Corpus. Experimental results show our model outperforms the baseline model by 1.5% in Recall-10-at-1 with the word vector enhancement.",
    "pdf_parse": {
        "paper_id": "2020",
        "_pdf_hash": "",
        "abstract": [
            {
                "text": "Aiming at the current problem of Deep Attention Matching Network(DAM) can not effectively match response details, and will cause semantic confusion, a Chinese question answering method based on multi-head attention and Bi-directional Long Short-Term Memory (BiLSTM) improved DAM model was proposed. This method can model longer multiple rounds of dialogue and handle the matching relationship between the response selection and the context. In addition, this paper uses the BiLSTM Network in the feature fusion process to improve the accuracy of multi-turn response selection tasks by capturing the time-dependent relation. In this paper, we test the improved DAM on two public multi-turn response selection datasets, the Douban Conversion Corpus and the E-commerce Dialogue Corpus. Experimental results show our model outperforms the baseline model by 1.5% in Recall-10-at-1 with the word vector enhancement.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Abstract",
                "sec_num": null
            }
        ],
        "body_text": [
            {
                "text": "1 \u5f15\u8a00 \u4eba\u673a\u5bf9\u8bdd\u7cfb\u7edf\u662f\u4e00\u4e2a\u590d\u6742\u7684\u7814\u7a76\u65b9\u5411\uff0c\u6784\u5efa\u4eba\u673a\u5bf9\u8bdd\u7cfb\u7edf\u7684\u65b9\u6cd5\u4e4b\u4e00\u662f\u68c0\u7d22\u5f0f\u65b9\u6cd5\u3002\u68c0\u7d22 \u5f0f\u65b9\u6cd5\u9996\u5148\u9700\u63d0\u53d6\u8f93\u5165\u5bf9\u8bdd\u7684\u7279\u5f81\uff0c\u968f\u540e\u5728\u5019\u9009\u56de\u590d\u5e93\u5339\u914d\u591a\u4e2a\u76ee\u6807\u5019\u9009\u56de\u590d\uff0c\u6309\u7167\u67d0\u79cd\u6307\u6807 \u8fdb\u884c\u6392\u5e8f\uff0c\u8f93\u51fa\u5f97\u5206\u6700\u9ad8\u7684\u56de\u590d\u3002\u5c06\u4e4b\u524d\u7684\u5bf9\u8bdd\u8f93\u51fa\u4f5c\u4e3a\u5386\u53f2\u5bf9\u8bdd\uff0c\u5373\u5f62\u6210\u591a\u8f6e\u5bf9\u8bdd\u7684\u5f62\u5f0f\u3002 \u8fd1\u5e74\u6765\uff0c\u968f\u7740\u6df1\u5ea6\u5b66\u4e60\u7684\u53d1\u5c55\uff0c\u5173\u4e8e\u4eba\u673a\u5bf9\u8bdd\u7684\u7814\u7a76\u91cd\u70b9\u9010\u6e10\u7531\u57fa\u4e8e\u6a21\u677f\u3001\u89c4\u5219\u7684\u4f20\u7edf\u65b9 \u6cd5\u8f6c\u53d8\u4e3a\u57fa\u4e8e\u7aef\u5230\u7aef\u7684\u6df1\u5ea6\u5b66\u4e60\u6a21\u578b\u65b9\u6cd5\u3002 (Wu Y et al.,2016) \u63d0\u51fa\u5e8f\u5217\u5339\u914d\u7f51\u7edc(Sequence Matching Network\uff0cSMN)\u6a21\u578b\uff0c\u6a21\u578b\u53ef\u5206\u4e3a\"\u8868\u793a-\u5339\u914d-\u878d\u5408\"\u4e09\u4e2a\u90e8\u5206\uff0c\u6574\u4f53\u4e0a\u57fa\u4e8e CNN \u548c RNN \u5b9e\u73b0\u4ee5\u8bed\u4e49\u878d\u5408\u4e3a\u4e2d\u5fc3\u7684\u591a\u8f6e\u5bf9\u8bdd\u56de\u590d\u9009\u62e9\u3002 (Zhang Z et al.,2018) \u63d0\u51fa\u6df1\u5ea6\u8868\u8fbe\u878d\u5408 (Deep Utterance Aggregation\uff0cDUA)\u6a21\u578b\uff0c\u9488\u5bf9 SMN \u6a21\u578b\u5c06\u5386\u53f2\u5bf9\u8bdd\u76f4\u63a5\u62fc\u63a5\u4e3a\u4e0a\u4e0b\u6587\u5b58\u5728 \u566a\u58f0\u548c\u5197\u4f59\u7684\u95ee\u9898\uff0c\u91c7\u7528\u6ce8\u610f\u529b\u673a\u5236\u6316\u6398\u5173\u952e\u4fe1\u606f\u5e76\u5ffd\u7565\u5197\u4f59\u4fe1\u606f\uff0c\u6700\u7ec8\u83b7\u5f97\u5bf9\u8bdd\u8868\u8fbe\u548c\u5019\u9009 \u54cd\u5e94\u7684\u5339\u914d\u5f97\u5206\u3002 (Zhou X et al.,2018) (Zhou X et al.,2018) \u6570\u636e\u9884\u5904\u7406\u7684\u7b2c\u4e00\u90e8\u5206\u662f\u6570\u636e\u6e05\u6d17\uff0c\u5229\u7528 response.txt \u4e2d\u7684\u5019\u9009\u56de\u590d\u8868\uff0c\u53ef\u7d22\u5f15 ",
                "cite_spans": [
                    {
                        "start": 189,
                        "end": 207,
                        "text": "(Wu Y et al.,2016)",
                        "ref_id": "BIBREF8"
                    },
                    {
                        "start": 307,
                        "end": 328,
                        "text": "(Zhang Z et al.,2018)",
                        "ref_id": "BIBREF9"
                    },
                    {
                        "start": 449,
                        "end": 469,
                        "text": "(Zhou X et al.,2018)",
                        "ref_id": null
                    },
                    {
                        "start": 470,
                        "end": 490,
                        "text": "(Zhou X et al.,2018)",
                        "ref_id": null
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            },
            {
                "text": "\u63d0\u51fa\u7684 DAM \u6a21\u578b\u7684\u6574\u4f53\u7ed3\u6784\u5982\u56fe 1 \u6240\u793a\uff0c\u53ef\u4ee5\u5206\u4e3a\u8f93\u5165\u3001\u8868\u793a\u3001\u5339 \u914d\u3001\u805a\u5408\u56db\u4e2a\u90e8\u5206\u3002\u6a21\u578b\u7684\u8f93\u5165\u662f\u591a\u8f6e\u5bf9\u8bdd\u548c\u5019\u9009\u56de\u590d\uff0c\u8f93\u51fa\u662f\u6bcf\u4e2a\u5019\u9009\u56de\u590d\u7684\u5f97\u5206\u3002 \u56fe 1. DAM \u6a21\u578b\u7684\u6574\u4f53\u7ed3\u6784 DAM \u6a21\u578b\u7684\u6ce8\u610f\u529b\u6a21\u5757\u542b\u6709\u67e5\u8be2\u5411\u91cf Q\u3001\u952e\u5411\u91cf K \u548c\u503c\u5411\u91cf V \u4e09\u4e2a\u8f93\u5165\u3002\u6a21\u5757\u9996\u5148\u5229\u7528 \u4ee5\u4e0b\u516c\u5f0f\u8ba1\u7b97\u8f93\u5165\u7684\u7f29\u653e\u70b9\u79ef\u6ce8\u610f\u529b\uff1a ) , , ( V att V K Q Attention \uf03d (1) \u4e4b\u540e\u6a21\u5757\u5c06 att V \u548c Q \u76f4\u63a5\u76f8\u52a0\uff0c\u4ea7\u751f\u7684\u548c\u5305\u542b\u4e8c\u8005\u7684\u8054\u5408\u8bed\u4e49\u4fe1\u606f\u3002\u4e3a\u9632\u6b62\u68af\u5ea6\u6d88\u5931\u6216\u68af \u5ea6\u7206\u70b8\uff0c\u5bf9 att V \u548c Q \u76f8\u52a0\u7684\u7ed3\u679c\u5e94\u7528\u5c42\u5f52\u4e00\u5316\uff0c\u7ed3\u679c\u8bb0\u4e3a ln V \u3002\u63a5\u7740\u5c06 ln V \u4f20\u5165\u4e00\u4e2a\u57fa\u4e8e ReLU \u51fd \u6570\u6fc0\u6d3b\u7684\u53cc\u5c42\u524d\u9988\u7f51\u7edc FFN\uff0c\u8fdb\u4e00\u6b65\u5904\u7406\u878d\u5408\u4fe1\u606f\u3002FFN \u7684\u8f93\u51fa\u4e0e\u8f93\u5165\u8fdb\u884c\u4e00\u6b21\u6b8b\u5dee\u8fde\u63a5\uff0c\u4ea7 \u751f\u7684\u7ed3\u679c\u518d\u6b21\u5e94\u7528\u5c42\u5f52\u4e00\u5316\uff0c\u6b64\u65f6\u7684 O \u662f\u6574\u4e2a\u6ce8\u610f\u529b\u6a21\u5757\u8ba1\u7b97\u8fc7\u7a0b\u7684\u6700\u7ec8\u8f93\u51fa\uff0c\u5c06\u5176\u8868\u793a\u4e3a\uff1a ) , , ( V K Q odule AttentionM O \uf03d (2) DAM \u6a21\u578b\u7684\u8bed\u4e49\u8868\u793a\u7f51\u7edc\u7531\u591a\u4e2a\u76f8\u540c\u7684\u6ce8\u610f\u529b\u6a21\u5757\u9996\u5c3e\u76f8\u8fde\uff0c\u5f62\u6210\u5806\u53e0\u7684\u7f51\u7edc\u7ed3\u6784\u3002\u7f51\u7edc \u4e2d\u6bcf\u4e2a\u6ce8\u610f\u529b\u6a21\u5757\u7684\u4e09\u4e2a\u8f93\u5165\u76f8\u540c\uff0c\u8ba1\u7b97\u81ea\u6ce8\u610f\u529b\uff0c\u516c\u5f0f\u8868\u793a\u4e3a\uff1a ) , , ( 1 i l i l i l i l U U U odule AttentionM U \uf03d \uf02b (3) ) , , ( 1 l l l l R R R odule AttentionM R \uf03d \uf02b (4) \u5176\u4e2d l \u7684\u8303\u56f4\u4ece 0 \u5230 1 \uf02d L \uff0cL \u8868\u793a\u6a21\u5757\u5806\u53e0\u7684\u6570\u91cf\uff0c