{
    "paper_id": "O01-1002",
    "header": {
        "generated_with": "S2ORC 1.0.0",
        "date_generated": "2023-01-19T08:08:51.003920Z"
    },
    "title": "2.2 n-gram \u6a21\u578b\u4e4b\u5efa\u7acb",
    "authors": [
        {
            "first": "T",
            "middle": [],
            "last": "\u4e26\u4e14\u662f\u7531\u4e00\u6bb5\u8a5e\u5e8f\u5217",
            "suffix": "",
            "affiliation": {},
            "email": ""
        },
        {
            "first": "\u2026w",
            "middle": [
                "T"
            ],
            "last": "\u6240\u7d44\u6210\uff0c\u5247",
            "suffix": "",
            "affiliation": {},
            "email": ""
        },
        {
            "first": "S",
            "middle": [],
            "last": "\u51fa\u73fe\u7684\u6a5f\u7387\u53ef\u4ee5\u5beb\u6210",
            "suffix": "",
            "affiliation": {},
            "email": ""
        }
    ],
    "year": "",
    "venue": null,
    "identifiers": {},
    "abstract": "",
    "pdf_parse": {
        "paper_id": "O01-1002",
        "_pdf_hash": "",
        "abstract": [],
        "body_text": [
            {
                "text": "\u53ef\u80fd\u7684\u5c0d\u61c9\u6587\u53e5(sentence) S * \uff0c\u4f7f\u7528\u8c9d\u5f0f\u5206\u985e\u67b6\u69cb\u662f\u627e\u51fa\u6700\u4f73\u4e8b\u5f8c\u6a5f\u7387\u7684\u6587\u53e5 * argmax ( | ) argmax ( | ) ( ) S S S PS X PX S PS = = \u22c5 (1) \u5176\u4e2d n-gram \u6a21\u578b P(S)\u626e\u6f14\u8457\u4e8b\u524d\u6a5f\u7387\u7684\u89d2\u8272\uff0c\u900f\u904e\u8072\u5b78\u6a21\u578b\u8a08\u7b97\u53ef\u7372\u5f97\u4e00\u6bb5\u6587 \u53e5\u7684\u8072\u5b78\u6a21\u578b\u5206\u6578 P(X|S)\uff0c\u518d\u900f\u904e\u8a9e\u8a00\u6a21\u578b\u8a08\u7b97\u53ef\u7372\u5f97\u6b64\u6587\u53e5\u7684\u8a9e\u8a00\u6a21\u578b\u5206\u6578 P(S)\uff0c\u5c07\u5169\u6a5f\u7387\u76f8\u4e58\u6c42\u5f97\u6700\u4f73\u5316\u4e4b\u6587\u53e5\uff0c\u5373\u70ba\u6b64\u8072\u5b78\u8a0a\u865f\u6700\u6709\u53ef\u80fd\u4e4b\u5c0d\u61c9\u6587\u53e5\u3002 \u5c31\u6587\u4ef6\u5206\u985e\u7684\u9818\u57df\u800c\u8a00\uff0c\u7d66\u5b9a\u4e00\u7bc7\u6587\u4ef6 d\uff0c\u76ee\u6a19\u662f\u53bb\u627e\u5c0b\u6b64\u7bc7\u6587\u4ef6\u6240\u5c6c\u7684\u985e \u5225 c (category)\uff0c\u5047\u8a2d\u6211\u5011\u7e3d\u5171\u5b9a\u7fa9\u4e86 k \u500b\u985e\u5225\uff0c\u4e26\u4e14\u4f7f\u7528\u9019\u4e9b\u985e\u5225\u6240\u5c6c\u7684\u6587\u4ef6\u8a13 \u7df4\u597d\u4e0d\u540c\u985e\u5225\u7684 n-gram \u6a21\u578b L 1 \u3001L 2 \u3001\u2026\u2026\u3001L k \uff0c \u4f7f\u7528\u8c9d\u5f0f\u5206\u985e\u5668\u6c42\u5f97\u6b64\u7bc7\u6587\u4ef6 \u6240\u5c6c\u7684\u985e\u5225 c * \u53ef\u5beb\u6210 (3) \u4f46\u662f\u6b64\u7a2e\u65b9\u6cd5\u5728\u8a08\u6bcf\u4e00\u500b\u8a5e\u7684\u689d\u4ef6\u6a5f\u7387\u6642\u90fd\u8981\u727d\u6d89\u5230\u524d\u9762\u6240\u6709\u7684\u8a5e\u5e8f\u5217\uff0c\u4f7f\u5f97\u8a08 \u7b97\u91cf\u592a\u5927\u800c\u7121\u6cd5\u5be6\u73fe\uff0c\u70ba\u89e3\u6c7a\u9019\u500b\u554f\u984c\u6240\u4ee5\u6709 n-gram \u6a21\u578b\u7684\u7522\u751f\uff0c\u5728 n-gram \u6a21 \u578b\u4e2d\uff0c\u5b83\u662f\u5047\u8a2d\u4e00\u500b\u8a5e\u51fa\u73fe\u7684\u6a5f\u7387\u53ea\u8ddf\u524d\u9762 n-1 \u500b\u8a5e\u6709\u95dc\uff0c\u56e0\u6b64(3)\u5f0f\u53ef\u4ee5\u8fd1\u4f3c\u70ba (4) \u5176\u4e2d 1 1 i i n W \u2212 \u2212 + \u4ee3\u8868 1 2 1 ...... i n i n i W W W \u2212 + \u2212 + \u2212 \u8a5e\u5e8f\u5217\uff0c\u5982\u6b64\u4e00\u4f86\u4f7f\u7528 n-gram \u53ef\u4ee5\u5927\u91cf\u7bc0\u7701\u8a08 \u7b97\u6642\u9593\u8207\u8a18\u61b6\u9ad4\uff0c\u8b93\u5be6\u7528\u6027\u5927\u70ba\u63d0\u9ad8\u3002\u800c\u4e00\u822c\u5728\u5efa\u7acb n-gram \u6a5f\u7387\u6a21\u578b 1 1 ( | ) i i i n P W W \u2212 \u2212 + \u6700\u76f4\u89ba\u7684\u65b9\u6cd5\u5c31\u662f\u7d71\u8a08\u5728\u8a5e\u5e8f\u5217 1 2 1 ...... i n i n i W W W \u2212 + \u2212 + \u2212 \u5f8c\u51fa\u73fe i W \u7684\u6b21\u6578\u518d\u9664\u4ee5\u8a5e\u5e8f \u5217 1 2 1 ...... i n i n i W W W \u2212 + \u2212 + \u2212 \u5728\u8a13\u7df4\u6587\u96c6\u4e2d\u51fa\u73fe\u7684\u6b21\u6578\uff0c\u4e5f\u5c31\u662f (5) \u5176\u4e2d ( ) j i C W \u4ee3\u8868 j i W \u5728\u8a13\u7df4\u6587\u96c6\u4e2d\u51fa\u73fe\u7684\u6b21\u6578\u3002",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            },
            {
                "text": "= = = \u220f 1 1 1 1 1 1 1 ( ) ( ) ( | ) ( ) ( ) i i i i i n i n i i n i i i n i n W C W C W P W W C W C W \u2212 \u2212 + \u2212 + \u2212 + \u2212 \u2212 + \u2212 + = = \u2211 1 2",
                "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": "\u80a1\"\u9019\u6bb5\u8a5e\u5e8f\u5217\u6703\u4e0d\u65b7\u51fa\u73fe\uff0c\u900f\u904e\u6211\u5011\u7684\u8a5e\u5178\uff0c\u6703\u5c07\u6b64\u8a5e\u5e8f\u5217\u65b7\u8a5e\u70ba\"\u91d1\u878d\"\u8207\"\u80a1\" \u5169\u500b\u8a5e\uff0c\u6b64\u6642\u82e5\u6211\u5011\u5728\u7b2c\u4e00\u6b21\u6e2c\u8a66\u5230\u6b64\u8a5e\u5e8f\u5217\u6642\uff0c\u5c07 \"\u91d1\u878d\" \u5f8c\u9762\u63a5 \"\u80a1\" \u7684\u6a5f \u7387\u63d0\u9ad8\uff0c\u81ea\u7136\u53ef\u4ee5\u589e\u5f37\u6211\u5011\u6a21\u578b\u7684\u6e96\u78ba\u6027\uff0c\u5728\u5feb\u53d6\u6a21\u578b\u4e2d\u6703\u4fdd\u7559\u4e00\u584a\u5feb\u53d6\u8a18\u61b6 \u9ad4\uff0c\u800c\u505a\u6587\u4ef6\u6e2c\u8a66\u6642\uff0c\u6703\u5c07\u6700\u8fd1\u6e2c\u8a66\u904e\u7684\u6587\u53e5\u62ff\u4f86\u8a13\u7df4\u51fa\u5feb\u53d6 n-gram \u6a21\u578b P c \u5c07 \u5176\u8207\u539f\u59cb\u7684\u7d71\u8a08\u6a21\u578b P S \u505a\u7d50\u5408\uff0c\u6211\u5011\u5c07\u6a21\u578b\u6a5f\u7387\u7528(9)\u5f0f\u8868\u793a (9) \u5176\u4e2d \u00b5 \u4ee3\u8868\u7d50\u5408\u6bd4\u91cd\u3002 \u800c\u5728\u672c\u8ad6\u6587\u4e2d\uff0c\u6211\u5011\u4f7f\u7528\u7684\u662f\u6587\u53e5\u968e\u5c64\u6df7\u5408\u5f0f n-gram \u6a21\u578b(sentence-level mixture n-gram model)\uff0c\u5728\u6bcf\u7d93\u904e\u4e00\u6587\u53e5\u5f8c\uff0c\u5c31\u5229\u7528\u6b64\u6587\u53e5\u6240\u63d0\u4f9b\u7684\u8cc7\u8a0a\u8abf\u6574\u6df7 \u5408\u6a21\u578b\u7684\u6bd4\u91cd\u53c3\u6578\u3002\u6211\u5011\u662f\u5229\u7528\u5947\u6469\u7db2\u7ad9\u5df2\u5206\u985e\u597d\u7684\u65b0\u805e\uff0c\u505a\u70ba\u6211\u5011\u7684\u5206\u985e\u7fa4 \u7d44\u3002\u800c\u6211\u5011\u6703\u4f9d\u64da\u5206\u7fa4\u904e\u5f8c\u4e4b\u6587\u96c6\u8a13\u7df4\u51fa\u5c0d\u61c9\u65bc\u5404\u7fa4\u7d44\u4e4b n-gram \u6a21\u578b\uff0c\u5728\u9019\u908a \u4ee5 P k \u8868\u793a\u7b2c k \u500b\u7fa4\u7d44\u7684 n-gram \u6a21\u578b\uff0c\u800c\u5728\u505a\u6e2c\u8a66\u6642\uff0c\u4f7f\u7528\u6b0a\u91cd k \u03bb \u5c07\u5404\u7fa4\u4e4b\u6a21 \u578b \u505a \u7d44 \u5408 \u6210 \u70ba \u6e2c \u8a66 \u7528 \u7684 n-gram \u6a21 \u578b \uff0c \u4e5f \u5c31 \u662f \u8aaa \u5047 \u8a2d \u6709 \u4e00 \u6587 \u53e5 S \u70ba W 1 W 2 W 3 \u2026\u2026W T \uff0c\u5247\u6b64\u6587\u53e5\u51fa\u73fe\u7684\u6a5f\u7387\u70ba (10) \u5176\u4e2d m \u4ee3\u8868\u6df7\u5408\u6578\u500b\u6578\uff0c\u4f46\u70ba\u6b64\u6a21\u578b\u9084\u9808\u505a\u5169\u9ede\u6539\u9032\uff0c\u7b2c\u4e00\u3001\u70ba\u4e86\u907f\u514d\u6bcf\u500b\u7fa4 \u7d44\u4e2d\u7684\u8a13\u7df4\u6587\u96c6\u592a\u5c11\uff0c\u9020\u6210\u8cc7\u6599\u7a00\u758f(data sparseness)\uff0c\u6bcf\u500b\u55ae\u4e00\u7fa4\u7d44\u6a21\u578b\u9700\u8981\u518d \u7d50\u5408\u4e00\u500b\u4e00\u822c\u5316\u7684\u6a21\u578b(general model)\uff0c\u7528\u4ee5\u589e\u52a0\u6a21\u578b\u7684\u53ef\u9760\u5ea6\uff0c\u7b2c\u4e8c\u3001\u5728\u6e2c\u8a66 \u6642\u53ef\u80fd\u6703\u6709\u7121\u9818\u57df(nontopic)\u7684\u6587\u96c6\u5b58\u5728\uff0c\u6240\u4ee5\u6211\u5011\u53c8\u5fc5\u9808\u5c07\u4e00\u822c\u5316\u6a21\u578b\u52a0\u5165\uff0c \u8996\u70ba\u4e00\u500b\u7121\u9818\u57df\u7684\u7fa4\u7d44\uff0c\u5728\u6b64\u6211\u5011\u5c07\u4e00\u822c\u5316\u6a21\u578b\u4ee5 P G \u8868\u793a\uff0c\u6545\u4e0a\u5f0f\u53ef\u6539\u5beb\u70ba , 1 1 1 1 1 1 1 ( ) [ ( | ) (1 ) ( | )] m G T i i k k k i in k G i in k i P S P W W P W W \u03bb \u03b1 \u03b1 + \u2212 \u2212 \u2212 + \u2212 + = = = + \u2212 \u2211 \u220f (11) \u5176\u4e2d k \u03b1 \u70ba\u7b2c k \u500b\u7fa4\u7d44\u6a21\u578b\u8207\u4e00\u822c\u5316\u6a21\u578b\u7684\u7d44\u5408\u6b0a\u91cd\u3002\u5728\u6df7\u5408\u5f0f n-gram \u6a21\u578b\u4e2d\uff0c \u6709\u5169\u500b\u6b0a\u91cd k \u03b1 \u53ca k \u03bb \u5b58\u5728\uff0c\u57fa\u672c\u4e0a\u6df7\u5408\u5f0f n-gram \u6a21\u578b\u662f\u4f9d\u64da\u524d\u6587\u4f86\u52d5\u614b\u7684\u8abf\u6574\u6b64 \u4e8c\u6b0a\u91cd\uff0c\u5728\u521d\u59cb\u6642\u6703\u4f7f\u7528\u5c11\u6578\u4fdd\u7559\u6587\u96c6\u4f30\u6e2c\u51fa\u5176\u521d\u59cb\u503c\uff0c\u6e2c\u8a66\u6642\u6703\u5728\u6bcf\u4e00\u6587\u53e5\u7d50 \u675f\u6642\u518d\u53bb\u505a\u4e00\u6b21\u6b0a\u91cd\u7684\u8abf\u6574\uff0c\u800c\u8abf\u6574\u7684\u52d5\u4f5c\u53ef\u4ee5\u5206\u5225\u5beb\u6210(12)(13)\u5f0f",
                        "eq_num": "(12)"
                    }
                ],
                "section": "",
                "sec_num": null
            },
            {
                "text": "1 1 1 1 1 1 1 1 1 1 ( | ) ( | ) (1 ) ( | ) k l k N T old i new k k i i n k N old i old i l i k k i i n k G i i n l l P W W P W W P W W T \u03b1 \u03b1 \u03b1 \u03b1 \u2212 \u2212 + \u2212 \u2212 = = \u2212 + \u2212 + = = + \u2212 \u2211\u2211 \u2211 1 1 1 1 2 1 1 1 ( ...... ) [(1 ) ( | ) ( | )] T s i c i T i i n i i n i P WW W P W W P W W \u00b5 \u00b5 + \u2212 \u2212 \u2212 + \u2212 + = = \u2212 + \u220f 1 1 1 1 1 1 1 11 ( ) ( | ) ( | ) T T m i i i i n k k i i n i",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            },
            {
                "text": "\u6a21\u578b\u4e2d\u6c92\u6709\u8a13\u7df4\u5230\u7684\u8a5e\u5e8f\u5217\u6a5f\u7387\u6a21\u578b\u4f7f\u7528(n-1)-gram \u6a21\u578b\u505a\u88dc\u511f\uff0c\u4e5f\u5c31\u662f 1 1 1 1 1 1 1 interp 1 1 interp 2 ( | ) ( | ) (1 ) ( | ) i i i N i N i i i i i N i i N i i N W W P W W P W W p W W \u03bb \u03bb \u2212 \u2212 \u2212 + \u2212 + \u2212 \u2212 \u2212 \u2212 + \u2212 + \u2212 + = + \u2212 (14) \u9019\u662f\u4e00\u500b\u905e\u8ff4\u5f0f\u7684\u5b9a\u7fa9\uff0c\u6240\u6709\u7684 n-gram \u6a21\u578b\u90fd\u5fc5\u9808\u5229\u7528(n-1)-gram \u6a21\u578b\u505a\u88dc\u511f\uff0c \u5176\u4e2d 1 1 i i N W \u03bb \u2212 \u2212 + \u4ee3\u8868\u7684\u662f\u5408\u4f75 n-gram \u8207 (n-1)-gram \u4e4b\u6b0a\u91cd\uff0c\u800c Witten-Bell \u5e73\u6ed1\u5316\u6280 \u8853\u5c0d\u6b64\u4e00\u6b0a\u91cd\u6709\u4e00\u500b\u7279\u6b8a\u7684\u4f30\u6e2c\u65b9\u5f0f\uff0c\u5728\u9019\u908a\u5148\u5c0d\u7b26\u865f\u505a\u4ee5\u4e0b\u7684\u5b9a\u7fa9 1 1 1 1 1 ( ) |{ : ( ) 0}| i i i n i i n i N W W C W W \u2212 \u2212 + \u2212 + \u2212 + \u22c5 = > (15) 1 1 1 ( ) i i n N W \u2212 + \u2212 + \u22c5 \u4ee3\u8868\u5728 1 1 i i n W \u2212 \u2212 + \u5f8c\u53ef\u63a5\u7684\u8a5e\u6578\uff0c\u5176\u4e2d\u4e0b\u6a19\u300c1+\u300d\u4ee3\u8868\u662f\u9023\u63a5\u4e00\u500b\u8a5e \u4ee5\u4e0a\u4e4b\u610f\u3002\u6b0a\u91cd\u56e0\u6578\u5b9a\u7fa9\u70ba (16) \u5373\u70ba Witten-Bell \u7684 n-gram \u6a21\u578b\u5efa\u7acb\u65b9\u5f0f\uff0c\u5176\u7269\u7406\u610f\u7fa9\u8868\u793a\u5728\u7d71\u8a08 1 1 i i n W \u2212 \u2212 + \u51fa\u73fe\u6b21\u6578 \u6642\uff0c\u5982\u679c 1 1 i i n W \u2212 \u2212 + \u5f8c\u9762\u53ef\u63a5\u7684\u8a5e\u6578\u8d8a\u5c11\uff0c\u6211\u5011\u7d66 1 1 ( | ) i i i n P W W \u2212 \u2212 + \u8f03\u5927\u7684\u6b0a\u91cd\uff0c\u53cd\u4e4b\u5247\u4f7f \u7528\u8f03\u591a\u7684(n-1)-gram \u505a\u88dc\u511f\uff0c\u5047\u8a2d\u5728\u505a bigram \u7d71\u8a08\u6642\uff0c\u8a5e\u5178\u4e2d\u6709\u4e00\u8a5e\u70ba \u300c\u985e\u795e\u7d93\u300d \uff0c \u6211\u5011\u767c\u73fe\u5728\u8a13\u7df4\u6587\u96c6\u4e2d\u300c\u985e\u795e\u7d93\u300d\u5f8c\u90fd\u63a5\u300c\u7db2\u8def\u300d\u4e00\u8a5e\uff0c\u6b64\u6642\u5c31\u4e0d\u9700\u8981\u592a\u591a\u7684 unigram \u505a\u88dc\u511f\uff0c\u9019\u662f\u56e0\u70ba\u6b64\u540d\u8a5e\u6709\u7368\u7279\u6027\uff0c\u5f8c\u9762\u5e7e\u4e4e\u90fd\u63a5\u5c11\u91cf\u7279\u5b9a\u7684\u8a5e\uff0c\u800c\u82e5 \u6b32\u7d71\u8a08\u4e00\u8a5e\u300c\u5e7e\u4e4e\u300d\u5f8c\u53ef\u63a5\u8a5e\u7684 bigram \u6a5f\u7387\uff0c\u53ef\u80fd\u6703\u767c\u73fe\u8a13\u7df4\u6587\u96c6\u4e2d\u5176\u5f8c\u53ef\u63a5 1 , 1 1 1 ( ,..., ) 1 ( ,..., ) i i old N k k T new k m G old i j j T j P W W N P W W \u03bb \u03bb \u03bb = = = \u2211 \u2211 1 1 1 1 1 1 1 1 1 ( ) 1 ( ) ( ) i i n i i i n i i W i n i n W N W N",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            },
            {
                "text": "(2)\u5927\u9678 \u5c0f\u4e09\u901a \u21d2 \u5169\u5cb8 confidence = 100% support = 2.25%",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            },
            {
                "text": "\u4ee5\u4e0a\u4f8b\u800c\u8a00\uff0c\u51fa\u73fe\"\u5c0f\u4e09\u901a\"\u5f8c\uff0c\u65b0\u805e\u8a9e\u6599\u6709 90 % \u7684\u6a5f\u7387\u4e5f\u6703\u51fa\u73fe \"\u5927\u9678\" \u9019\u500b \u8a5e\uff0c\u9019\u500b\u95dc\u806f\u6cd5\u5247\u4f54\u4e86\u7e3d\u6587\u53e5 6.25%\uff0c\u5982\u679c\"\u5927\u9678\"\u8207\"\u5c0f\u4e09\u901a\"\u540c\u6642\u51fa\u73fe\u5f8c\uff0c\u65b0\u805e \u8a9e\u6599\u6703\u6709 100 %\u7684\u6a5f\u7387\u4e5f\u6703\u51fa\u73fe\"\u5169\u5cb8\"\uff0c\u800c\u6b64\u95dc\u806f\u6cd5\u5247\u4f54\u4e86\u7e3d\u6587\u53e5 2.25%\u3002 ",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            }
        ],
        "back_matter": [],
        "bib_entries": {
            "BIBREF0": {
                "ref_id": "b0",
                "title": "Fast Algorithms for Mining Association Rules",
                "authors": [
                    {
                        "first": "R",
                        "middle": [],
                        "last": "Agrawal",
                        "suffix": ""
                    },
                    {
                        "first": "R",
                        "middle": [],
                        "last": "Srikant",
                        "suffix": ""
                    }
                ],
                "year": 1994,
                "venue": "Proceedings of the 20 th VLDB Conference",
                "volume": "",
                "issue": "",
                "pages": "487--499",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "R. Agrawal and R. Srikant, \"Fast Algorithms for Mining Association Rules\", Proceedings of the 20 th VLDB Conference, Santiago-Chile, pp.487-499, 1994\u3002",
                "links": null
            },
            "BIBREF1": {
                "ref_id": "b1",
                "title": "An Empirical Study of Smoothing Techniques for Language Modeling",
                "authors": [
                    {
                        "first": "S",
                        "middle": [
                            "F"
                        ],
                        "last": "Chen",