i i u U \uf03d 0 \u4ee5\u53ca r R \uf03d 0 \u662f\u539f\u59cb\u8f93\u5165\u3002 DAM \u6a21\u578b\u4e2d\u7279\u5f81\u5339\u914d\u7684\u8f93\u5165\u662f\u8bed\u4e49\u8868\u793a\u7f51\u7edc\u7684\u8f93\u51fa i U \u548c R\u3002\u9488\u5bf9\u4e0d\u540c\u7c92\u5ea6 l\uff0c\u7f51\u7edc\u5c06\u4ea7\u751f \u4e24\u79cd\u5339\u914d\u77e9\u9635\uff0c\u4e00\u79cd\u662f\u81ea\u5339\u914d\u77e9\u9635 l r u self i M , , \uff0c\u53e6\u4e00\u79cd\u662f\u4e92\u5339\u914d\u77e9\u9635 l r u cross i M , , \u3002 l r u self i M , , \u662f i U \u548c R \u7684\u70b9 \u79ef\uff0c\u77e9\u9635\u4e2d\u542b\u6709 i U \u548c R \u5143\u7d20\u95f4\u7684\u8bed\u4e49\u4f9d\u8d56\u5173\u7cfb\u3002 l r u cross i M , , \u7684\u4ea7\u751f\u57fa\u4e8e\u5bf9\u6ce8\u610f\u529b\u6a21\u5757\u8f93\u5165\u7684\u4fee\u6539\uff0c \u901a\u8fc7\u4ee4 i U \u548c R \u4e2d\u5bf9\u5e94\u5143\u7d20\u4e92\u76f8\u5173\u6ce8\uff0c\u6784\u9020\u51fa\u65b0\u7684\u8bed\u4e49\u8868\u793a l i \u0168 \u548c l R \uff0c\u7528\u4e8e\u6355\u6349\u8de8\u8d8a\u5bf9\u8bdd\u8868\u8fbe\u4e0e \u5019\u9009\u56de\u590d\u4e4b\u95f4\u7684\u4ea4\u53c9\u5173\u8054\u7279\u5f81\u3002\u4e8c\u8005\u7684\u8ba1\u7b97\u516c\u5f0f\u4e3a\uff1a ) , , ( i l l l i l R R U odule AttentionM U \uf03d (5) ) , , ( l l i l i l U U R odule AttentionM R \uf03d (6) \u901a\u8fc7\u8ba1\u7b97 l i \u0168 \u548c l R \u7684\u70b9\u79ef\uff0c \u5f97\u5230 l r u cross i M , , \uff0c DAM \u6a21\u578b\u4e2d\u7684\u7279\u5f81\u878d\u5408\u7f51\u7edc\u5c06\u591a\u4e2a\u7c92\u5ea6\u7684 l r u self i M , , \u548c l r u cross i M , , \u4f5c\u4e3a\u8f93\u5165\uff0c\u5728 i \u548c l \u4e24\u4e2a\u7ef4\u5ea6\u62fc\u63a5\u4e24\u4e2a\u77e9\u9635\uff0c\u5f97\u5230\u77e9\u9635 l i P , \u5982\u4e0b\uff1a l r u cross l r u self i i M M P , , , , l , i \uf0c5 \uf03d (7) \u5728 DAM \u6a21\u578b\u4e2d\uff0c l i P , \u88ab\u79f0\u4e3a\u50cf\u7d20\u70b9\uff0c\u7531 l i P , \u7ec4\u5408\u5f62\u6210\u7684\u9ad8\u7ef4\u77e9\u9635 P \u88ab\u79f0\u4e3a\u56fe\u50cf\uff0c\u8fd9\u6837\u7684\u547d \u540d\u662f\u4e3a\u4e86\u65b9\u4fbf\u4f7f\u7528 CNN\u3002P \u4e2d\u7684\u56fe\u50cf\u6df1\u5ea6\u5bf9\u5e94\u4e8e\u591a\u8f6e\u5bf9\u8bdd\u7684\u8f6e\u6b21\uff0c\u56fe\u50cf\u5bbd\u5ea6\u5bf9\u5e94\u4e8e\u6bcf\u8f6e\u5bf9\u8bdd\u548c \u5019\u9009\u56de\u590d\u5728\u53e5\u5b50\u5c42\u7ea7\u7684\u5339\u914d\u4fe1\u606f\uff0c\u56fe\u50cf\u9ad8\u5ea6\u5bf9\u5e94\u4e8e\u6bcf\u8f6e\u5bf9\u8bdd\u548c\u5019\u9009\u56de\u590d\u5728\u5355\u8bcd\u5c42\u7ea7\u7684\u5339\u914d\u4fe1\u606f\u3002 \u7531\u4e8e P \u542b\u6709\u4e09\u4e2a\u7ef4\u5ea6\u7684\u7279\u5f81\uff0cDAM \u6a21\u578b\u91c7\u7528 3D \u5377\u79ef\u8fdb\u884c\u7279\u5f81\u63d0\u53d6\u3002\u7ecf\u8fc7\u4e24\u6b21 3D \u5377\u79ef\u548c\u6700\u5927 \u6c60\u5316\uff0cP \u6700\u7ec8\u53d8\u6210\u4e00\u7ef4\u7279\u5f81 f\uff0c\u518d\u7ecf\u8fc7\u4e00\u4e2a\u7ebf\u6027\u5206\u7c7b\u5668\u5373\u53ef\u83b7\u5f97\u5339\u914d\u5206\u6570 ) , ( r c g \u3002 3 \u57fa\u4e8e\u591a\u5934\u6ce8\u610f\u529b\u673a\u5236\u548c BiLSTM \u7f51\u7edc\u7684 Ex-DAM \u6a21\u578b 3.1 \u57fa\u4e8e\u591a\u5934\u6ce8\u610f\u529b\u548c BiLSTM \u7684 Ex-DAM \u6a21\u578b \u4e3a\u4f7f DAM \u6a21\u578b\u66f4\u9002\u5408\u5904\u7406\u542b\u6709\u7ec6\u5fae\u53d8\u5316\u7684\u6570\u636e\uff0c \u8fdb\u4e00\u6b65\u63d0\u9ad8\u9009\u62e9\u76ee\u6807\u5019\u9009\u56de\u590d\u7684\u51c6\u786e\u7387\uff0c \u672c\u6587\u5229\u7528\u591a\u5934\u6ce8\u610f\u529b\u8868\u793a\u7f51\u7edc\u548c\u53cc\u901a\u9053\u7279\u5f81\u878d\u5408\u7f51\u7edc\uff0c\u7ed3\u5408 DAM \u6a21\u578b\u4e2d\u7684\u4ea4\u4e92\u5339\u914d\u7f51\u7edc\uff0c\u57fa \u4e8e\u6b64\u6784\u6210\u4e00\u4e2a\u65b0\u7684\u7aef\u5230\u7aef\u68c0\u7d22\u5f0f\u591a\u8f6e\u5bf9\u8bdd\u7cfb\u7edf\u6a21\u578b\uff0c\u5c06\u8be5\u6a21\u578b\u547d\u540d\u4e3a\u57fa\u4e8e\u591a\u5934\u6ce8\u610f\u529b\u548c BiLSTM \u7684 Ex-DAM \u6a21\u578b\u3002\u6a21\u578b\u7684\u6574\u4f53\u7ed3\u6784\u5982\u56fe 2 \u6240\u793a\u3002 Matching Score M ru M u M rl M un-1, l M u0, l R l r u 0 i=1~L Multi-Head Attention Layer i Multi-Head Attention Layer i Multi-Head Attention Layer i u n-1 U 0, l \u2026 \u2026 U n-1, l Attention Layer Attention Layer Attention Layer \u2026 BiLSTM \uf0c5 BiLSTM \uf0c5 BiLSTM \uf0c5 M ur F MLP Representation Matching Aggregation Prediction \u56fe 2. Ex-DAM \u6a21\u578b\u6574\u4f53\u7ed3\u6784 \u6a21\u578b\u7684\u8f93\u5165\u662f\u8bcd\u5411\u91cf\u5f62\u5f0f\u7684\u591a\u8f6e\u5bf9\u8bdd\u548c\u5019\u9009\u56de\u590d\uff0c\u9996\u5148\u7ecf\u8fc7 L \u4e2a\u591a\u5934\u6ce8\u610f\u529b\u5c42\u4ee5\u83b7\u53d6\u5b83\u4eec \u7684\u591a\u7c92\u5ea6\u8868\u793a\u3002\u5728\u8fd9\u4e9b\u8868\u793a\u5411\u91cf\u4e2d\uff0c\u6bcf\u4e00\u8f6e\u5bf9\u8bdd\u90fd\u548c\u5019\u9009\u56de\u590d\u8fdb\u884c\u4e00\u6b21\u666e\u901a\u7684\u6ce8\u610f\u529b\u8ba1\u7b97\uff0c\u5f97 \u5230\u591a\u4e2a\u4e3b\u5339\u914d\u77e9\u9635\u3002\u6b64\u5916\uff0c\u5019\u9009\u56de\u590d\u518d\u989d\u5916\u5730\u4e0e\u6700\u540e\u4e00\u8f6e\u5bf9\u8bdd\u8ba1\u7b97\u4e00\u6b21\u6ce8\u610f\u529b\uff0c\u4ee5\u83b7\u5f97\u6b21\u5339\u914d \u77e9\u9635\u3002\u968f\u540e\uff0c\u4e3b\u3001\u6b21\u5339\u914d\u77e9\u9635\u5206\u522b\u4f5c\u4e3a\u4e24\u4e2a\u901a\u9053\u7684\u8f93\u5165\u8fdb\u884c\u7279\u5f81\u878d\u5408\u3002\u5728\u8fd9\u4e2a\u8fc7\u7a0b\u4e2d\uff0c\u6240\u6709\u7684 \u5339\u914d\u77e9\u9635\u7ecf\u8fc7 BiLSTM \u548c\u62fc\u63a5\u64cd\u4f5c\u4f9d\u6b21\u8fdb\u884c\u5e8f\u5217\u7279\u5f81\u63d0\u53d6\u548c\u7ef4\u5ea6\u7edf\u4e00\u3002\u6700\u540e\uff0c\u628a\u4e24\u4e2a\u901a\u9053\u7684\u8f93 \u51fa\u5411\u91cf\u9996\u5c3e\u62fc\u63a5\uff0c\u7ecf\u8fc7\u591a\u5c42\u611f\u77e5\u5668\u5c31\u80fd\u83b7\u5f97\u6bcf\u4e2a\u5019\u9009\u56de\u590d\u4e0e\u591a\u8f6e\u5bf9\u8bdd\u4e4b\u95f4\u7684\u5339\u914d\u5206\u6570\u3002 3.2 Ex-DAM \u6a21\u578b\u4e2d\u7684\u591a\u5934\u6ce8\u610f\u529b\u6a21\u5757\u548c\u8bed\u4e49\u8868\u793a\u7f51\u7edc \u666e\u901a\u7684\u6ce8\u610f\u529b\u673a\u5236\u5728\u6587\u672c\u5e8f\u5217\u4e2d\u53ef\u4ee5\u5f88\u597d\u5730\u63d0\u53d6\u8bcd\u5411\u91cf\u89d2\u5ea6\u7684\u5173\u952e\u4fe1\u606f\uff0c\u4f46\u51e0\u4e4e\u65e0\u6cd5\u8bc6\u522b \u5bf9\u8bcd\u5411\u91cf\u8fdb\u884c\u7edf\u4e00\u4fee\u6539\u7684\u64cd\u4f5c\u3002\u591a\u5934\u6ce8\u610f\u529b\u673a\u5236\u6b63\u597d\u53ef\u4ee5\u89e3\u51b3\u6b64\u7c7b\u95ee\u9898\uff0c\u5728\u8ba1\u7b97\u65f6\u9996\u5148\u8f93\u5165\u591a \u6b21\u6620\u5c04\uff0c\u6bcf\u4e2a\u6620\u5c04\u4f7f\u7528\u4e0d\u540c\u53c2\u6570\u8fdb\u884c\u76f8\u540c\u8ba1\u7b97\uff0c\u6700\u540e\u5c06\u5404\u4e2a\u8f93\u51fa\u5408\u5e76\u5728\u4e00\u8d77\uff0c\u56e0\u6b64\u6bd4\u7f29\u653e\u70b9\u79ef \u6ce8\u610f\u529b\u66f4\u9002\u5408\u5904\u7406\u542b\u6709\u7ec6\u5fae\u53d8\u5316\u7684\u6570\u636e\uff0c\u672c\u6587\u4f7f\u7528\u591a\u5934\u6ce8\u610f\u529b\u6a21\u5757\u7ed3\u6784\u5982\u56fe 3 \u6240\u793a\u3002 \u56fe 3. \u591a\u5934\u6ce8\u610f\u529b\u6a21\u5757 \u591a\u5934\u6ce8\u610f\u529b\u6a21\u5757\u542b\u6709\u67e5\u8be2\u5411\u91cf Q\u3001\u952e\u5411\u91cf K \u548c\u503c\u5411\u91cf V \u4e09\u4e2a\u8f93\u5165\uff0c\u6574\u4f53\u8ba1\u7b97\u516c\u5f0f\u4e3a\uff1a )) , , ( ( ) , , ( V K Q MultiHead Q LayerNorm V K Q MHAModule \uf02b \uf03d (8) \u672c\u6587\u5728\u6a21\u5757\u4e2d\u5e94\u7528\u4e86\u6b8b\u5dee\u8fde\u63a5\uff0c\u5c06\u8f93\u5165 Q \u4e0e\u524d\u9988\u5c42\u7684\u8f93\u51fa\u6052\u7b49\u53e0\u52a0\uff0c\u4e0d\u4f1a\u5f15\u5165\u989d\u5916\u7684\u53c2\u6570\uff0c \u4e5f\u4e0d\u4f1a\u589e\u52a0\u6a21\u578b\u7684\u8ba1\u7b97\u590d\u6742\u5ea6\uff0c\u5728\u53e0\u52a0\u8fc7\u7a0b\u4e2d\u53ef\u5f3a\u5316\u8f93\u5165\u4e2d\u7684\u91cd\u70b9\u5185\u5bb9\uff0c\u63d0\u5347\u8bad\u7ec3\u6548\u679c\u3002\u591a\u5934 \u6ce8\u610f\u529b\u7684\u5934\u90e8\u80fd\u5728\u4e0d\u540c\u5b50\u7a7a\u95f4\u5904\u7406\u540c\u4e00\u5e8f\u5217\uff0c\u4ece\u800c\u83b7\u5f97\u66f4\u4e30\u5bcc\u7684\u8bed\u4e49\u8868\u793a\u4fe1\u606f\u3002\u7531\u591a\u5934\u6ce8\u610f\u529b \u6a21\u5757\u7ec4\u6210 Ex-DAM \u6a21\u578b\u7684\u8bed\u4e49\u8868\u793a\u7f51\u7edc\uff0c\u7ed3\u6784\u5982\u56fe 2 \u4e2d\u7684 Representation \u6a21\u5757\u6240\u793a\u3002 \u8f93\u5165 = 1 [ ] i T i i u \uf03d \u548c 2 10 1 [ ] i i i r \uf0a3 \uf0a3 \uf03d \u5206\u522b\u662f\u6bcf\u8f6e\u5bf9\u8bdd\u548c\u5019\u9009\u56de\u590d\u7684\u8bcd\u5411\u91cf\uff0c\u8fd9\u4e9b\u8bcd\u5411\u91cf\u5df2\u7ecf\u8fc7\u589e\u5f3a\u5904\u7406\uff0c \u540c\u65f6\u542b\u6709\u610f\u56fe\u4fe1\u606f\u548c\u8bed\u4e49\u4fe1\u606f\u3002\u8bcd\u5411\u91cf\u8fdb\u5165\u591a\u5934\u6ce8\u610f\u529b\u6a21\u5757\u8ba1\u7b97\u8bed\u4e49\u8868\u793a\u7684\u516c\u5f0f\u4e3a\uff1a ) , , ( 1 i l i l i l i l U U U MHAModule U \uf03d \uf02b (9) ) , , ( 1 l l l l R R R MHAModule R \uf03d \uf02b (10) \u4ee5 i i u U \uf03d 0 \u548c r R \uf03d 0 \u4e3a\u591a\u5934\u6ce8\u610f\u529b\u6a21\u5757\u7684\u539f\u59cb\u8f93\u5165\uff0c\u7ecf 1 \uf02d L \u4e2a\u5806\u53e0\u7684\u76f8\u540c\u6a21\u5757\u53ef\u9010\u5c42\u83b7\u5f97\u591a \u7c92\u5ea6\u7684\u8bed\u4e49\u8868\u793a\uff0c\u5373 ] ,..., [ 0 L i i i U U U \uf03d \u548c ] ,..., [ 0 L i i R R R \uf03d \u3002\u7531\u4e8e\u540e\u7eed\u5904\u7406\u7684\u8ba1\u7b97\u91cf\u8f83\u5927\uff0c\u4e3a\u907f\u514d\u5185 \u5b58\u6ea2\u51fa\u9700\u63a7\u5236\u8f93\u5165\u6570\u91cf\uff0c\u672c\u6587\u9488\u5bf9 i U \u548c R \u4e2d\u7684\u5143\u7d20\u9009\u62e9\u5236\u5b9a\u4e86\u4ee5\u4e0b\u4e09\u4e2a\u5904\u7406\u7b56\u7565\uff1a \u7b2c\u4e00\u4e2a\u7b56\u7565\u662f\u5f53 L \u6570\u503c\u8f83\u5c0f\u65f6\uff0c\u4fdd\u7559\u6240\u6709 i U \u548c R\uff0c\u5373\u4f7f\u7528\u6240\u6709\u7c92\u5ea6\u7684\u8bed\u4e49\u8868\u793a\u4f5c\u4e3a\u7279\u5f81\u5339 \u914d\u7f51\u7edc\u7684\u8f93\u5165\uff0c\u8bb0\u4f5c Ex-DAML\uff1b\u7b2c\u4e8c\u4e2a\u7b56\u7565\u662f\u5f53 L \u6570\u503c\u8f83\u5927\u65f6\uff0c\u4fdd\u7559 i U \u548c R \u4e2d\u7684\u540e m \u4e2a\u5143\u7d20\uff0c \u5373\u4ec5\u4f7f\u7528\u6df1\u5c42\u7c92\u5ea6\u7684\u8bed\u4e49\u8868\u793a\u4f5c\u4e3a\u7279\u5f81\u5339\u914d\u7f51\u7edc\u7684\u8f93\u5165\uff0c\u8bb0\u4f5c Ex-DAML-m\uff1b\u7b2c\u4e09\u4e2a\u7b56\u7565\u662f\u5f53 L \u6570\u503c\u8f83\u5927\u65f6\uff0c\u4fdd\u7559 i U \u548c R \u4e2d\u7684\u7b2c\u4e00\u4e2a\u5143\u7d20\u548c\u540e m \u4e2a\u5143\u7d20\uff0c\u5c06\u539f\u59cb\u8f93\u5165\u540c\u65f6\u4f5c\u4e3a\u8bed\u4e49\u8868\u793a\u7f51\u7edc\u548c \u7279\u5f81\u5339\u914d\u7f51\u7edc\u7684\u8f93\u5165\uff0c\u800c\u540e\u8005\u7684\u8f93\u5165\u8fd8\u5305\u542b\u539f\u59cb\u8f93\u5165\u7684\u591a\u7c92\u5ea6\u8bed\u4e49\u8868\u793a\uff0c\u8bb0\u4f5c Ex-DAML-0-m\u3002 3.3 Ex-DAM \u6a21\u578b\u4e2d\u7684\u53cc\u901a\u9053 BiLSTM \u7279\u5f81\u878d\u5408\u7f51\u7edc \u672c\u6587\u63d0\u51fa\u7684 Ex-DAM \u6a21\u578b\u7684\u7279\u5f81\u5339\u914d\u7f51\u7edc\u4eff\u7167 DAM \u6a21\u578b\uff0c\u4ee5\u8bed\u4e49\u8868\u793a\u7f51\u7edc\u7684\u8f93\u51fa i U \u548c R \u4f5c\u4e3a\u8f93\u5165\uff0c\u5229\u7528 DAM \u6a21\u578b\u4e2d\u7684\u6ce8\u610f\u529b\u6a21\u5757\u6784\u9020\u81ea\u5339\u914d\u77e9\u9635 l r u self i M , , \u548c\u4e92\u5339\u914d\u77e9\u9635 l r u cross i M , , \uff0c\u5982\u56fe 2 \u4e2d\u7684 Matching \u6a21\u5757\uff0c\u7531\u4e8e\u672c\u6587\u5728\u8be5\u7f51\u7edc\u4e2d\u672a\u505a\u6539\u52a8\uff0c\u8fd9\u91cc\u4e0d\u505a\u8d58\u8ff0\u3002\u5728 DAM \u6a21\u578b\u7684\u7279\u5f81\u878d\u5408 \u7f51\u7edc\u4e2d\uff0c3D \u5377\u79ef\u4f5c\u4e3a\u6ce8\u610f\u529b\u673a\u5236\u7684\u4e00\u79cd\u8f85\u52a9\u7b56\u7565\uff0c\u5728\u8fdb\u4e00\u6b65\u63d0\u53d6\u7279\u5f81\u7684\u540c\u65f6\u7f29\u5c0f\u6570\u636e\u7ef4\u5ea6\u3002\u7531 \u4e8e\u672c\u6587\u5df2\u5c06\u7f29\u653e\u70b9\u79ef\u6ce8\u610f\u529b\u66ff\u6362\u4e3a\u591a\u5934\u6ce8\u610f\u529b\uff0c\u56e0\u6b64\u672c\u6587\u5728 Ex-DAM \u6a21\u578b\u4e2d\u4e5f\u5c06 3D \u5377\u79ef\u8fdb\u884c \u4e86\u66ff\u6362\uff0cEx-DAM \u6a21\u578b\u7684\u7279\u5f81\u878d\u5408\u7f51\u7edc\u5982\u56fe 2 \u4e2d Aggregation \u6a21\u5757\u6240\u793a\u3002 \u5728 Ex-DAM \u6a21\u578b\u7684\u7279\u5f81\u878d\u5408\u7f51\u7edc\u4e2d\uff0c\u6587\u672c\u9996\u5148\u5bf9 l r u self i M , , \u548c l r u cross i M , , \u8fdb\u884c\u52a0\u6743\u6c42\u548c\uff0c\u63d0\u53d6\u5339\u914d \u77e9\u9635\u4e2d\u7684\u91cd\u8981\u5339\u914d\u7279\u5f81\uff0c\u5176\u4e2d l w \u662f\u5171\u4eab\u6743\u91cd\uff0c\u4f5c\u7528\u662f\u589e\u5f3a\u6a21\u578b\u7684\u6cdb\u5316\u6027\u5e76\u51cf\u5c11\u8ba1\u7b97\u5f00\u9500\u3002\u8ba1\u7b97 \u5f97\u5230\u7684 i self M \u542b\u6709\u6bcf\u8f6e\u5bf9\u8bdd\u4e0e\u524d\u4e00\u8f6e\u5bf9\u8bdd\u4e4b\u95f4\u7684\u5339\u914d\u4f9d\u8d56\u4fe1\u606f\uff0c\u800c i cross M \u4e2d\u542b\u6709\u6bcf\u8f6e\u5bf9\u8bdd\u4e0e\u5019\u9009 \u56de\u590d\u4e4b\u95f4\u7684\u5339\u914d\u4f9d\u8d56\u4fe1\u606f\u3002 \u63a5\u4e0b\u6765\uff0c\u672c\u6587\u5229\u7528\u4e24\u4e2a\u4e0d\u540c\u7684 BiLSTM \u7f51\u7edc\u5206\u522b\u5904\u7406 i self M \u548c i cross M \uff0c\u63d0\u53d6\u5176\u4e2d\u7ec6\u5fae\u7247\u6bb5\u4e4b \u95f4\u7684\u5339\u914d\u5173\u7cfb\u3002\u4e0e\u539f\u59cb DAM \u6a21\u578b\u4e2d\u4f7f\u7528\u7684 3D \u5377\u79ef\u4e0d\u540c\uff0cEx-DAM \u6a21\u578b\u653e\u5f03\u4ece\"\u8f6e\u6b21-\u5bf9\u8bdd- \u5019\u9009\u56de\u590d\"\u7684\u89d2\u5ea6\u5165\u624b\uff0c\u8f6c\u800c\u8003\u8651\u4ee5\u8f6e\u6b21\u4e3a\u5355\u4e00\u4e3b\u7ebf\uff0c\u5206\u522b\u5bf9\"\u5bf9\u8bdd-\u5bf9\u8bdd\"\u4ee5\u53ca\"\u5bf9\u8bdd-\u5019\u9009\u56de \u590d\"\u8fdb\u884c\u72ec\u7acb\u8ba1\u7b97\u3002\u540c\u65f6\uff0c\u7531\u4e8e\u5c06\u5728\u5b9e\u9a8c\u4e2d\u5f15\u5165\u57fa\u4e8e\u610f\u56fe\u8bc6\u522b\u7684\u8bcd\u5411\u91cf\u589e\u5f3a\uff0c\u4ee5\u7d2f\u52a0\u5f62\u5f0f\u5d4c\u5165\u5230 \u5bf9\u8bdd\u4e2d\u7684\u610f\u56fe\u5d4c\u5165\u5411\u91cf\u4e5f\u8f83\u5bb9\u6613\u5f15\u8d77 BiLSTM \u7684\u5173\u6ce8\u3002\u5c06\u4ee3\u8868\u4e0d\u540c\u8f6e\u6b21\u7684\u9690\u85cf\u72b6\u6001\u77e9\u9635\u9996\u5c3e\u62fc \u63a5\uff0c\u4fbf\u53ef\u5f97\u5230\u4e24\u79cd\u4e0d\u540c\u7684\u878d\u5408\u7279\u5f81\u77e9\u9635--agr self M \u548c agr oss M cr \u3002\u7279\u5f81\u878d\u5408\u7f51\u7edc\u7684\u672b\u5c3e\u90e8\u5206\u4e0e\u4e4b\u524d\u7684 \u76f8\u5173\u7814\u7a76\u4fdd\u6301\u4e00\u81f4\uff0c\u91c7\u7528\u5168\u8fde\u63a5\u795e\u7ecf\u7f51\u7edc\u5904\u7406 agr self M \u548c",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            },