                        "suffix": ""
                    },
                    {
                        "first": "J",
                        "middle": [],
                        "last": "Goodman",
                        "suffix": ""
                    }
                ],
                "year": 1999,
                "venue": "Computer Speech and Language",
                "volume": "13",
                "issue": "",
                "pages": "359--394",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "S. F. Chen and J. Goodman, \"An Empirical Study of Smoothing Techniques for Language Modeling\", Computer Speech and Language , vol.13, 359-394 , 1999\u3002",
                "links": null
            },
            "BIBREF2": {
                "ref_id": "b2",
                "title": "Language Model Adaptation Using Mixtures and an Exponentially Decaying Cache",
                "authors": [
                    {
                        "first": "P",
                        "middle": [
                            "R"
                        ],
                        "last": "Clarkson",
                        "suffix": ""
                    },
                    {
                        "first": "A",
                        "middle": [
                            "J"
                        ],
                        "last": "Robinson",
                        "suffix": ""
                    }
                ],
                "year": 1997,
                "venue": "Proc. of ICASSP",
                "volume": "",
                "issue": "",
                "pages": "799--802",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "P. R. Clarkson and A. J. Robinson, \"Language Model Adaptation Using Mixtures and an Exponentially Decaying Cache\", Proc. of ICASSP, pp.799-802 , 1997\u3002",
                "links": null
            },
            "BIBREF3": {
                "ref_id": "b3",
                "title": "Relevance weighting for combining multi-domain data for n-gram language modeling",
                "authors": [
                    {
                        "first": "R",
                        "middle": [],
                        "last": "Iyer",
                        "suffix": ""
                    },
                    {
                        "first": "M",
                        "middle": [],
                        "last": "Ostendorf",
                        "suffix": ""
                    }
                ],
                "year": 1999,
                "venue": "Computer Speech and Language",
                "volume": "13",
                "issue": "",
                "pages": "267--282",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "R. Iyer and M. Ostendorf, \"Relevance weighting for combining multi-domain data for n-gram language modeling\", Computer Speech and Language, vol.13, pp.267-282, 1999\u3002",
                "links": null
            },
            "BIBREF4": {
                "ref_id": "b4",
                "title": "Modeling long distance dependence in language : Topic Mixtures Versus dynamic cache models",
                "authors": [
                    {
                        "first": "R",
                        "middle": [
                            "M"
                        ],
                        "last": "Iyer",
                        "suffix": ""
                    },
                    {
                        "first": "M",
                        "middle": [],
                        "last": "Ostendorf",
                        "suffix": ""
                    }
                ],
                "year": 1999,
                "venue": "IEEE Transaction on speech and audio processing",
                "volume": "7",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "R. M. Iyer and M. Ostendorf, \"Modeling long distance dependence in language : Topic Mixtures Versus dynamic cache models\", IEEE Transaction on speech and audio processing , vol.7 , January 1999\u3002",
                "links": null
            },
            "BIBREF5": {
                "ref_id": "b5",
                "title": "Interpolation estimation of Markov source parameters from sparse data",
                "authors": [
                    {
                        "first": "F",
                        "middle": [],
                        "last": "Jelinek",
                        "suffix": ""
                    },
                    {
                        "first": "R",
                        "middle": [
                            "L"
                        ],
                        "last": "Mercer",
                        "suffix": ""
                    }
                ],
                "year": 1980,
                "venue": "Proceedings of the workshop on pattern recognition in Practice",
                "volume": "",
                "issue": "",
                "pages": "381--397",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "F. Jelinek and R. L. Mercer, \"Interpolation estimation of Markov source parameters from sparse data\", Proceedings of the workshop on pattern recognition in Practice, North-Holland, Amsterdam, The Netherlands, pp.381-397, May 1980\u3002",
                "links": null
            },
            "BIBREF6": {
                "ref_id": "b6",
                "title": "Selecting Articles from the Language Model Training Corpus",
                "authors": [
                    {
                        "first": "D",
                        "middle": [],
                        "last": "Klakow",
                        "suffix": ""
                    }
                ],
                "year": 2000,
                "venue": "Proc of ICASSP",
                "volume": "",
                "issue": "",
                "pages": "1695--1698",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "D. Klakow, \"Selecting Articles from the Language Model Training Corpus\", Proc of ICASSP, pp.1695 -1698, 2000\u3002",
                "links": null
            },
            "BIBREF7": {
                "ref_id": "b7",
                "title": "A maximum entropy approach to adaptive statistical language model",
                "authors": [
                    {
                        "first": "R",
                        "middle": [],
                        "last": "Rosenfeld",
                        "suffix": ""
                    }
                ],
                "year": 1996,
                "venue": "Computer Speech and Language",
                "volume": "10",
                "issue": "",
                "pages": "187--228",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "R. Rosenfeld, \"A maximum entropy approach to adaptive statistical language model\", Computer Speech and Language, vol 10 , pp.187-228 , 1996.",
                "links": null
            },
            "BIBREF8": {
                "ref_id": "b8",
                "title": "Trigger-based language models: A maximum entropy approach",
                "authors": [
                    {
                        "first": "R",
                        "middle": [],
                        "last": "Lau",
                        "suffix": ""
                    },
                    {