            {
                "text": "T i \u548c F i \u5f97\u5230 2~10 \u4e2a\u5019\u9009\u56de\u590d\uff0c\u8bb0\u4f5c 2 10 1 [ ] i i i R r \uf0a3 \uf0a3 \uf03d \uf03d \u3002\u7edf\u8ba1 R \u4e2d\u6bcf\u4e2a\u5019\u9009\u56de\u590d i r \u7684\u5355\u8bcd\u6570 r i n \uff0c\u82e5\u6ee1\u8db3 max W n r i \uf0a3 (11) \u8868\u793a i r \u4e2d\u7684\u5355\u8bcd\u6570\u7b26\u5408\u5e38\u89c4\uff0c\u5176\u4e2d max W \u662f\u672c\u6587\u81ea\u884c\u8bbe\u7f6e\u7684\u5355\u8bcd\u5904\u7406\u6700\u5927\u503c\uff0c\u672c\u6587\u8bbe\u4e3a 100\u3002 \u672c\u6587\u5229\u7528 C \u4e2d\u7684<EOS>\u5c06 C \u5206\u5272\u6210\u591a\u4e2a\u5355\u8f6e\u5bf9\u8bdd\uff0c \u540c\u65f6\u6839\u636e<EOS>\u7684\u4e2a\u6570\u5feb\u901f\u7edf\u8ba1 C \u7684\u8f6e\u6b21\u6570\uff0c \u5f97\u5230 1 [ ] i t i i C c \uf03d \uf03d \uf03d \uff0c\u5176\u4e2d t \u662f\u8f6e\u6b21\u6570\uff0c\u82e5\u6ee1\u8db3 max T t \uf0a3 (12) \u8868\u793a C \u7684\u8f6e\u6b21\u6570\u7b26\u5408\u5e38\u89c4\uff0c\u5176\u4e2d max T \u662f\u672c\u6587\u81ea\u884c\u8bbe\u7f6e\u7684\u8f6e\u6b21\u5904\u7406\u6700\u5927\u503c\u3002\u6b64\u5916\uff0c\u8fd8\u9700\u9488\u5bf9 \u6ee1\u8db3\u6761\u4ef6\u7684 C \u7edf\u8ba1\u5176\u4e2d i c \u7684\u5355\u8bcd\u6570 c i n \uff0c\u82e5\u6ee1\u8db3 max min W n W c i \uf0a3 \uf0a3 (13) \u8868\u793a c i n \u4e2d\u7684\u5355\u8bcd\u6570\u7b26\u5408\u5e38\u89c4\uff0c\u5176\u4e2d ] , [ max min W W \u662f\u5355\u8bcd\u5904\u7406\u9608\u503c\u533a\u95f4\uff0c max W \u4e0e\u672c\u6587\u5bf9 r i n \u7684 \u9650\u5236\u76f8\u540c\uff0c min W \u662f\u7279\u522b\u9488\u5bf9 i c \u8bbe\u7f6e\u7684\u5355\u8bcd\u5904\u7406\u6700\u5c0f\u503c\uff0c\u672c\u6587\u8bbe\u4e3a 2\u3002 \u968f\u540e\uff0c\u672c\u6587\u5c06\u6e05\u6d17\u540e\u5f97\u5230\u7684\u6570\u636e\u8fdb\u884c\u6570\u636e\u89c4\u8303\u3002\u9488\u5bf9\u5019\u9009\u56de\u590d 2 10 1 [ ] i i i R r \uf0a3 \uf0a3 \uf03d \uf03d \uff0c\u672c\u6587\u5728 ] , [ max min W W \u533a\u95f4\u5185\u8bbe\u7f6e\u4e00\u4e2a\u8868\u8fbe\u957f\u5ea6\u503c W\uff0c\u901a\u8fc7\u5904\u7406\u6240\u6709 i r \uff0c\u4f7f\u5176\u6ee1\u8db3 W n r i \uf03d (14) \u82e5 r i n W \uf03c \uff0c\u5c06\u82e5\u5e72\u7279\u6b8a\u6807\u8bc6\u7b26<PAD>\u589e\u52a0\u81f3 i r \u5c3e\u90e8\uff0c\u4f7f\u5176\u957f\u5ea6\u8fbe\u6807\u3002<PAD>\u672c\u8eab\u4e0d\u5177\u5907\u8bed \u4e49\uff0c\u56e0\u6b64\u4e5f\u4e0d\u4f1a\u5f71\u54cd\u5230 i r \u7684\u8bed\u4e49\u3002\u82e5 r i n W \uf03e \uff0c\u5220\u9664\u8d85\u51fa\u90e8\u5206\u7684\u5355\u8bcd\u3002 \u7ecf\u8fc7\u4e0a\u8ff0\u5904\u7406\uff0c\u6bcf\u4e2a\u591a\u8f6e\u5bf9\u8bdd\u90fd\u7531 T \u8f6e\u5bf9\u8bdd\u7ec4\u6210\uff0c\u6bcf\u8f6e\u5bf9\u8bdd\u548c\u6bcf\u4e2a\u5019\u9009\u56de\u590d\u90fd\u7531 W \u4e2a\u5355\u8bcd \u7ec4\u6210\u3002\u501f\u52a9 vocab.txt \u4e2d\u5355\u8bcd\u4e0e\u8bcd\u5e8f\u4e4b\u95f4\u7684\u5bf9\u5e94\u5173\u7cfb\u8868\uff0c\u672c\u6587\u5c06\u6570\u636e\u4e2d\u7684\u6240\u6709\u5355\u8bcd(\u5305\u62ec<EOS>) \u8f6c\u6362\u4e3a\u6570\u5b57\uff0c\u5b9e\u73b0\u6587\u672c\u6570\u636e\u7684\u5411\u91cf\u5316\u8fc7\u7a0b\u3002 4.2 \u57fa\u4e8e\u610f\u56fe\u8bc6\u522b\u7684\u8bcd\u5411\u91cf\u589e\u5f3a \u5f97\u5230\u4e0a\u8ff0\u89c4\u8303\u6570\u636e\uff0c\u5176\u4e2d\u5c06\u5355\u8bcd\u8f6c\u5316\u6210\u7684\u6570\u5b57\u540c\u65f6\u4e0e word2vec.txt \u6587\u4ef6\u4e2d\u7684\u9884\u8bad\u7ec3\u8bcd\u5411\u91cf \u4e00\u4e00\u5bf9\u5e94\uff0c\u6839\u636e\u8fd9\u79cd\u5bf9\u5e94\u5173\u7cfb\u53ef\u5c06\u6bcf\u4e2a\u6570\u5b57\u8f6c\u6362\u6210 200 \u7ef4\u7684\u8bcd\u5411\u91cf\uff0c\u5176\u4e2d<PAD>\u7684\u8bcd\u5411\u91cf\u662f 200 \u7ef4\u7684\u96f6\u5411\u91cf\u3002 i r \u548c i c \u5747\u662f\u7ef4\u5ea6\u4e3a ) 200 , (W \u7684\u5411\u91cf\uff0cC \u7684\u7ef4\u5ea6\u662f ) 200 , , ( W T \u3002 \u53d7\u4f4d\u7f6e\u7f16\u7801\u7684\u542f\u53d1\uff0c\u672c\u6587\u5f15\u5165\u610f\u56fe\u5d4c\u5165\u5411\u91cf\u4ee5\u5904\u7406\u610f\u56fe\u8bc6\u522b\u7ed3\u679c\u3002\u9488\u5bf9\u5019\u9009\u56de\u590d\u96c6\u5408 2 10 1 [ ] i i i R r \uf0a3 \uf0a3 \uf03d \uf03d \uff0c\u53d6\u5176\u6b63\u786e\u5019\u9009\u56de\u590d\u96c6\u5408 T R \uff0c\u5bf9\u542b\u6709\u67d0\u79cd\u610f\u56fe\u7684 C \u548c T R \u540c\u65f6\u91c7\u7528\u5982\u4e0b\u7684\u7f16\u7801\u65b9\u5f0f\uff1a ) 10000 / 1 sin( 10 1 ) ( 1 dm d a ID \uf02d \uf03d (15) ) 10000 / 1 sin( 10 1 ) ( m d d b ID \uf03d (16) \u672c\u6587\u5c06 200 \u7ef4\u7684\u610f\u56fe\u5d4c\u5165\u5411\u91cf\u76f4\u63a5\u4e0e\u76f8\u5e94\u7684\u8bcd\u5411\u91cf\u76f8\u52a0\uff0c\u7531<PAD>\u6784\u6210\u7684\u5355\u8bcd\u548c\u5bf9\u8bdd\u4fdd\u6301 \u4e0d\u53d8\uff0c\u5f97\u5230\u7684\u7ed3\u679c\u5373\u4e3a\u8bcd\u5411\u91cf\u589e\u5f3a\u7684\u7ed3\u679c\u3002\u8fd9\u6837\u8bbe\u8ba1\u7684\u597d\u5904\u662f\uff0c\u5bf9\u4e8e\u76f8\u540c\u610f\u56fe\u7684 C \u548c T R \uff0c\u4e8c\u8005 \u7684\u610f\u56fe\u5d4c\u5165\u5411\u91cf\u5b8c\u5168\u76f8\u540c\uff0c\u589e\u5f3a\u4e86\u4e8c\u8005\u7684\u76f8\u5173\u7a0b\u5ea6\u3002\u6b64\u5916\uff0c\u9884\u8bad\u7ec3\u8bcd\u5411\u91cf\u7684\u6570\u503c\u8303\u56f4\u662f 0~1\uff0c \u800c\u610f\u56fe\u5d4c\u5165\u5411\u91cf\u7684\u6570\u503c\u8303\u56f4\u662f 0~0.1\uff0c\u4e8c\u8005\u76f4\u63a5\u76f8\u52a0\u7684\u6570\u636e\u53d8\u5316\u5bf9\u9884\u8bad\u7ec3\u8bcd\u5411\u91cf\u5f71\u54cd\u5f88\u5c0f\u3002 4.3 \u5b9e\u9a8c\u8bbe\u7f6e\u4e0e\u7ed3\u679c\u5206\u6790 4.3.1 \u8bc4\u4ef7\u6307\u6807\u53ca\u5b9e\u9a8c\u8bbe\u7f6e \u672c\u6587\u91c7\u7528\u68c0\u7d22\u5f0f\u591a\u8f6e\u5bf9\u8bdd\u7cfb\u7edf\u5e38\u7528\u7684\u51e0\u79cd\u8bc4\u4ef7\u6307\u6807\uff0c\u7528\u6765\u8861\u91cf\u6a21\u578b\u7684\u6027\u80fd\u3002\u5047\u8bbe\u591a\u8f6e\u5bf9\u8bdd \u6570\u636e\u96c6 C \u7531 N \u4e2a\u96c6\u5408 c \u7ec4\u6210\uff0c\u6bcf\u4e2a\u96c6\u5408 c \u5305\u542b\u6b63\u786e\u56de\u590d t \u4e2a\u3001\u9519\u8bef\u56de\u590d f \u4e2a\u3002\u5728\u6574\u4e2a\u6570\u636e\u96c6\u4e0a \u8ba1\u7b97\u5404\u9879\u8bc4\u4ef7\u6307\u6807\u7684\u5e73\u5747\u503c\uff0c\u5f97\u5230\u5e73\u5747\u7cbe\u5ea6\u5747\u503c(Mean Average Precision\uff0cMAP) \u3001\u5012\u6570\u6392\u5e8f\u5747 \u503c (Mean Reciprocal Rank\uff0c MRR) \u3001 \u9996\u4f4d\u51c6\u786e\u7387 (Precision-at-1\uff0c P@1) \u548c\u8ba1\u7b97\u53ec\u56de\u7387 (Recall-n-at-k\uff0c k R n @ ) \uff1a \uf0e5 \uf0ce \uf03d ) ( 1 c AP N MAP C c (17) \uf0e5 \uf0ce \uf03d ) ( 1 c RR N MRR C c (18) \uf0e5 \uf0ce \uf03d @1(c) 1 1 @ P N P C c (19) ) ( @ 1 @ c k R N k R n C c n \uf0e5 \uf0ce \uf03d (20) \u5176\u4e2d AP \u4e3a\u5e73\u5747\u7cbe\u5ea6(Average Precision) \uff0cRR \u4e3a\u5012\u6570\u6392\u5e8f\u6307\u6570(Reciprocal Rank) \u3002 \u672c\u6b21\u5b9e\u9a8c\u4e2d\u4f7f\u7528\u7684\u6570\u636e\u5747\u5df2\u7ecf\u8fc7\u6570\u636e\u9884\u5904\u7406\uff0c\u8bad\u7ec3\u8fc7\u7a0b\u76f8\u5173\u914d\u7f6e\u4f7f\u7528 Adam \u4f18\u5316\u5668\u8c03\u8282\u6a21 \u578b\u53c2\u6570\uff0cDAM \u6a21\u578b\u548c Ex-DAM \u6a21\u578b\u8d85\u53c2\u6570\u7684\u53d6\u503c\u5747\u5982\u8868 2 \u6240\u793a\u3002 \u8868 2.DAM(Ex-DAM)",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            }
        ],
        "back_matter": [],
        "bib_entries": {
            "BIBREF1": {
                "ref_id": "b1",
                "title": "Modern information retrieval",
                "authors": [
                    {
                        "first": "R",
                        "middle": [],
                        "last": "Baeza-Yates",
                        "suffix": ""
                    },
                    {
                        "first": "Ribeiro-Neto B ;",
                        "middle": [],
                        "last": "\u9648\u6668",
                        "suffix": ""
                    },
                    {
                        "first": "\u4e25\u777f",
                        "middle": [],
                        "last": "\u6731\u6674\u6674",
                        "suffix": ""
                    },
                    {
                        "first": "\u7b49",
                        "middle": [],
                        "last": "",
                        "suffix": ""
                    }
                ],
                "year": 1999,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Baeza-Yates R, Ribeiro-Neto B. Modern information retrieval[M]. New York: ACM press, 1999. \u9648\u6668, \u6731\u6674\u6674, \u4e25\u777f, \u7b49. \u57fa\u4e8e\u6df1\u5ea6\u5b66\u4e60\u7684\u5f00\u653e\u9886\u57df\u5bf9\u8bdd\u7cfb\u7edf\u7814\u7a76\u7efc\u8ff0[J].",
                "links": null
            },
            "BIBREF3": {
                "ref_id": "b3",
                "title": "Deep Sparse Rectifier Neural Networks[C]. international conference on artificial intelligence and statistics",
                "authors": [
                    {
                        "first": "X",
                        "middle": [],
                        "last": "Glorot",
                        "suffix": ""
                    },
                    {
                        "first": "A",
                        "middle": [],
                        "last": "Bordes",
                        "suffix": ""
                    },
                    {
                        "first": "Y",
                        "middle": [],
                        "last": "Bengio",
                        "suffix": ""
                    }
                ],
                "year": 2011,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "315--323",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Glorot X, Bordes A, Bengio Y, et al. Deep Sparse Rectifier Neural Networks[C]. international conference on artificial intelligence and statistics, 2011: 315-323.",
                "links": null
            },
            "BIBREF4": {
                "ref_id": "b4",
                "title": "Deep Residual Learning for Image Recognition",
                "authors": [
                    {
                        "first": "K",
                        "middle": [],
                        "last": "He",
                        "suffix": ""
                    },
                    {
                        "first": "X",
                        "middle": [],
                        "last": "Zhang",
                        "suffix": ""
                    },
                    {
                        "first": "S",
                        "middle": [],
                        "last": "Ren",
                        "suffix": ""
                    }
                ],
                "year": 2016,
                "venue": "computer vision and pattern recognition",
                "volume": "",
                "issue": "",
                "pages": "770--778",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "He K, Zhang X, Ren S, et al. Deep Residual Learning for Image Recognition[C]. computer vision and pattern recognition, 2016: 770-778.",
                "links": null
            },
            "BIBREF5": {
                "ref_id": "b5",
                "title": "A Method for Stochastic Optimization",
                "authors": [
                    {
                        "first": "D P",
                        "middle": [],
                        "last": "Kingma",
                        "suffix": ""
                    },
                    {
                        "first": "J",
                        "middle": [],
                        "last": "Ba",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Adam",
                        "suffix": ""
                    }
                ],
                "year": 2014,
                "venue": "J]. Computer ence",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Kingma D P, Ba J. Adam: A Method for Stochastic Optimization[J]. Computer ence, 2014",
                "links": null
            },
            "BIBREF6": {
                "ref_id": "b6",
                "title": "Spatiotemporal Features with 3D Convolutional Networks[C]. international conference on computer vision",
                "authors": [
                    {
                        "first": "D",
                        "middle": [],
                        "last": "Tran",
                        "suffix": ""
                    },
                    {
                        "first": "L",
                        "middle": [],
                        "last": "Bourdev",
                        "suffix": ""
                    },
                    {
                        "first": "R",
                        "middle": [],
                        "last": "Fergus",
                        "suffix": ""
                    }
                ],
                "year": 2015,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "4489--4497",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Tran D, Bourdev L, Fergus R, et al. Learning Spatiotemporal Features with 3D Convolutional Networks[C]. international conference on computer vision, 2015: 4489-4497.",
                "links": null
            },
            "BIBREF7": {
                "ref_id": "b7",
                "title": "The TREC-8 question answering track",
                "authors": [
                    {
                        "first": "E",
                        "middle": [],
                        "last": "Voorhees",
                        "suffix": ""
                    }
                ],
                "year": 1999,
                "venue": "",
                "volume": "99",
                "issue": "",
                "pages": "77--82",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Voorhees E M. The TREC-8 question answering track report[C]//Trec. 1999, 99: 77-82.",
                "links": null
            },
            "BIBREF8": {
                "ref_id": "b8",
                "title": "Sequential matching network: A new architecture for multi-turn response selection in retrieval-based chatbots",