                        "first": "R",
                        "middle": [],
                        "last": "Rosenfeld",
                        "suffix": ""
                    },
                    {
                        "first": "S",
                        "middle": [],
                        "last": "Roukos",
                        "suffix": ""
                    }
                ],
                "year": 1993,
                "venue": "Proc. Int. Conf. Acoustics, Speech, Signal Processing",
                "volume": "II",
                "issue": "",
                "pages": "45--48",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "R. Lau, R. Rosenfeld, and S. Roukos, \"Trigger-based language models: A maximum entropy approach\" , in Proc. Int. Conf. Acoustics, Speech, Signal Processing, vol. II, pp. 45-48. , 1993",
                "links": null
            },
            "BIBREF9": {
                "ref_id": "b9",
                "title": "Statistical language modeling combining N -gram and context free grammars",
                "authors": [
                    {
                        "first": "M",
                        "middle": [],
                        "last": "Meteer",
                        "suffix": ""
                    },
                    {
                        "first": "J",
                        "middle": [
                            "R"
                        ],
                        "last": "Rohlicek",
                        "suffix": ""
                    }
                ],
                "year": 1993,
                "venue": "Proc. Int. Conf. Acoustics, Speech, Signal Processing",
                "volume": "II",
                "issue": "",
                "pages": "37--40",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "M. Meteer and J. R. Rohlicek, \"Statistical language modeling combining N -gram and context free grammars\" , in Proc. Int. Conf. Acoustics, Speech, Signal Processing, vol. II, pp. 37-40 , 1993.",
                "links": null
            },
            "BIBREF10": {
                "ref_id": "b10",
                "title": "Foundations of statistical natural language processing",
                "authors": [
                    {
                        "first": "C",
                        "middle": [
                            "D"
                        ],
                        "last": "Manning",
                        "suffix": ""
                    },
                    {
                        "first": "H",
                        "middle": [],
                        "last": "Schutze",
                        "suffix": ""
                    }
                ],
                "year": 1999,
                "venue": "Massachusetts Institute of Technology",
                "volume": "",
                "issue": "",
                "pages": "315--407",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "C. D. Manning, H. Schutze, \"Foundations of statistical natural language processing\", Massachusetts Institute of Technology pp.315-407, 1999\u3002",
                "links": null
            },
            "BIBREF11": {
                "ref_id": "b11",
                "title": "Funadamental of Speech Recognition",
                "authors": [
                    {
                        "first": "L",
                        "middle": [],
                        "last": "Rabiner",
                        "suffix": ""
                    },
                    {
                        "first": "B",
                        "middle": [
                            "H"
                        ],
                        "last": "Juang",
                        "suffix": ""
                    }
                ],
                "year": 1993,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "321--387",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "L. Rabiner and B.H. Juang, \"Funadamental of Speech Recognition\", Prentice Hall, pp.321-387, 1993\u3002",
                "links": null
            },
            "BIBREF12": {
                "ref_id": "b12",
                "title": "The zero-frequency problem : Estimating the probabilities of novel events in adaptive text compression",
                "authors": [
                    {
                        "first": "I",
                        "middle": [
                            "H"
                        ],
                        "last": "Witten",
                        "suffix": ""
                    },
                    {
                        "first": "T",
                        "middle": [
                            "C"
                        ],
                        "last": "Bell",
                        "suffix": ""
                    }
                ],
                "year": 1991,
                "venue": "IEEE Transactions on Information Theory",
                "volume": "37",
                "issue": "",
                "pages": "1085--1094",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "I. H. Witten and T. C. Bell, \"The zero-frequency problem : Estimating the probabilities of novel events in adaptive text compression.\", IEEE Transactions on Information Theory , vol.37, pp.1085-1094, 1991\u3002",
                "links": null
            },
            "BIBREF13": {
                "ref_id": "b13",
                "title": "Interpolation of n-gram and mutual-information based trigger pair language models for Mandarin speech recognition",
                "authors": [
                    {
                        "first": "G",
                        "middle": [
                            "D"
                        ],
                        "last": "Zhou",
                        "suffix": ""
                    },
                    {
                        "first": "K",
                        "middle": [
                            "T"
                        ],
                        "last": "Lua",
                        "suffix": ""
                    }
                ],
                "year": 1999,
                "venue": "Computer Speech and Language",
                "volume": "13",
                "issue": "",
                "pages": "125--141",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "G. D. Zhou and K. T. Lua, \"Interpolation of n-gram and mutual-information based trigger pair language models for Mandarin speech recognition\", Computer Speech and Language, vol. 13, pp.125-141, 1999\u3002",
                "links": null
            }
        },
        "ref_entries": {
            "FIGREF0": {
                "type_str": "figure",
                "num": null,
                "uris": null,