                "authors": [
                    {
                        "first": "Y",
                        "middle": [],
                        "last": "Wu",
                        "suffix": ""
                    },
                    {
                        "first": "W",
                        "middle": [],
                        "last": "Wu",
                        "suffix": ""
                    },
                    {
                        "first": "C",
                        "middle": [],
                        "last": "Xing",
                        "suffix": ""
                    }
                ],
                "year": 2016,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {
                    "arXiv": [
                        "arXiv:1612.01627"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Wu Y, Wu W, Xing C, et al. Sequential matching network: A new architecture for multi-turn response selection in retrieval-based chatbots[J]. arXiv preprint arXiv:1612.01627, 2016.",
                "links": null
            },
            "BIBREF9": {
                "ref_id": "b9",
                "title": "Modeling multi-turn conversation with deep utterance aggregation",
                "authors": [
                    {
                        "first": "Z",
                        "middle": [],
                        "last": "Zhang",
                        "suffix": ""
                    },
                    {
                        "first": "J",
                        "middle": [],
                        "last": "Li",
                        "suffix": ""
                    },
                    {
                        "first": "P",
                        "middle": [],
                        "last": "Zhu",
                        "suffix": ""
                    }
                ],
                "year": 2018,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {
                    "arXiv": [
                        "arXiv:1806.09102"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Zhang Z, Li J, Zhu P, et al. Modeling multi-turn conversation with deep utterance aggregation[J]. arXiv preprint arXiv:1806.09102, 2018.",
                "links": null
            },
            "BIBREF10": {
                "ref_id": "b10",
                "title": "Multi-turn response selection for chatbots with deep attention matching network",
                "authors": [
                    {
                        "first": "X",
                        "middle": [],
                        "last": "Zhou",
                        "suffix": ""
                    },
                    {
                        "first": "L",
                        "middle": [],
                        "last": "Li",
                        "suffix": ""
                    },
                    {
                        "first": "D",
                        "middle": [],
                        "last": "Dong",
                        "suffix": ""
                    }
                ],
                "year": null,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Zhou X, Li L, Dong D, et al. Multi-turn response selection for chatbots with deep attention matching network[C].",
                "links": null
            },
            "BIBREF11": {
                "ref_id": "b11",
                "title": "Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics",
                "authors": [],
                "year": 2018,
                "venue": "",
                "volume": "1",
                "issue": "",
                "pages": "1118--1127",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2018: 1118-1127.",
                "links": null
            }
        },
        "ref_entries": {
            "TABREF0": {
                "type_str": "table",
                "num": null,
                "content": "<table><tr><td>2 DAM \u6a21\u578b</td></tr></table>",
                "text": "\u63d0\u51fa\u6df1\u5ea6\u6ce8\u610f\u529b\u5339\u914d(Deep Attention Matching\uff0cDAM) \u6a21\u578b\u3002\u5728 SMN \u6a21\u578b\u7684\u57fa\u7840\u4e0a\uff0c\u7701\u53bb CNN \u548c RNN \u7b49\u7ed3\u6784\uff0c\u4ec5\u4f9d\u9760\u6ce8\u610f\u529b\u673a\u5236\u5b8c\u6210\u591a\u8f6e\u5bf9\u8bdd\u7684 \u56de\u590d\u9009\u62e9\uff0c\u4f7f\u6a21\u578b\u53c2\u6570\u5927\u5e45\u5ea6\u51cf\u5c11\uff0c\u5927\u5e45\u63d0\u5347\u4e86\u8bad\u7ec3\u901f\u5ea6\u3002\u800c\u8fd9\u7c7b\u65b9\u6cd5\u7684\u5c40\u9650\u6027\u5728\u4e8e\u5019\u9009\u96c6\u4e2d \u88ab\u9009\u5b9a\u7684\u56de\u590d\u4ec5\u9002\u7528\u4e8e\u672c\u8f6e\u5bf9\u8bdd\uff0c\u4e0e\u4e0a\u4e0b\u6587\u5e76\u4e0d\u80fd\u5f62\u6210\u826f\u597d\u7684\u5339\u914d\uff0c\u6216\u5728\u5339\u914d\u6a21\u578b\u4e2d\u6ca1\u6709\u5b66\u4e60 \u5230\u771f\u6b63\u7684\u8bed\u4e49\u5173\u7cfb\uff0c\u5bf9\u591a\u8f6e\u5bf9\u8bdd\u7684\u5185\u5bb9\u4ea7\u751f\u4e86\u6df7\u6dc6\uff0c\u96be\u4ee5\u9009\u62e9\u6b63\u786e\u7684\u5019\u9009\u56de\u590d\u3002 \u5728\"\u8868\u793a-\u5339\u914d-\u878d\u5408\"\u8fd9\u4e00\u6846\u67b6\u4e0b\uff0c\u540c\u65f6\u4f18\u5316\u4e09\u4e2a\u90e8\u5206\u662f\u73b0\u9636\u6bb5\u7684\u7814\u7a76\u96be\u70b9\u3002\u5728\u524d\u4eba\u7814\u7a76 \u7684\u57fa\u7840\u4e0a\uff0c\u672c\u6587\u5bf9\u6bd4 DAM \u6a21\u578b\uff0c\u901a\u8fc7\u5f15\u5165\u591a\u5934\u6ce8\u610f\u529b\u673a\u5236\uff0c\u4f7f\u6a21\u578b\u66f4\u9002\u5408\u5904\u7406\u542b\u6709\u7ec6\u5fae\u53d8\u5316 \u7684\u6570\u636e\uff0c\u80fd\u8ba9\u9009\u5b9a\u7684\u76ee\u6807\u5019\u9009\u56de\u590d\u4e0e\u4e0a\u4e0b\u6587\u5f62\u6210\u826f\u597d\u7684\u5339\u914d\u5173\u7cfb\u3002\u6b64\u5916\uff0c\u672c\u6587\u5728\u7279\u5f81\u878d\u5408\u8fc7\u7a0b \u4e2d\u91c7\u7528 BiLSTM \u6a21\u578b\uff0c\u901a\u8fc7\u6355\u83b7\u591a\u8f6e\u5bf9\u8bdd\u4e2d\u7684\u5e8f\u5217\u4f9d\u8d56\u5173\u7cfb\uff0c\u5e2e\u52a9\u6a21\u578b\u5efa\u7acb\u6bcf\u8f6e\u5bf9\u8bdd\u4e0e\u524d\u4e00\u8f6e \u5bf9\u8bdd\u3001\u5019\u9009\u56de\u590d\u4e4b\u95f4\u7684\u5339\u914d\u4fe1\u606f\uff0c\u4f7f\u5339\u914d\u6a21\u578b\u5b66\u4e60\u5230\u771f\u6b63\u7684\u8bed\u4e49\u5173\u7cfb\uff0c\u8fdb\u4e00\u6b65\u63d0\u9ad8\u9009\u62e9\u76ee\u6807\u5019 \u9009\u56de\u590d\u7684\u51c6\u786e\u7387\uff0c\u57fa\u4e8e\u6b64\u5efa\u7acb\u57fa\u4e8e\u591a\u5934\u6ce8\u610f\u529b\u548c BiLSTM \u6539\u8fdb\u7684 DAM \u6a21\u578b Ex-DAM\uff0c\u5728\u8c46\u74e3\u3001 \u7535\u5546\u8fd9\u4e24\u4e2a\u591a\u8f6e\u4e2d\u6587\u5bf9\u8bdd\u6570\u636e\u96c6\u4e0a\u8fdb\u884c\u7814\u7a76\u3002 \u8bba\u6587\u7684\u7ed3\u6784\u5b89\u6392\u5982\u4e0b\uff1a\u7b2c 1 \u8282\u4ecb\u7ecd\u4e86\"\u68c0\u7d22\u5f0f\u4eba\u673a\u591a\u8f6e\u5bf9\u8bdd\"\u7684\u6982\u5ff5\u53ca\u7279\u70b9\uff0c\u6982\u8ff0\u4e86\u8fd1\u51e0\u5e74 \u6df1\u5ea6\u5b66\u4e60\u6a21\u578b\u65b9\u6cd5\uff1b\u7b2c 2 \u8282\u4ecb\u7ecd\u4e86\u6df1\u5ea6\u6ce8\u610f\u529b\u673a\u5236\u6a21\u578b DAM \u7684\u6574\u4f53\u7ed3\u6784\uff1b\u7b2c 3 \u8282\u4ecb\u7ecd\u57fa\u4e8e\u591a \u5934\u6ce8\u610f\u529b\u548c BiLSTM \u6539\u8fdb\u7684 DAM \u6a21\u578b Ex-DAM\uff0c\u4e3b\u8981\u5305\u62ec\u591a\u5934\u6ce8\u610f\u529b\u6a21\u5757\u3001\u8bed\u4e49\u8868\u793a\u7f51\u7edc\u548c \u53cc\u901a\u9053 BiLSTM \u7279\u5f81\u878d\u5408\u7f51\u7edc\uff1b \u7b2c 4 \u8282\u4ecb\u7ecd\u5b9e\u9a8c\u6570\u636e\u3001 \u5b9e\u9a8c\u5185\u5bb9\u3001 \u5b9e\u9a8c\u7ed3\u679c\u53ca\u5206\u6790\uff0c \u9a8c\u8bc1 Ex-DAM \u6a21\u578b\u7684\u6709\u6548\u6027\uff1b\u6700\u540e\u5728\u7b2c 5 \u8282\u8fdb\u884c\u603b\u7ed3\u3002",