                "text": ",b)\uff0c(a,c)\uff0c(b,c)\uff0c\u4e26\u4e14\u5c0d\u8cc7\u6599\u5eab\u641c\u5c0b\u6bcf\u4e00\u5e8f\u5c0d\uff0c\u662f \u5426\u540c\u6642\u51fa\u73fe\u5728\u65bc\u540c\u4e00\u6587\u53e5\u4e2d\uff0c\u5047\u8a2d\u53ea\u6709(a,b)\uff0c(b,c)\u5e8f\u5c0d\u7b26\u5408\u9019\u9805\u689d\u4ef6\uff0c\u5247\u5c07(a,c) \u522a\u9664\uff0c\u6b64\u6642\u6211\u5011\u5efa\u7acb(a,b)\uff0c(b,c)\u7684\u95dc\u806f\u6cd5\u5247\uff0c\u6b64\u95dc\u806f\u6cd5\u5247\u7684\u5c64\u7d1a(step) \u70ba\u4e8c \uff0c\u4e0d \u904e\u6211\u5011\u5fc5\u9808\u8a08\u7b97\u5176\u4fe1\u8cf4\u5ea6\u8207\u652f\u63f4\u5ea6\uff0c\u4f8b\u5982\u6211\u5011\u53ef\u4ee5\u8a08\u7b97\u540c\u4e00\u7bc7\u6587\u7ae0\u51fa\u73fe\u8a5e a \u4e14\u51fa \u73fe\u8a5e b \u7684\u6a5f\u7387\uff0c\u5373\u70ba\u5176\u4fe1\u8cf4\u5ea6\u3002\u800c\u6211\u5011\u6703\u518d\u5c07\u5269\u9918\u4e0b\u4e4b\u5e8f\u5c0d(a,b)\uff0c(b,c)\u505a\u7d50\u5408\u6210 \u70ba(a,b,c) \u4e26\u641c\u5c0b\u8a13\u7df4\u6587\u96c6\u4e2d(a,b,c\u7684\u6587\u7ae0\u4e2d\u6709 c%\u7684\u6a5f\u7387\u6703\u51fa\u73fe Y \uff0c\u800c\u6709 s% \u7684\u6587\u7ae0\u540c\u6642\u5305\u542b\u4e86 WordSeq \u8207 Y\u3002\u4ee5\u4e0b\u70ba\u5229\u7528\u897f\u5143\u4e8c\u5343\u96f6\u4e00\u5e74\u5341\u4e8c\u6708\u4e8c\u5341\u516b\u865f\u5230\u897f\u5143\u4e8c\u5343\u96f6\u4e00\u5e74\u5341\u4e8c\u6708\u4e09\u5341 \u4e00\u865f\u671f\u9593\u7684\u653f\u6cbb\u65b0\u805e\u6240\u64f7\u53d6\u51fa\u7684\u5169\u689d\u95dc\u806f\u6cd5\u5247\u7bc4\u4f8b\uff0c\u7b2c\u4e00\u689d\u95dc\u806f\u6cd5\u5247\u7684\u5c64\u7d1a\u70ba \u4e8c\uff0c\u7b2c\u4e8c\u689d\u7684\u5c64\u7d1a\u5247\u70ba\u4e09 \u5c0f\u4e09\u901a \u21d2 \u5927\u9678 confidence = 90% support = 6.25%"
            },
            "FIGREF1": {
                "type_str": "figure",
                "num": null,
                "uris": null,
                "text": "ig ra m + A s s o c ia tio n ru le B ig ra m + T rig g e r p a ir perplexity a s s o c ia tio n s te p"
            },
            "TABREF0": {
                "text": "",
                "content": "<table><tr><td>\u7684\u9ed1\u624b\u5c31\u662f\u5efa\u7acb\u5728 \"\u4f7f\u4eba\u4fbf\u5229\" \u7684\u57fa\u790e\u4e4b\u4e0a\uff0c\u4f46\u662f\u96fb\u8166\u81ea\u5f9e\u5728\u767c\u660e\u4e4b\u521d\u5c31\u5b58\u5728\u4e00 \u65b0\u805e\u6587\u4ef6\u5206\u985e\u7684\u7cfb\u7d71\u4e0a\u6709\u4e00\u5b9a\u5e45\u5ea6\u7684\u5e6b\u52a9\u3002</td></tr><tr><td>\u500b\u8207\u4eba\u6027\u80cc\u9053\u800c\u99b3\u7684\u7f3a\u9ede\uff0c\u8207\u5b83\u5011\u7684\u6e9d\u901a\u9700\u8981\u900f\u904e\u4e00\u500b\u7279\u5b9a\u7684\u6309\u9375\u88dd\u7f6e\uff0c\u6bd4\u65b9\u8aaa</td></tr><tr><td>\u8981\u8207\u500b\u4eba\u96fb\u8166\u6e9d\u901a\u5c31\u5fc5\u9808\u900f\u904e\u9375\u76e4\u6216\u6ed1\u9f20\u7b49\u88dd\u7f6e\uff0c\u4e8b\u5be6\u4e0a\u9019\u662f\u4f7f\u8a31\u591a\u4eba\u5c0d\u96fb\u8166\u671b 2. n-gram \u8a9e\u8a00\u6a21\u578b\u7c21\u4ecb</td></tr><tr><td>\u4e4b\u537b\u6b65\u7684\u539f\u56e0\uff0c\u8981\u5b78\u7fd2\u5982\u4f55\u4f7f\u7528\u9375\u76e4\u8207\u96fb\u8166\u505a\u6e9d\u901a\u5c31\u7b49\u65bc\u662f\u5f37\u8feb\u4eba\u53bb\u5b78\u7fd2\u3127\u7a2e \u76ee\u524d n-gram[11]\u6a21\u578b\u7684\u63a2\u8a0e\u65bc\u5404\u76f8\u95dc\u5b78\u8853\u6703\u8b70\u53ca\u671f\u520a\u8ad6\u6587\u4e0a\u5df2\u6709\u76f8\u7576\u591a\u7684</td></tr><tr><td>\"\u96fb\u8166\u8a9e\u8a00\"\uff0c\u9019\u8207 \"\u4f7f\u4eba\u4fbf\u5229\" \u7684\u539f\u5247\u7576\u7136\u662f\u4e92\u76f8\u9055\u80cc\u7684\uff0c\u4f46\u662f\u53cd\u904e\u4f86\u8aaa\u5982\u80fd\u8b93 \u6587\u737b\u767c\u8868\uff0c\u986f\u793a\u5404\u7a2e\u7814\u7a76\u6a5f\u69cb\u5c0d\u6b64\u4e00\u9818\u57df\u7684\u767c\u5c55\u6709\u76f8\u7576\u5927\u7684\u671f\u8a31\uff0c\u6545\u6295\u8eab\u65bc\u5176</td></tr><tr><td>\u96fb\u8166\u5b78\u7fd2\u4eba\u985e\u7684\u8a9e\u8a00\uff0c\u4f7f\u96fb\u8166\u80fd\u66f4\u63a5\u8fd1\u4eba\u985e\uff0c\u4e5f\u5c31\u80fd\u4f7f\u5176\u8207\u4eba\u985e\u751f\u6d3b\u7684\u7d50\u5408\u66f4\u52a0 \u4e2d\uff0c\u800c\u5728\u5404\u65b9\u90fd\u81f4\u529b\u65bc\u6539\u9032 n-gram \u6a21\u578b\u4e4b\u4e0b\uff0cn-gram \u6a21\u578b\u5728\u6548\u80fd\u4e0a\u5df2\u7372\u5f97\u76f8\u7576</td></tr><tr><td>\u7c21\u4ec1\u5b97 \u9673\u9d3b\u5100 \u7dca\u5bc6\uff0c\u9032\u4e00\u6b65\u5982\u679c\u96fb\u8166\u80fd\u900f\u904e\u8a9e\u8a00\u7684\u5b78\u7fd2\u800c\u5177\u5099\u4e86\u95b1\u8b80\u7684\u80fd\u529b\uff0c\u6211\u5011\u5c31\u53ef\u4ee5\u8b93\u96fb \u4e0d\u932f\u4e4b\u6210\u679c\uff0c\u5728\u672c\u7ae0\u4e2d\u6211\u5011\u5c07\u6703\u5c0d n-gram \u6a21\u578b\u7684\u57fa\u672c\u6982\u5ff5\u505a\u4e00\u7c21\u55ae\u4e4b\u4ecb\u7d39\u3002</td></tr><tr><td>\u570b\u7acb\u6210\u529f\u5927\u5b78\u8cc7\u8a0a\u5de5\u7a0b\u5b78\u7cfb \u8166\u70ba\u6211\u5011\u904e\u6ffe\u4ea6\u6216\u5206\u985e\u6bcf\u5929\u6240\u9700\u95b1\u8b80\u7684\u6587\u4ef6\uff0c\u6bd4\u65b9\u8aaa\u53ef\u4ee5\u61c9\u7528\u5728\u65bc e-mail \u5ee3\u544a\u904e</td></tr><tr><td>Email\uff1ajtchien@mail.ncku.edu.tw \u6ffe\u6216\u662f\u65b0\u805e\u6587\u4ef6\u5206\u985e\u7b49\u7b49\uff0c\u5c31\u53ef\u4ee5\u8b93\u96fb\u8166\u70ba\u6211\u5011\u7701\u4e0b\u66f4\u591a\u7684\u6642\u9593\u3002 2.1 n-gram \u6a21\u578b\u4e4b\u61c9\u7528</td></tr><tr><td>\u8981\u514b\u670d\u96fb\u8166\u8207\u4eba\u5728\u8a9e\u8a00\u4e0a\u7684\u9d3b\u6e9d\uff0c\u5728\u8a9e\u8a00\u6280\u8853\u7684\u9818\u57df\u6709\u4e86\u8072\u5b78\u6a21\u578b(acoustic \u4e00\u822c\u800c\u8a00 n-gram \u8a9e\u8a00\u6a21\u578b\u901a\u5e38\u61c9\u7528\u65bc\u8c9d\u5f0f\u5206\u985e\u5668(Bayes classifier)\uff0c\u626e\u6f14\u8457</td></tr><tr><td>\u6458\u8981 model)\u8207\u81ea\u7136\u8a9e\u8a00\u6a21\u578b(natural language model)\u7684\u7522\u751f\uff0c\u800c\u9019\u5169\u9805\u6280\u8853\u7684\u767c\u5c55\u5728\u570b \u4e8b\u524d\u6a5f\u7387(priori probability)\u6216\u662f\u53ef\u80fd\u6027( likelihood )\u7684\u89d2\u8272\uff0c\u4ee5\u8a9e\u97f3\u8fa8\u8b58\u70ba\u4f8b\u5b50\u800c</td></tr><tr><td>\u672c\u8ad6\u6587\u63d0\u51fa\u4e00\u7a2e\u80fd\u64f7\u53d6\u9577\u8ddd\u96e2\u8cc7\u8a0a\u7684\u8a9e\u8a00\u6a21\u578b\uff0c\u5b83\u53ef\u4ee5\u64f7\u53d6\u591a\u8a5e\u5f59\u4e4b\u9593\u7684\u95dc \u5916\u5df2\u7d93\u884c\u4e4b\u6709\u5e74\uff0c\u53f0\u7063\u81ea\u897f\u5143\u4e00\u4e5d\u516b\u4e8c\u5e74\u8d77\u4fbf\u958b\u59cb\u6709\u4e86\u4e2d\u6587\u8072\u5b78\u6a21\u578b\u65b9\u9762\u7684\u7814 \u8a00\uff0c\u5047\u8a2d\u6709\u4e00\u6bb5\u8072\u5b78\u8a0a\u865f(acoustic signal)X\uff0c\u6211\u5011\u7684\u76ee\u6a19\u662f\u53bb\u627e\u5c0b\u51fa\u6b64\u8a0a\u865f\u6700\u6709</td></tr><tr><td>\u806f\u6027\uff0c\u64f7\u53d6\u7684\u65b9\u5f0f\u662f\u4f7f\u7528\u8cc7\u6599\u63a2\u52d8\u4e2d\u5341\u5206\u6d41\u884c\u7684 Apriori \u6f14\u7b97\u6cd5\uff0c\u50b3\u7d71\u4e0a n-gram \u7a76\uff0c\u8a31\u591a\u7814\u7a76\u55ae\u4f4d\u5305\u62ec\u53f0\u6e05\u4ea4\u6210\u7b49\u5927\u5c08\u9662\u6821\uff0c\u4ee5\u53ca\u5de5\u7814\u9662\u3001\u4ea4\u901a\u90e8\u3001\u4e2d\u7814\u9662\u3001\u4e2d</td></tr><tr><td>\u8a9e\u8a00\u6a21\u578b\u53ea\u80fd\u5728 n-gram \u8996\u7a97\u5167\u64f7\u53d6\u5230\u6709\u9650\u8ddd\u96e2\u7684\u8cc7\u8a0a\uff0c\u8f03\u9577\u8ddd\u96e2\u7684\u8cc7\u8a0a\u4e5f\u5c31\u56e0 \u83ef\u96fb\u4fe1\u7b49\u90fd\u7a4d\u6975\u7684\u6295\u5165\u7814\u7a76\u7684\u5de5\u4f5c\u4e26\u4e14\u64c1\u6709\u4e86\u5341\u5206\u8c50\u78a9\u7684\u7814\u7a76\u6210\u679c\uff0c\u800c\u5728\u8072\u5b78\u6a21</td></tr><tr><td>\u6b64\u800c\u6d41\u5931\uff0c\u7136\u800c\u9019\u4e9b\u5931\u53bb\u7684\u9577\u8ddd\u96e2\u8cc7\u8a0a\u5c0d\u65bc\u8a9e\u8a00\u6a21\u578b\u662f\u5341\u5206\u91cd\u8981\u7684\uff0c\u6240\u4ee5\u5982\u4f55\u514b \u578b\u5df2\u65e5\u76ca\u6210\u719f\u7684\u57fa\u790e\u4e0b\uff0c\u81ea\u7136\u8a9e\u8a00\u6a21\u578b\u7684\u767c\u5c55\u4e5f\u5099\u53d7\u77da\u76ee\uff0c\u8aa0\u5982\u524d\u6587\u6240\u8ff0\uff0c\u8a9e\u97f3</td></tr><tr><td>\u670d n-gram \u6a21\u578b\u7f3a\u4e4f\u9577\u8ddd\u96e2\u8cc7\u8a0a\u4e00\u76f4\u662f\u975e\u5e38\u71b1\u9580\u7684\u7814\u7a76\u8ab2\u984c\uff0c\u89f8\u767c\u5e8f\u5c0d\u5c31\u662f\u5176\u4e2d \u6280\u8853\u767c\u5c55\u7684\u6700\u7d42\u76ee\u7684\u5c31\u662f\u8981\u5c07\u96fb\u8166\u8207\u4eba\u985e\u7684\u6e9d\u901a\u4fbf\u5229\u5316\uff0c\u800c\u8981\u9054\u5230\u9019\u500b\u76ee\u7684\uff0c\u5c07</td></tr><tr><td>\u4e00\u7a2e\u6709\u6548\u7684\u65b9\u6cd5\uff0c\u5176\u4e3b\u8981\u529f\u80fd\u662f\u5728\u64f7\u53d6\u9577\u8ddd\u96e2\u4e4b\u8a5e\u5e8f\u5c0d\u8cc7\u8a0a\uff0c\u4e5f\u5c31\u662f\u5efa\u7acb\u8d77\u8a5e\u8207 \u8a9e\u97f3\u6a21\u578b\u8207\u81ea\u7136\u8a9e\u8a00\u6a21\u578b\u505a\u7d50\u5408\u662f\u5fc5\u9808\u7684\uff0c\u6211\u5011\u7684\u8ad6\u6587\u4e3b\u8981\u5c31\u662f\u8457\u58a8\u65bc\u81ea\u7136\u8a9e\u8a00</td></tr><tr><td>\u8a5e\u4e4b\u9593\u7684\u95dc\u806f\u6027\uff0c\u7136\u800c\u6211\u5011\u6240\u63d0\u51fa\u7684\u95dc\u806f\u6cd5\u5247\u6280\u8853\u80fd\u64f7\u53d6\u591a\u5143\u8a5e\u7d44\u9593\u7684\u95dc\u806f\u6027\uff0c \u6a21\u578b\u7684\u63a2\u8a0e\uff0c\u6211\u5011\u5c07\u6703\u5c0d\u81ea\u7136\u8a9e\u8a00\u6a21\u578b\u4e2d\u7684\u4e00\u9805\u5341\u5206\u6210\u529f\u4e14\u5ee3\u6cdb\u904b\u7528\u7684\u6280\u8853</td></tr><tr><td>\u53ef\u4ee5\u8aaa\u662f\u9032\u4e00\u6b65\u6539\u826f\u8a5e\u7d44\u6578\u4e26\u5efa\u7acb\u66f4\u9577\u8ddd\u96e2\u8cc7\u8a0a\uff0c\u800c\u5be6\u9a57\u7d50\u679c\u4e5f\u986f\u793a\u672c\u8ad6\u6587\u65b9\u6cd5 n-gram \u8a9e\u8a00\u6a21\u578b\u505a\u4ecb\u7d39\uff0c\u4e26\u4e14\u5206\u6790\u5176\u5728\u50b3\u7d71\u4e0a\u7684\u7f3a\u9ede\u8207\u6539\u9032\u6280\u8853\uff0c\u800c\u672c\u8ad6\u6587\u4e5f</td></tr><tr><td>\u6bd4\u8d77\u50b3\u7d71\u89f8\u767c\u5e8f\u5c0d\u7372\u5f97\u8f03\u4f4e\u7684 perplexity\uff0c\u6b64\u95dc\u806f\u6cd5\u5247\u6280\u8853\u4e5f\u53ef\u4ee5\u6709\u6548\u7684\u8207\u5176\u4ed6 \u5c07\u6703\u91dd\u5c0d n-gram \u6a21\u578b\u5176\u4e2d\u4e00\u9805\u7f3a\u9ede-\u9577\u8ddd\u96e2\u8cc7\u8a0a\u7684\u7f3a\u4e4f\uff0c\u63d0\u51fa\u4e00\u5957\u65b0\u7684\u6539\u9032\u65b9</td></tr><tr><td>\u6a21\u578b\u8abf\u6574\u53ca\u6a21\u578b\u5e73\u6ed1\u5316\u7684\u6280\u8853\u7d50\u5408\uff0c\u5728\u8a9e\u8a00\u6a21\u578b\u7684\u6548\u7387\u6539\u5584\u65b9\u9762\u80fd\u6709\u66f4\u826f\u597d\u7684\u6548 \u6cd5\uff0c\u4e26\u4e14\u7d50\u5408\u5176\u4ed6\u6539\u9032\u65b9\u6cd5\uff0c\u9032\u800c\u767c\u5c55\u51fa\u4e00\u5957\u8f03\u6709\u6548\u7387\u7684 n-gram \u6a21\u578b\uff0c\u6211\u5011\u5c07</td></tr><tr><td>\u679c\uff0c\u6700\u5f8c\u672c\u8ad6\u6587\u4e5f\u5c07\u63d0\u51fa\u7684\u8a9e\u8a00\u6a21\u578b\u6210\u529f\u7684\u61c9\u7528\u5728\u8a9e\u97f3\u8fa8\u8b58\u8207\u6587\u4ef6\u5206\u985e\u4e0a\uff0c\u4e26\u5efa \u6703\u5c07\u5176\u61c9\u7528\u5728\u7d50\u5408\u8072\u5b78\u6a21\u578b\u505a\u8a9e\u97f3\u8fa8\u8b58\u548c\u6587\u4ef6\u5206\u985e\u7684\u9818\u57df\u4e4b\u4e0a\uff0c\u671f\u671b\u5c0d\u5176\u6b63\u78ba\u7387</td></tr><tr><td>\u7acb\u4e00\u5957\u500b\u4eba\u5316\u4e4b\u65b0\u805e\u700f\u89bd\u5668\u4e4b\u5c55\u793a\u7cfb\u7d71\u3002 \u6709\u4e00\u5b9a\u5e45\u5ea6\u7684\u6539\u5584\u3002</td></tr><tr><td>\u800c\u81ea\u7136\u8a9e\u8a00\u6a21\u578b\u65b9\u9762\u5728\u73fe\u4eca\u6709\u8a31\u591a\u4e0d\u540c\u7684\u767c\u5c55\uff0c\u4f9d\u5176\u5167\u5bb9\u4e3b\u8981\u5206\u70ba\u4e09\u500b\u65b9</td></tr><tr><td>1. \u7c21\u4ecb \u5411\uff0c\u4e00\u3001\u6839\u64da\u8a9e\u8a00\u5b78\u6240\u767c\u5c55\u51fa\u7684\u6587\u6cd5(grammar)\u5206\u6790\uff0c\u4e8c\u3001\u4ee5\u77e5\u8b58\u70ba\u57fa\u790e\u800c\u767c\u5c55</td></tr><tr><td>\u62dc\u786c\u9ad4\u6280\u8853\u4e0d\u65b7\u9032\u6b65\u7684\u8ca2\u737b\u4e4b\u4e0b\uff0c\uff0c\u4e00\u822c\u4eba\u6703\u5f88\u7406\u6240\u7576\u7136\u7684\u4f7f\u7528\u81ea\u52d5\u6ac3\u54e1\u6a5f \u7684\u8a9e\u8a00\u8cc7\u6599\u5eab\uff0c\u4e09\u3001\u8457\u91cd\u65bc\u7d71\u8a08\u800c\u767c\u5c55\u51fa\u7684 n-gram \u6a21\u578b\u3002\u800c\u6211\u5011\u4e3b\u8981\u662f\u8457\u58a8\u65bc</td></tr><tr><td>\u63d0\u6b3e\u6216\u662f\u5229\u7528\u81ea\u52d5\u7a7a\u8abf\u8a2d\u5099\u4f86\u63a7\u5236\u5ba4\u5167\u7684\u6eab\u5ea6\uff0c\u800c\u9019\u90fd\u662f\u7531\u65bc\u96fb\u8166\u7684\u81ea\u52d5\u5316\u7ba1\u7406 \u7d71\u8a08\u5f0f\u7684 n-gram \u6a21\u578b\uff0c\u5728\u7b2c\u4e8c\u7ae0\u4e2d\uff0c\u6211\u5011\u5c07\u5c0d n-gram \u6a21\u578b\u505a\u8a73\u7d30\u7684\u4ecb\u7d39\uff0c\u4e26\u5c0d</td></tr><tr><td>\u8b93\u751f\u6d3b\u8b8a\u7684\u5982\u6b64\u4fbf\u5229\uff0c\u6b63\u6240\u8b02 \"\u79d1\u6280\u59cb\u7d42\u4f86\u81ea\u4eba\u6027\" \uff0c\u63a8\u52d5\u79d1\u6280\u9032\u6b65\u7684\u90a3\u96bb\u5e55\u5f8c \u5176\u7f3a\u9ede\u52a0\u4ee5\u63a2\u8a0e\uff0c\u7b2c\u4e09\u7ae0\u4e2d\u5c07\u6703\u4ecb\u7d39\u50b3\u7d71\u4e0a\u91dd\u5c0d n-gram \u6a21\u578b\u7684\u7f3a\u9ede\u6240\u884d\u751f\u51fa\u7684</td></tr><tr><td>\u6539\u9032\u65b9\u6cd5\uff0c\u4e26\u4e14\u63d0\u51fa\u4e00\u7a2e\u80fd\u64f7\u53d6\u9577\u8ddd\u96e2\u8cc7\u8a0a\u7684\u8a9e\u8a00\u6a21\u578b\uff0c\u5c07\u5b83\u61c9\u7528\u5728\u8a9e\u97f3\u8fa8\u8b58\u6216</td></tr></table>",
                "html": null,
                "num": null,
                "type_str": "table"
            },
            "TABREF1": {
                "text": "",
                "content": "<table><tr><td colspan=\"8\">1 ( , ,....., ) 2 T P S P W W ( ) W</td></tr><tr><td/><td/><td colspan=\"6\">1 ( ) ( | )..... ( | 2 1 T P W P W W P W W W 1 , 2</td><td>,.....,</td><td>T W</td><td>1 \u2212</td><td>)</td></tr><tr><td/><td/><td colspan=\"2\">T</td><td/><td/><td/></tr><tr><td/><td/><td>i</td><td>1 =</td><td colspan=\"4\">1 2 P W WW ( | i</td><td>,......,</td><td>i W</td><td>1 \u2212</td><td>)</td></tr><tr><td colspan=\"8\">\u57fa\u672c\u4e0a\u5728\u8a55\u4f30\u4e00\u500b n-gram \u6a21\u578b\u7684\u6548\u679c\u6642\u5e38\u4f7f\u7528 perplexity[12]\u9019\u500b\u8a55\u4f30\u6a19</td></tr><tr><td colspan=\"8\">\u6e96\uff0c\u800c\u4e8b\u5be6\u4e0a\u5b83\u662f\u5728\u8a08\u7b97\u6a5f\u7387\u6a21\u578b\u7684 entropy\uff0centropy \u5728\u8a0a\u606f\u7406\u8ad6\u4e0a\u6307\u7684\u662f\u5c07\u6a5f</td></tr><tr><td colspan=\"8\">\u7387 P \u4e58\u4ee5\u8cc7\u8a0a-logP\uff0c\u61c9\u7528\u5728 n-gram \u6a21\u578b\u7684\u8a55\u4f30\u5247\u8868\u793a\u70ba:</td></tr><tr><td>H</td><td>p</td><td colspan=\"3\">= \u2212</td><td>P</td><td>log</td><td>P</td></tr><tr><td/><td/><td/><td/><td/><td/><td/><td>(6)</td></tr><tr><td colspan=\"8\">\u5176\u7269\u7406\u610f\u7fa9\u8868\u793a\u5728\u8a08\u7b97\u4e00\u500b n-gram \u6a21\u578b\u7684 entropy \u6642\uff0c\u5fc5\u9808\u5148\u5c07\u8a5e\u5178\u4e2d\u7684\u8a5e\u505a</td></tr><tr><td colspan=\"8\">\u7d44\u5408\uff0c\u5f62\u6210\u70ba\u7121\u9650\u9577\u7684\u8a5e\u5e8f\u5217 1 2 ...... Q WW W \uff0c\u4e26\u4e14\u5c07\u6240\u6709\u7684\u53ef\u80fd\u8a5e\u5e8f\u5217\u8a08\u7b97\u5176\u6a5f\u7387</td></tr><tr><td colspan=\"8\">\u8207\u8cc7\u8a0a\u7684\u4e58\u7a4d\u5f8c\u52a0\u7e3d\uff0c\u5373\u53ef\u5f97\u5230\u6b64 n-gram \u6a21\u578b\u7684 entropy\u3002\u4f46\u4e8b\u5be6\u4e0a\u4e0d\u5bb9\u6613\u5be6\u73fe</td></tr><tr><td colspan=\"8\">\u5982\u6b64\u8907\u96dc\u7684\u8a08\u7b97\uff0c\u5fc5\u9808\u5047\u8a2d\u53ef\u4ee5\u63d0\u4f9b\u4e00\u6bb5\u8db3\u5920\u9577\u7684\u8a5e\u5e8f\u5217\u4f86\u4ee3\u8868\u6240\u6709\u7684\u8a5e\u5e8f\u5217\u7d44</td></tr></table>",
                "html": null,
                "num": null,
                "type_str": "table"
            },
            "TABREF2": {