                "html": null
            },
            "TABREF2": {
                "type_str": "table",
                "num": null,
                "content": "<table><tr><td colspan=\"11\">4.3.2 DAM \u6a21\u578b\u5b9e\u9a8c\u7ed3\u679c \u7531\u8868 4 \u548c\u8868 5 \u770b\u51fa\uff0c\u65e0\u8bba\u662f\u5426\u5bf9\u6570\u636e\u96c6\u8fdb\u884c\u8bcd\u5411\u91cf\u589e\u5f3a\uff0cEx-DAM \u6a21\u578b\u7684\u5b9e\u9645\u8868\u73b0\u90fd\u4f18\u4e8e</td></tr><tr><td colspan=\"11\">\u57fa\u7ebf\u6a21\u578b\u3002\u5373\u4f7f\u8bcd\u5411\u91cf\u589e\u5f3a\u66fe\u5728\u4e4b\u524d\u7684\u5b9e\u9a8c\u4e2d\u5bfc\u81f4\u57fa\u7ebf\u6a21\u578b\u7684\u6027\u80fd\u4e0d\u5347\u53cd\u964d\uff0c\u5374\u5e2e\u52a9 Ex-DAM</td></tr><tr><td colspan=\"11\">\u672c\u6587\u5148\u5728\u4e24\u4e2a\u6570\u636e\u96c6\u7684\u8bad\u7ec3\u96c6\u4e0a\u8bad\u7ec3 DAM \u6a21\u578b\uff0c\u7136\u540e\u5728\u6d4b\u8bd5\u96c6\u4e0a\u8bc4\u4f30\u6a21\u578b\u7684\u6027\u80fd\uff0c\u5b9e\u9a8c \u6a21\u578b\u8fbe\u5230\u4e86\u6700\u4f73\u6027\u80fd\uff0c\u8bf4\u660e\u591a\u5934\u6ce8\u610f\u529b\u673a\u5236\u4e0e BiLSTM \u7684\u5171\u540c\u4f5c\u7528\u8981\u4f18\u4e8e\u666e\u901a\u81ea\u6ce8\u610f\u529b\u673a\u5236\u3002</td></tr><tr><td colspan=\"11\">\u7ed3\u679c\u5982\u8868 3 \u6240\u793a\u3002 \u5728\u8868 4 \u4e2d\uff0cEx-DAM5 \u6a21\u578b\u6027\u80fd\u4f18\u4e8e\u5176\u4f59\u6a21\u578b\uff0c\u8be5\u6a21\u578b\u4e0e\u5176\u4f59\u6a21\u578b\u7684\u4e0d\u540c\u4e4b\u5904\u5728\u4e8e\u8bed\u4e49\u8868\u793a</td></tr><tr><td colspan=\"11\">\u8868 3. DAM \u6a21\u578b\u5b9e\u9a8c\u7ed3\u679c \u7f51\u7edc\u7684\u8f93\u51fa\u542b\u6709 5 \u79cd\u7c92\u5ea6\u7684\u8bed\u4e49\u8868\u793a\u3002\u82e5\u5c06\u6700\u5e95\u5c42\u7c92\u5ea6\u8bed\u4e49\u8868\u793a\u53bb\u9664\u6216\u4ee5\u539f\u59cb\u8f93\u5165\u66ff\u6362\u6700\u5e95\u5c42</td></tr><tr><td colspan=\"11\">\u8c46\u74e3\u5bf9\u8bdd\u6570\u636e\u96c6 \u8bed\u4e49\u8868\u793a\uff0c\u90fd\u5c06\u635f\u5931\u4e00\u90e8\u5206\u6a21\u578b\u6027\u80fd\u3002\u7136\u800c\u5728\u8868 5 \u4e2d\uff0c\u4ee5\u539f\u59cb\u8f93\u5165\u66ff\u6362\u6700\u5e95\u5c42\u8bed\u4e49\u8868\u793a\u7684 \u7535\u5546\u5bf9\u8bdd\u6570\u636e\u96c6 \u5b9e\u9a8c\u5185\u5bb9 MAP MRR P@1 R10@1 R10@1 R10@2 R10@5 Ex-DAM5-0-4 \u6a21\u578b\u5374\u6bd4 Ex-DAM5 \u6a21\u578b\u6027\u80fd\u66f4\u4f18\u3002\u8fd9\u662f\u7531\u4e8e\u5bf9\u539f\u59cb\u8f93\u5165\u8fdb\u884c\u4e86\u8bcd\u5411\u91cf\u589e\u5f3a\uff0c\u5bfc\u81f4</td></tr><tr><td colspan=\"11\">DAM\u6a21\u578b \u539f\u59cb\u8f93\u5165\u542b\u6709\u989d\u5916\u7684\u610f\u56fe\u7279\u5f81\uff0c\u800c\u8bed\u4e49\u8868\u793a\u7f51\u7edc\u4e2d\u6bcf\u4e2a\u7c92\u5ea6\u7684\u8bed\u4e49\u8868\u793a\u90fd\u6e90\u4e8e\u539f\u59cb\u8f93\u5165\uff0c\u76f8\u5f53 0.550 0.601 0.427 0.254 0.406 0.547 0.810</td></tr><tr><td colspan=\"8\">+\u8bcd\u5411\u91cf\u589e\u5f3a \u4e8e\u5728\u8ba1\u7b97\u8fc7\u7a0b\u4e2d\u4e0d\u65ad\u5f3a\u5316\u8fd9\u79cd\u610f\u56fe\u7279\u5f81\uff0c\u4fc3\u4f7f\u6a21\u578b\u91cd\u70b9\u5bf9\u610f\u56fe\u7279\u5f81\u5efa\u6a21\u3002 0.539 0.583 0.409 0.238 0.392</td><td/><td>0.530</td><td>0.798</td></tr><tr><td colspan=\"11\">+\u6570\u636e\u9884\u5904\u7406 \u672c\u6587\u8fdb\u884c\u7684\u5b9e\u9a8c\u5747\u5c06\u5806\u53e0\u6ce8\u610f\u529b\u6a21\u5757\u6570\u8bbe\u7f6e\u4e3a 5\uff0c\u4e3a\u63a2\u7a76 Ex-DAML-0-m \u6a21\u578b\u4e2d L \u548c m \u7684\u53d6 0.556 0.606 0.434 0.259 0.414 0.557 0.819</td></tr><tr><td colspan=\"11\">+\u6570\u636e\u9884\u5904\u7406+\u8bcd\u5411\u91cf\u589e\u5f3a \u503c\u5bf9\u6a21\u578b\u6027\u80fd\u7684\u5f71\u54cd\uff0c\u672c\u6587\u4f7f\u7528\u7ecf\u6570\u636e\u9884\u5904\u7406\u548c\u8bcd\u5411\u91cf\u589e\u5f3a\u7684\u7535\u5546\u5bf9\u8bdd\u6570\u636e\u96c6\u8fdb\u884c\u4e86\u989d\u5916\u7684\u5b9e 0.548 0.587 0.425 0.248 0.396 0.539 0.807</td></tr><tr><td colspan=\"11\">\u7531\u8868 3 \u53ef\u4ee5\u770b\u51fa\uff0c\u6570\u636e\u9884\u5904\u7406\u6709\u52a9\u4e8e\u6a21\u578b\u6027\u80fd\u7684\u63d0\u5347\uff0c\u5bf9\u539f\u59cb\u6570\u636e\u96c6\u8fdb\u884c\u6570\u636e\u9884\u5904\u7406\uff0c\u5728 \u9a8c\uff0c\u5c06\u8bc4\u4ef7\u6307\u6807 R10@1 \u7ed3\u679c\u7ed8\u5236\u6210\u6298\u7ebf\u56fe 4 \u6240\u793a\u3002</td></tr><tr><td colspan=\"11\">\u8c46\u74e3\u5bf9\u8bdd\u6570\u636e\u96c6\u4e0a\u7684\u5404\u9879\u8bc4\u4ef7\u6307\u6807\u83b7\u5f97\u4e86 0.5%~0.7%\u7684\u63d0\u5347\uff0c\u5728\u7535\u5546\u5bf9\u8bdd\u6570\u636e\u96c6\u4e0a\u7684\u5404\u9879\u8bc4\u4ef7</td></tr><tr><td colspan=\"11\">\u6307\u6807\u83b7\u5f97\u4e86 0.8%~1%\u7684\u63d0\u5347\u3002\u8fd8\u53ef\u4ee5\u770b\u51fa\uff0c\u5c06\u57fa\u4e8e\u610f\u56fe\u8bc6\u522b\u7684\u8bcd\u5411\u91cf\u589e\u5f3a\u76f4\u63a5\u5e94\u7528\u4e8e DAM \u6a21</td></tr><tr><td colspan=\"11\">\u578b\u4ea7\u751f\u4e86\u4e0d\u7406\u60f3\u7684\u6548\u679c\uff0c\u5728\u4e24\u4e2a\u6570\u636e\u96c6\u4e0a\u7684\u5404\u9879\u8bc4\u4ef7\u6307\u6807\u5747\u4e0b\u964d\u4e86 1%\u4ee5\u4e0a\uff0c\u5373\u4f7f\u7ecf\u8fc7\u6570\u636e\u9884\u5904</td></tr><tr><td colspan=\"11\">\u7406\uff0c\u6a21\u578b\u6548\u679c\u6709\u4e86\u7565\u5fae\u63d0\u5347\uff0c\u4f46\u4e5f\u59cb\u7ec8\u4f4e\u4e8e\u57fa\u7ebf\u6c34\u5e73\u3002\u6b64\u7ed3\u679c\u7684\u4ea7\u751f\u539f\u56e0\u4e3b\u8981\u662f DAM \u6a21\u578b\u5b8c</td></tr><tr><td colspan=\"11\">\u5168\u7531\u81ea\u6ce8\u610f\u529b\u673a\u5236\u6784\u9020\u800c\u6210\uff0c\u5176\u4e2d\u7684\u8ba1\u7b97\u8fc7\u7a0b\u4f9d\u8d56\u4e8e\u8bcd\u5411\u91cf\uff0c\u672c\u6587\u5728\u8bcd\u5411\u91cf\u5c42\u9762\u8fdb\u884c\u7684\u4efb\u4f55\u6539</td></tr><tr><td colspan=\"8\">\u52a8\u90fd\u5c06\u9010\u5c42\u5e72\u6270\u6ce8\u610f\u529b\u673a\u5236\u7684\u8ba1\u7b97\uff0c\u4ece\u800c\u5bfc\u81f4 DAM \u6a21\u578b\u6027\u80fd\u6025\u5267\u4e0b\u964d\u3002</td><td/><td/></tr><tr><td colspan=\"8\">4.3.3 Ex-DAM \u6a21\u578b\u5b9e\u9a8c\u7ed3\u679c \u56fe 4. \u4e0d\u540c\u53c2\u6570\u642d\u914d\u5bf9 Ex-DAML-0-m \u6a21\u578b\u7684\u5f71\u54cd</td><td/><td/></tr><tr><td colspan=\"11\">\u672c\u6587\u5728\u5b9e\u9a8c\u4e2d\u59cb\u7ec8\u4fdd\u6301\u8bed\u4e49\u8868\u793a\u7f51\u7edc\u7684\u8f93\u51fa\u7c92\u5ea6\u4e0d\u8d85\u8fc7 5\uff0c\u8fd9\u662f\u7531\u4e8e\u6a21\u578b\u5360\u7528\u7684\u663e\u5b58\u6240\u81f4\uff0c</td></tr><tr><td colspan=\"11\">\u4e3a\u4e86\u9a8c\u8bc1\u672c\u6587\u63d0\u51fa\u7684 Ex-DAM \u6a21\u578b\u662f\u5426\u6709\u6548\uff0c\u5c06\u7ecf\u8fc7\u6570\u636e\u9884\u5904\u7406\u7684\u4e24\u79cd\u5b9e\u9a8c\u6a21\u578b\u4f5c\u4e3a\u57fa\u7ebf \u82e5\u8d85\u51fa\u6b64\u503c\u5fc5\u987b\u4fee\u6539\u8d85\u53c2\u6570\uff0c\u800c\u8ba1\u7b97\u65b9\u9762\u7684\u6d88\u8017\u5c06\u5448\u6307\u6570\u7ea7\u589e\u957f\uff0c\u5f88\u96be\u4e0e\u4e4b\u524d\u7684\u5b9e\u9a8c\u505a\u5bf9\u6bd4\u3002</td></tr><tr><td colspan=\"11\">\u6a21\u578b\uff0c\u4ece\u662f\u5426\u8fdb\u884c\u8bcd\u5411\u91cf\u589e\u5f3a\u7684\u89d2\u5ea6\u8fdb\u884c\u4e86\u72ec\u7acb\u5b9e\u9a8c\u3002\u5176\u4e2d Ex-DAM5 \u8868\u793a\u5806\u53e0\u6ce8\u610f\u529b\u6a21\u5757\u6570\u8bbe \u56fe 4 \u6807\u8bb0\u7684\u6a21\u578b\u5bf9\u6bd4\u70b9\u662f\u4e0a\u4e00\u5b9e\u9a8c\u4e2d\u7684 Ex-DAM5-0-4 \u6a21\u578b\uff0c\u7531\u56fe\u53ef\u77e5\uff0c\u591a\u7c92\u5ea6\u8bed\u4e49\u8868\u793a\u786e\u5b9e\u80fd\u5728</td></tr><tr><td colspan=\"11\">\u7f6e\u4e3a 5\uff0c\u4fdd\u7559\u6240\u6709 i \u4e00\u5b9a\u7a0b\u5ea6\u4e0a\u63d0\u5347\u6a21\u578b\u6027\u80fd\uff0c\u5f53\u6a21\u578b\u5c06 4 \u4e2a\u5806\u53e0\u591a\u5934\u6ce8\u610f\u529b\u6a21\u5757\u7684\u8f93\u51fa\u4e0e\u539f\u59cb\u8f93\u5165\u5171\u540c\u4f5c\u4e3a\u8bed\u4e49 U \u548c R\uff0c\u5373\u4f7f\u7528\u6240\u6709\u7c92\u5ea6\u7684\u8bed\u4e49\u8868\u793a\u4f5c\u4e3a\u7279\u5f81\u5339\u914d\u7f51\u7edc\u7684\u8f93\u5165\uff1bEx-DAM5-4 \u8868\u793a\u7f51\u7edc\u7684\u8f93\u51fa\u65f6\uff0c\u901a\u5e38\u80fd\u53d6\u5f97\u6700\u9ad8\u6027\u80fd\u3002\u82e5\u5806\u53e0\u591a\u5934\u6ce8\u610f\u529b\u6a21\u5757\u6570\u8d85\u8fc7 5\uff0c\u65e0\u8bba\u5982\u4f55\u9009\u62e9 m</td></tr><tr><td colspan=\"11\">\u8868\u793a\u6a21\u5757\u6570\u8bbe\u7f6e\u4e3a 5\uff0c\u4fdd\u7559 i U \u548c R \u4e2d\u7684\u540e 4 \u4e2a\u5143\u7d20\uff1bEx-DAM5-0-4 \u8868\u793a\u6a21\u5757\u6570\u8bbe\u7f6e\u4e3a 5\uff0c\u4fdd\u7559 i \u7684\u503c\uff0c\u6a21\u578b\u90fd\u5c06\u9010\u6e10\u51fa\u73b0\u8fc7\u62df\u5408\u73b0\u8c61\uff0c\u8fd9\u662f\u7531\u4e8e\u968f\u7740\u5806\u53e0\u591a\u5934\u6ce8\u610f\u529b\u6a21\u5757\u6570\u7684\u589e\u52a0\uff0c\u88ab\u591a\u5934\u6ce8 U \u610f\u529b\u673a\u5236\u91cd\u70b9\u5173\u6ce8\u7684\u4fe1\u606f\u4f1a\u4ece\u524d\u4e00\u5c42\u4e0d\u65ad\u7d2f\u52a0\u5230\u4e0b\u4e00\u5c42\uff0c\u5bfc\u81f4\u8fd9\u4e9b\u4fe1\u606f\u5728\u6df1\u5c42\u8ba1\u7b97\u8fc7\u7a0b\u4e2d\u57fa\u672c</td></tr><tr><td colspan=\"8\">\u548c R \u4e2d\u7684\u7b2c 1 \u4e2a\u5143\u7d20\u548c\u540e 4 \u4e2a\u5143\u7d20\uff0c\u5b9e\u9a8c\u7ed3\u679c\u5206\u522b\u5982\u8868 4 \u548c\u8868 5 \u6240\u793a\uff1a \u4fdd\u6301\u4e0d\u53d8\uff0c\u4e25\u91cd\u5f71\u54cd\u6a21\u578b\u8bad\u7ec3\u3002</td><td/><td/></tr><tr><td/><td/><td/><td colspan=\"5\">\u8868 4. \u4e0d\u542b\u8bcd\u5411\u91cf\u589e\u5f3a\u7684 Ex-DAM \u6a21\u578b\u5b9e\u9a8c\u7ed3\u679c</td><td/><td/></tr><tr><td>5</td><td colspan=\"2\">\u5b9e\u9a8c\u5185\u5bb9</td><td/><td colspan=\"2\">\u8c46\u74e3\u5bf9\u8bdd\u6570\u636e\u96c6</td><td/><td/><td colspan=\"2\">\u7535\u5546\u5bf9\u8bdd\u6570\u636e\u96c6</td></tr><tr><td/><td/><td/><td>MAP</td><td>MRR</td><td>P@1</td><td>R10@1</td><td colspan=\"2\">R10@1</td><td>R10@2</td><td>R10@5</td></tr><tr><td>DAM\u6a21\u578b</td><td/><td/><td>0.556</td><td>0.606</td><td>0.434</td><td>\u6a21\u578b\u8d85\u53c2\u6570\u8868 0.259</td><td>0.414</td><td/><td>0.557</td><td>0.819</td></tr><tr><td>Ex-DAM5</td><td/><td>\u8d85\u53c2\u6570</td><td>0.562</td><td>\u53c2\u6570\u542b\u4e49 0.610</td><td>0.441</td><td>0.264</td><td colspan=\"2\">\u53c2\u6570\u503c 0.423</td><td>0.563</td><td>0.822</td></tr><tr><td>Ex-DAM5-4</td><td/><td>W</td><td>0.557</td><td>\u8868\u8fbe\u957f\u5ea6\u503c 0.605</td><td>0.437</td><td>0.260</td><td>0.419</td><td>50</td><td>0.561</td><td>0.819</td></tr><tr><td colspan=\"2\">Ex-DAM5-0-4</td><td>T</td><td>0.559</td><td>\u5bf9\u8bdd\u8f6e\u6b21\u503c 0.607</td><td>0.438</td><td>0.259</td><td>0.420</td><td>9</td><td>0.561</td><td>0.820</td></tr><tr><td/><td/><td>epoch</td><td/><td colspan=\"2\">\u6570\u636e\u96c6\u8fed\u4ee3\u6b21\u6570</td><td/><td/><td>3</td><td/></tr><tr><td/><td/><td>layer</td><td colspan=\"5\">\u5806\u53e0\u591a\u5934\u6ce8\u610f\u529b\u6a21\u5757\u6570 \u8868 5. \u542b\u6709\u8bcd\u5411\u91cf\u589e\u5f3a\u7684 Ex-DAM \u6a21\u578b\u5b9e\u9a8c\u7ed3\u679c</td><td>5</td><td/></tr><tr><td/><td/><td>batch_size</td><td/><td colspan=\"2\">\u5355\u6b21\u8bad\u7ec3\u6837\u672c\u6570 \u8c46\u74e3\u5bf9\u8bdd\u6570\u636e\u96c6</td><td/><td/><td colspan=\"2\">128 \u7535\u5546\u5bf9\u8bdd\u6570\u636e\u96c6</td></tr><tr><td/><td colspan=\"2\">learning_rate \u5b9e\u9a8c\u5185\u5bb9</td><td>MAP</td><td>\u521d\u59cb\u5b66\u4e60\u7387 MRR</td><td>P@1</td><td>R10@1</td><td colspan=\"2\">0.001 R10@1</td><td>R10@2</td><td>R10@5</td></tr><tr><td>DAM\u6a21\u578b</td><td/><td>decay_step</td><td>0.548</td><td colspan=\"2\">\u5b66\u4e60\u7387\u8870\u51cf\u6b65\u957f 0.587 0.425</td><td>0.248</td><td>0.396</td><td>500</td><td>0.539</td><td>0.807</td></tr><tr><td>Ex-DAM5</td><td/><td>decay_rate</td><td>0.568</td><td colspan=\"2\">\u5b66\u4e60\u7387\u8870\u51cf\u7387 0.607 0.442</td><td>0.265</td><td>0.425</td><td>0.9</td><td>0.564</td><td>0.830</td></tr><tr><td>Ex-DAM5-4</td><td/><td/><td>0.565</td><td>0.605</td><td>0.440</td><td>0.262</td><td>0.421</td><td/><td>0.561</td><td>0.828</td></tr><tr><td colspan=\"2\">Ex-DAM5-0-4</td><td/><td>0.570</td><td>0.615</td><td>0.448</td><td>0.270</td><td>0.427</td><td/><td>0.566</td><td>0.831</td></tr></table>",
                "text": "\u7ed3\u8bba\u53ca\u540e\u7eed\u5de5\u4f5c \u672c\u6587\u63d0\u51fa\u4e86\u57fa\u4e8e\u591a\u5934\u6ce8\u610f\u529b\u548c BiLSTM \u6539\u8fdb\u7684 DAM \u6a21\u578b Ex-DAM\uff0c\u8be5\u6a21\u578b\u7528\u4e8e\u5904\u7406\u4e2d\u6587 \u591a\u8f6e\u5bf9\u8bdd\u95ee\u7b54\u5339\u914d\u95ee\u9898\u3002\u672c\u6587\u5c06 DAM \u6a21\u578b\u4f5c\u4e3a\u57fa\u7ebf\u6a21\u578b\uff0c\u5229\u7528\u591a\u5934\u6ce8\u610f\u529b\u673a\u5236\u5728\u591a\u4e2a\u4e0d\u540c\u5b50 \u7a7a\u95f4\u5185\u8ba1\u7b97\u7279\u5f81\uff0c\u4ece\u800c\u6709\u80fd\u529b\u5efa\u6a21\u8f83\u957f\u7684\u591a\u8f6e\u5bf9\u8bdd\u3002Ex-DAM \u6a21\u578b\u7684\u5377\u79ef\u6838\u4f7f\u7528 BiLSTM \u6765\u6355 \u83b7\u5e8f\u5217\u4e0a\u7684\u4f9d\u8d56\u5173\u7cfb\u3002\u5b9e\u9a8c\u8bc1\u660e\uff0cEx-DAM \u6a21\u578b\u6027\u80fd\u5728\u7535\u5546\u548c\u8c46\u74e3\u7684\u6570\u636e\u96c6\u4e0a\u5747\u4f18\u4e8e\u57fa\u7ebf\u6a21\u578b\u3002 \u5728\u672a\u6765\u7684\u7814\u7a76\u4e2d\uff0c\u5c06\u5c1d\u8bd5\u52a0\u5165\u547d\u540d\u5b9e\u4f53\u8bc6\u522b\u3001\u60c5\u611f\u5206\u6790\u7b49\u591a\u79cd\u8f85\u52a9\u624b\u6bb5\uff0c\u4f7f\u5f97 Ex-DAM \u6a21 \u578b\u53ef\u4ee5\u5728\u6587\u672c\u7247\u6bb5\u4e2d\u5c3d\u53ef\u80fd\u63d0\u53d6\u66f4\u591a\u7684\u7279\u5f81\u3002 \u81f4\u8c22 *\u901a\u4fe1\u4f5c\u8005(chongzhy@vip.sina.com) \uff0c\u672c\u6587\u627f\u56fd\u5bb6\u6559\u80b2\u90e8\u4eba\u6587\u793e\u4f1a\u79d1\u5b66\u7814\u7a76\u89c4\u5212\u57fa\u91d1\u8d44\u52a9 \u9879\u76ee(16YJAZH072) \u3001\u56fd\u5bb6\u793e\u4f1a\u79d1\u5b66\u57fa\u91d1\u91cd\u5927\u9879\u76ee(14ZDB156) \u3001\u98df\u54c1\u5b89\u5168\u5927\u6570\u636e\u6280\u672f\u5317\u4eac\u5e02 \u91cd\u70b9\u5b9e\u9a8c\u5ba4\u8d44\u52a9\u3002",
                "html": null
            }
        }
    }
}