                "text": "\u9577\u8ddd\u96e2\u8cc7\u8a0a(long distance)\u7f3a\u4e4f\u4e4b\u554f\u984c : n-gram \u6a21\u578b\u5728\u8a08\u7b97\u4e0a\u7684\u512a\u52e2\u662f\u5728\u65bc\u5b83\u4f7f\u7528\u4e86 n-gram \u8996\u7a97(n-gram window)\u505a",
                "content": "<table><tr><td>\u5408\uff0c\u9019\u7a2e\u5047\u8a2d\u5728\u7d71\u8a08\u5b78\u4e0a\u7a31\u70ba\u6b64\u8a5e\u5e8f\u5217\u70ba ergodic\uff0c\u6545(8)\u5f0f\u53ef\u6539\u5beb\u70ba 1 2 1 ( ) log ( ...... ) p Q H PW W W Q = \u2212 3.\u70ba\u57fa\u790e\uff0c\u7bc0\u7701\u4e86\u5927\u91cf\u7684\u8a18\u61b6\u9ad4\u8207\u904b\u7b97\u6642\u9593\uff0c\u4f46\u4e5f\u56e0\u70ba\u4f7f\u7528\u4e86\u9019\u500b\u6982\u5ff5\u4f7f\u5f97 n-gram (7)</td></tr><tr><td>\u800c perplexity \u7684\u5b9a\u7fa9\u70ba \u6a21\u578b\u53ea\u80fd\u64f7\u53d6\u5230\u8996\u7a97\u4e4b\u5167\u7684\u8cc7\u8a0a\uff0c\u9577\u8ddd\u96e2\u7684\u8cc7\u8a0a\u5c31\u56e0\u6b64\u800c\u6d41\u5931\uff0c\u800c\u9019\u4e9b\u6d41\u5931\u7684\u8cc7</td></tr><tr><td>2 p H \u8a0a\u5f88\u53ef\u80fd\u6703\u9020\u6210 n-gram \u6a21\u578b\u6e2c\u8a66\u6642\u76f8\u7576\u7a0b\u5ea6\u7684\u8aa4\u5dee\uff0c\u6545\u5982\u4f55\u64f7\u53d6\u9577\u8ddd\u96e2\u7684\u8cc7\u8a0a perplexity = (8)</td></tr><tr><td>perplexity \u4ee3\u8868\u4e86 n-gram \u6a21\u578b\u4e2d\u7684\u5e73\u5747\u5206\u652f\u56e0\u6578(average branching factor)\uff0c \u4e00\u76f4\u90fd\u662f n-gram \u6a21\u578b\u4e2d\u76f8\u7576\u53d7\u5230\u77da\u76ee\u7684\u7814\u7a76\u7684\u8ab2\u984c\u3002\u4e00\u822c\u800c\u8a00\u76ee\u524d n-gram \u6a21</td></tr><tr><td>perplexity \u8d8a\u4f4e\u4ee3\u8868 n-gram \u6a21\u578b\u5728\u505a\u6a5f\u7387\u8a55\u4f30\u6642\uff0c\u6240\u9047\u5230\u7684\u5206\u652f\u8d8a\u5c11\uff0c\u4e5f\u5c31\u662f\u6b64 \u578b\u7684\u7814\u7a76\u5747\u4ee5\u89e3\u6c7a\u6b64\u4e09\u9805\u554f\u984c\u70ba\u4e3b\uff0c\u672c\u8ad6\u6587\u91dd\u5c0d\u4e0a\u8ff0\u7b2c\u4e09\u9805\u9577\u8ddd\u96e2\u8cc7\u8a0a\u7684\u64f7\u53d6\u63d0</td></tr><tr><td>1 i n i PW W \u2212 ( | i \u2212 + 1 ) \u2245 \u220f 1 T i = \u51fa\u6539\u9032\u65b9\u6cd5\uff0c\u671f\u671b\u80fd\u63d0\u6607 n-gram \u6a21\u578b\u7684\u6548\u679c\u3002 1 2 3 ( ) ( , , ,....., ) T P S PW W W W = \u6a21\u578b\u7684\u6548\u7387\u8d8a\u597d\u3002</td></tr><tr><td>2.4 n-gram \u6a21\u578b\u7684\u7f3a\u9ede 3. n-gram \u6a21\u578b\u6539\u9032\u65b9\u5411</td></tr><tr><td>n-gram \u6a21\u578b\u9577\u4e45\u5df2\u4f86\u5c31\u5b58\u5728\u8457\u4e09\u500b\u91cd\u8981\u7684\u554f\u984c\uff0c\u4e5f\u662f\u7814\u7a76 n-gram \u6a21\u578b\u7684\u4eba</td></tr><tr><td>\u4e00\u76f4\u52aa\u529b\u7684\u76ee\u6a19\uff0c\u6211\u5011\u5206\u8ff0\u5982\u4e0b:</td></tr><tr><td>1.\u8a13\u7df4\u6587\u96c6\u8207\u6e2c\u8a66\u6587\u96c6\u9818\u57df\u4e0a\u4e4b\u5dee\u8ddd(domain mismatch) :</td></tr><tr><td>n-gram \u6a21\u578b\u5728\u5efa\u7acb\u6642\uff0c\u5fc5\u9808\u8981\u6709\u4e00\u8a13\u7df4\u6587\u96c6\u4f86\u7d71\u8a08\u51fa\u6b64\u6a21\u578b\u7684\u6a5f\u7387\uff0c\u56e0\u6b64</td></tr><tr><td>n-gram \u6a21\u578b\u53d7\u5236\u65bc\u5b83\u7684\u8a13\u7df4\u6587\u96c6\uff0c\u7576\u8a13\u7df4\u6587\u96c6\u4e0d\u5e73\u5747\u6642\u53ef\u80fd\u6703\u4f7f n-gram \u6a21\u578b\u8f03</td></tr><tr><td>\u504f\u5411\u67d0\u7a2e\u9818\u57df(domain)\uff0c\u5047\u8a2d\u6211\u5011\u7684\u8a13\u7df4\u6587\u96c6\u662f\u8ca1\u7d93\u985e\u7684\u65b0\u805e\uff0c\u4f46\u662f\u6b64 n-gram \u6a21 (Trigger pair)[8][9]\u7684\u7c21\u4ecb\uff0c\u89f8\u767c\u5e8f\u5c0d\u662f\u5728\u64f7\u53d6\u9577\u8ddd\u96e2\u8cc7\u8a0a\u7684\u4e00\u7a2e\u6709\u6548\u7684\u65b9\u6cd5\uff0c\u53ef</td></tr><tr><td>\u578b\u7684\u76ee\u7684\u662f\u7528\u4f86\u6e2c\u8a66\u653f\u6cbb\u65b0\u805e\uff0c\u90a3\u9ebc\u5c31\u6703\u9020\u6210\u8f03\u5927\u7684\u8aa4\u5dee\uff0c\u5728\u9019\u65b9\u9762\u901a\u5e38\u6703\u4f7f\u7528 \u4ee5\u7528\u4f86\u88dc\u511f n-gram \u6a21\u578b\u9577\u8ddd\u96e2\u8cc7\u8a0a\u7684\u4e0d\u8db3\uff0c\u800c\u672c\u8ad6\u6587\u4e5f\u5c07\u63d0\u51fa\u4e00\u7a2e\u6539\u9032\u89f8\u767c\u5e8f</td></tr><tr><td>\u8f03\u4e00\u822c\u5316\u7684\u7684\u5e73\u8861\u6587\u96c6\u4f5c\u70ba\u8a13\u7df4\u6587\u96c6\u4f86\u89e3\u6c7a\u9019\u500b\u554f\u984c\u3002\u4f46\u77db\u76fe\u7684\u662f\u5982\u679c\u6211\u5011\u4f7f\u7528 \u5c0d\u7684\u65b9\u6cd5\u70ba\u5c0d\u7167\u7d44\uff0c\u4e26\u5728\u5be6\u9a57\u4e2d\u505a\u6bd4\u8f03\u7814\u7a76\u3002</td></tr><tr><td>\u8f03\u70ba\u5e73\u8861\u7684\u6587\u96c6\u8a13\u7df4\u51fa\u6211\u5011\u7684 n-gram \u6a21\u578b\uff0c\u6b64 n-gram \u6a21\u578b\u7528\u4f86\u6e2c\u8a66\u67d0\u4e9b\u7279\u5b9a\u9818</td></tr><tr><td>\u57df\u7684\u65b0\u805e\u662f\u5426\u6070\u7576?\u4e8b\u5be6\u4e0a\u6211\u5011\u5e0c\u671b\u5728\u6e2c\u8a66\u653f\u6cbb\u65b0\u805e\u6642\u6211\u5011\u7684 n-gram \u6a21\u578b\u662f\u504f\u5411 3.1 \u5feb\u53d6(cache)n-gram \u6a21\u578b\u8207\u6df7\u5408\u5f0f(mixture)n-gram \u6a21\u578b</td></tr><tr><td>\u653f\u6cbb\u985e\u7684\uff0c\u6e2c\u8a66\u8ca1\u7d93\u65b0\u805e\u6642 n-gram \u6a21\u578b\u662f\u504f\u5411\u8ca1\u7d93\u985e\u7684\uff0c\u70ba\u4e86\u8981\u5b8c\u6210\u9019\u9805\u9700\u6c42\uff0c \u70ba\u4e86\u8981\u4f7f n-gram \u6a21\u578b\u80fd\u5920\u66f4\u7b26\u5408\u6e2c\u8a66\u6642\u7684\u9818\u57df\uff0c\u6240\u4ee5\u7522\u751f\u4e86\u6a21\u578b\u8abf\u6574\u7684\u6982</td></tr><tr><td>\u5c31\u5fc5\u9808\u5c0d n-gram \u6a21\u578b\u518d\u505a\u6539\u9032\uff0c\u4f7f\u5176\u5177\u6709\u8abf\u6574\u4e4b\u6548\u679c[3][4][5][7]\u3002 \u5ff5\uff0c\u5b83\u7684\u6982\u5ff5\u662f\u57fa\u65bc\u3127\u7bc7\u6587\u7ae0\u6216\u662f\u4e00\u6bb5\u6587\u53e5\u6703\u6709\u4e00\u500b\u8fd1\u4f3c\u7684\u4e3b\u984c\uff0c\u6bd4\u65b9\u8aaa\u68d2\u7403\u985e</td></tr><tr><td>\u7684\u65b0\u805e\u5c31\u6bd4\u8f03\u504f\u5411\u904b\u52d5\u985e\u7684\u9818\u57df\uff0c\u8207\u5176\u4ed6\u985e\u5225\u7684\u65b0\u805e(\u5982\u8ca1\u7d93\u65b0\u805e)\u5c31\u6709\u4e00\u6bb5\u76f8\u7576</td></tr><tr><td>1 2 P WW W ( ... )log ( Q P WW W 1 2 ... ) Q \u5927\u7684\u5dee\u8ddd\uff0c\u800c\u5e0c\u671b\u80fd\u5728\u505a\u6e2c\u8a66\u6642\uff0c\u5229\u7528\u6587\u7ae0\u524d\u9762\u51fa\u73fe\u7684\u8cc7\u8a0a\uff0c\u52d5\u614b\u7684\u8abf\u6574\u6211\u5011\u7684 ... 1 lim Q Q WW W Q \u2192\u221e = \u2212 2.\u8a13\u7df4\u6587\u96c6\u4e0d\u8db3(data spareness) : \u2211 n-gram \u6a21\u578b\u5728\u8a13\u7df4\u6642\uff0c\u4e26\u4e0d\u80fd\u4fdd\u8b49\u8a13\u7df4\u6587\u96c6\u80fd\u5920\u5305\u542b\u6240\u6709\u8a5e\u7684\u7d44\u5408\uff0c\u4ee5\u81f3 \u65bc\u6240\u8a13\u7df4\u51fa\u4f86\u7684\u6a5f\u7387\u6a21\u578b\u67d0\u4e9b\u8a5e\u7d44\u76f8\u9023\u7684\u6a5f\u7387\u70ba\u96f6\uff0c\u6216\u662f\u56e0\u70ba\u8a13\u7df4\u6587\u96c6\u7684\u4e0d\u5e73 n-gram \u6a21\u578b\uff0c\u4f7f\u5f97\u6211\u5011\u7684\u6a21\u578b\u66f4\u80fd\u7b26\u5408\u6211\u5011\u6e2c\u8a66\u6587\u96c6\u7684\u9818\u57df\uff0c\u57fa\u65bc\u6b64\u7a2e\u6982\u5ff5\uff0c</td></tr><tr><td>\u5c31\u6709\u5feb\u53d6\u6a21\u578b\u8207\u6df7\u5408\u5f0f\u6a21\u578b\u7684\u6280\u8853\u7522\u751f\u3002</td></tr><tr><td>\u8861\uff0c\u9020\u6210\u7d71\u8a08\u51fa\u4f86\u7684\u6a5f\u7387\u6a21\u578b\u4e26\u4e0d\u5920\u4e00\u822c\u5316\uff0c\u800c\u70ba\u4e86\u89e3\u6c7a\u9019\u500b\u554f\u984c\uff0c\u5c31\u6709\u5e73\u6ed1\u5316</td></tr><tr><td>\u6280\u8853\u7684\u7522\u751f\uff0c\u5728\u53c3\u8003\u6587\u737b[2]\u4e2d\u5c0d\u50b3\u7d71\u4e0a\u53d7\u6b61\u8fce\u7684\u5e73\u6ed1\u5316\u6280\u8853\u6709\u8a73\u76e1\u7684\u8aaa\u660e\u3002 \u5feb\u53d6 n-gram \u6a21\u578b\u9867\u540d\u601d\u7fa9\u5c31\u662f\u76f8\u540c\u7684\u8a5e\u5e8f\u5217\u6703\u5728\u9130\u8fd1\u7684\u6642\u9593\u9ede\u4e0a\u4e0d\u65b7\u51fa</td></tr><tr><td>\u73fe\uff0c\u6bd4\u65b9\u8aaa\u6211\u5011\u7684\u6e2c\u8a66\u6587\u96c6\u662f\u4e00\u7bc7\u6709\u95dc\u91d1\u878d\u80a1\u7684\u65b0\u805e\uff0c\u4e5f\u5c31\u662f\u8aaa\u6b64\u7bc7\u6587\u4ef6\"\u91d1\u878d</td></tr></table>",
                "html": null,
                "num": null,
                "type_str": "table"
            },
            "TABREF3": {
                "text": "\u4ee3\u8868\u5728\u6587\u53e5 l \u7684\u8a5e\u6578\uff0cN k \u8868\u793a\u5728\u7fa4\u7d44 k \u7684\u7e3d\u6587\u53e5\u6578\u3002",
                "content": "<table><tr><td colspan=\"10\">\u5176\u4e2d T l (13)</td></tr><tr><td colspan=\"10\">\u5176\u4e2d N \u4ee3\u8868\u8abf\u6574\u7684\u7e3d\u6587\u53e5\u6578\u3002\u6b0a\u91cd\u7684\u8abf\u6574\u7684\u4e3b\u8981\u6839\u64da\u6e2c\u8a66\u6642\u6587\u4ef6\u6240\u51fa\u73fe\u7684\u8cc7\u8a0a\uff0c</td></tr><tr><td colspan=\"10\">\u6df7\u5408\u5f0f n-gram \u6a21\u578b\u6703\u4f9d\u524d\u6587\u5728\u6bcf\u500b\u7fa4\u7d44\u6a21\u578b\u51fa\u73fe\u7684\u6a5f\u7387\u70ba\u6b0a\u91cd\uff0c\u52d5\u614b\u7684\u8abf\u6574\u6e2c</td></tr><tr><td colspan=\"10\">\u8a66\u6a21\u578b\u7684\u7d44\u5408\u6b0a\u91cd\uff0c\u6bd4\u65b9\u8aaa\u5728\u6e2c\u8a66\u6587\u4ef6\u4e2d\u4e0d\u65b7\u63d0\u5230\u91d1\u878d\u6d88\u606f\uff0c\u6df7\u5408\u5f0f n-gram \u6a21</td></tr><tr><td colspan=\"10\">\u578b\u5c31\u6703\u5c07\u6a21\u578b\u9010\u6b65\u7684\u8abf\u6574\u5230\u8ca1\u7d93\u9818\u57df\uff0c\u518d\u5229\u7528\u9019\u8abf\u6574\u904e\u5f8c\u4e4b n-gram \u6a21\u578b\u7e7c\u7e8c\u6e2c</td></tr><tr><td colspan=\"10\">\u8a66\u5f8c\u9762\u7684\u6587\u53e5\uff0c\u7136\u5f8c\u518d\u5c07\u6e2c\u8a66\u800c\u5f97\u7684\u65b0\u8cc7\u8a0a\u7e7c\u7e8c\u505a\u8abf\u6574\uff0c\u9019\u7a2e\u905e\u8ff4\u5f0f\u7684\u505a\u6cd5\u662f\u4e00</td></tr><tr><td colspan=\"7\">\u7a2e\u7a31\u70ba\u8cc7\u8a0a\u7d50\u69cb(Information structure)\u7684\u6982\u5ff5\u3002</td><td/><td/><td/></tr><tr><td>P S</td><td>=</td><td>+ = \u220f</td><td>P W W</td><td>\u2212 \u2212 +</td><td>=</td><td>ik + == \u220f\u2211</td><td>\u03bb</td><td>P W W</td><td>\u2212 \u2212 +</td></tr></table>",
                "html": null,
                "num": null,
                "type_str": "table"
            },
            "TABREF4": {
                "text": "\u76f8\u7368\u7acb\uff0c\u5982\u679c\u5728\u6240\u6709\u5305\u542b X \u7684\u6587\u7ae0\u4e2d\u6709 c% \u540c\u6642\u4e5f\u5305\u542b\u4e86 Y\uff0c\u5247\u6211\u5011\u53ef\u4ee5\u7a31\u95dc\u806f \u6cd5\u5247 X Y \u21d2 \u5b58\u5728\u65bc\u8cc7\u6599\u5eab D \u4e2d\u7684\u4fe1\u8cf4\u5ea6(confidence)\u70ba c\uff0c\u6b64\u5916\u82e5\u6709 s% \u7684\u6587\u7ae0\u540c \u6642\u5305\u542b X \u8207 Y\uff0c\u5247\u6211\u5011\u53ef\u7a31\u95dc\u806f\u6cd5\u5247 X Y \u21d2 \u4ee5\u652f\u6301\u5ea6(support) s \u5b58\u5728\u65bc\u8cc7\u6599\u5eab D",
                "content": "<table><tr><td colspan=\"10\">\u7684\u8a5e\u975e\u5e38\u591a\uff0c\u6b64\u6642 unigram \u7684\u6b0a\u91cd\u53ef\u4ee5\u9069\u5ea6\u52a0\u5927\uff0c\u4ee5\u5f4c\u88dc\u53ef\u80fd\u8f03\u591a\u7684\u8cc7\u8a0a\u640d\u5931\uff0c \u6bd4\u8d77 n-gram \u6a21\u578b\u591a\u4e86\u9577\u8ddd\u96e2\u7684\u8cc7\u8a0a\uff0c\u70ba\u4e86\u65b9\u4fbf\u8d77\u898b\u4f7f\u7528\u5c0d\u6578\u8868\u793a\u70ba</td></tr><tr><td>\u4f7f\u8a9e\u8a00\u6a21\u578b\u7684\u6e96\u78ba\u6027\u63d0\u9ad8\u3002 1 log ( ) log ( ) T i i P S PW = = \u2211</td><td>+</td><td>max(1, i ws 1 j i \u2212 = \u2212 \u2211 \u2211 2 i T =</td><td>)</td><td>( M I T r i g g e rW \u2212</td><td>j</td><td>\u2192</td><td>i W</td><td>)</td><td>(19)</td></tr><tr><td colspan=\"10\">3.3 \u89f8\u767c\u5e8f\u5c0d (Trigger Pair) \u6f14\u7b97\u6cd5 \u5176\u4e2d logP(W i ) \u5373\u70ba unigram \u6a21\u578b\u6a5f\u7387\uff0cws \u4ee3\u8868 window size\uff0c\u73fe\u5728\u5728\u6211\u5011\u7684\u8ad6\u6587 \u4e2d\uff0c\u63db\u53e5\u8a71\u8aaa\uff0c\u4fe1\u8cf4\u5ea6\u662f\u4e00\u7a2e\u91cf\u6e2c\u95dc\u806f\u6cd5\u5247\u5f37\u5f31\u7684\u6a19\u6e96\uff0c\u800c\u652f\u6301\u5ea6\u5247\u662f\u8868\u793a\u7d71\u8a08</td></tr><tr><td colspan=\"10\">\u5728\u81ea\u7136\u8a9e\u8a00\u4e2d\uff0c\u5b58\u5728\u8457\u8a31\u591a\u9ad8\u5ea6\u95dc\u806f\u6027\u7684\u8a5e\u7d44\uff0c\u6bd4\u65b9\u8aaa\"\u91ab\u751f\u3001\"\u8b77\u58eb\"\u6216\u662f\"\u967d \u4e2d\u5c07 window size \u5b9a\u70ba\u6587\u53e5\u9577\u5ea6\uff0c\u4e5f\u5c31\u662f\u8aaa\u5728\u6211\u5011\u8ad6\u6587\u4e2d\u7684\u89f8\u767c\u5e8f\u5c0d\u662f\u6587\u53e5\u968e\u5c64 \u4e0a\u51fa\u73fe\u7684\u983b\u7387\uff0c\u4e8b\u5be6\u4e0a\u6211\u5011\u5be6\u4f5c\u6642\u6703\u8a02\u5b9a\u4fe1\u8cf4\u5ea6\u8207\u652f\u6301\u5ea6\u7684\u9580\u6abb\uff0c\u6211\u5011\u64f7\u53d6\u51fa\u4f86</td></tr><tr><td colspan=\"10\">\u5149\"\u3001\"\u71b1\"\u7b49\u5c31\u7d93\u5e38\u51fa\u73fe\u65bc\u65bc\u540c\u4e00\u53e5\u5b50\u4e4b\u4e2d\uff0c\u4f46\u7531\u65bc\u5b83\u5011\u901a\u5e38\u5728\u53e5\u5b50\u4e2d\u4e26\u4e0d\u76f8\u9023\uff0c \u7684\u89f8\u767c\u5e8f\u5c0d(sentence-level trigger pair) \uff0c\u4ee3\u8868\u6211\u5011\u53ea\u80fd\u64f7\u53d6\u540c\u4e00\u6587\u53e5\u4e2d\u7684\u89f8\u767c\u5e8f\u5c0d \u4e4b\u95dc\u806f\u6cd5\u5247\u7684\u4fe1\u8cf4\u5ea6\u8207\u652f\u6301\u5ea6\u5747\u5fc5\u9808\u5927\u65bc\u6b64\u9580\u6abb\u3002</td></tr><tr><td colspan=\"10\">\u6240\u4ee5 n-gram \u6a21\u578b\u4e26\u6c92\u6709\u8fa6\u6cd5\u64f7\u53d6\u5230\u9019\u4e9b\u8a5e\u4e4b\u9593\u7684\u95dc\u806f\u8cc7\u8a0a\uff0c\u56e0\u6b64\u5c31\u6709\u4e86\u89f8\u767c\u5e8f \u8cc7\u8a0a\u3002\u73fe\u5728\u6211\u5011\u5fc5\u9808\u5c07\u89f8\u767c\u5e8f\u5c0d\u52a0\u5165 n-gram \u6a21\u578b\u4e4b\u4e2d\u505a\u70ba\u9577\u8ddd\u96e2\u8cc7\u8a0a\u64f7\u53d6\u4e4b\u8f14 \u4ee5\u4e0b\u5373\u70ba\u64f7\u53d6\u95dc\u806f\u6cd5\u5247\u7684\u6f14\u7b97\u6cd5\u6d41\u7a0b\uff0c\u662f\u4ee5\u8cc7\u6599\u63a2\u52d8\u4e2d\u7684 Apriori \u6f14\u7b97\u6cd5\u505a</td></tr><tr><td colspan=\"10\">\u5c0d\u7684\u7522\u751f\uff0c\u89f8\u767c\u5e8f\u5c0d\u7684\u8a2d\u8a08\u4e3b\u8981\u5728\u65bc\u89e3\u6c7a\u9577\u8ddd\u96e2\u8cc7\u8a0a\u5f4c\u88dc n-gram \u6a21\u578b\u7684\u4e0d\u8db3\u7684 \u52a9\uff0c\u900f\u904e\u7dda\u6027\u63d2\u88dc(linear interpolation)\u7684\u65b9\u5f0f\uff0c\u6211\u5011\u53ef\u4ee5\u4e00\u6b0a\u91cd a i \u5c07\u5176\u505a\u5408\u4f75\uff0c \u70ba\u57fa\u790e\u6240\u6539\u5beb\u800c\u6210\u82e5\u6211\u5011\u4ee5\u7c21\u55ae\u7684\u4f8b\u5b50\u8aaa\u660e\u4e4b\uff0c\u5047\u8a2d\u6211\u5011\u5171\u6709\u4e09\u8a5e\uff0c\u5206\u5225\u4ee5 a\u3001</td></tr><tr><td colspan=\"10\">\u554f\u984c\uff0c\u89f8\u767c\u5e8f\u5c0d\u7531\u65bc\u5176\u6c92\u6709\u6f14\u7b97\u6cd5\u8207\u8cc7\u6599\u7d50\u69cb\u53ef\u4ee5\u5feb\u901f\u7684\u5c0d\u8cc7\u6599\u5eab\u505a\u6c42\u53d6\uff0c\u6545\u89f8 \u4e5f\u5c31\u662f b\u3001c \u4ee3\u8868\uff0cApriori \u6f14\u7b97\u6cd5\u5c31\u662f\u5728\u627e\u5c0b\u6b64\u4e09\u8a5e\u7684\u95dc\u806f\u6027\uff0c\u5b83\u7684\u6982\u5ff5\u5c31\u662f\u5148\u5c07\u9019\u4e9b</td></tr><tr><td colspan=\"10\">\u767c\u5e8f\u5c0d\u6703\u9650\u5236\u672c\u8eab\u70ba \"\u5e8f\u5c0d\" \u3001\u5373\u82e5\u6709\u4e00\u8fad\u5178 V\uff0c\u89f8\u767c\u5e8f\u5c0d\u6703\u5c0d\u5176\u4e2d\u6240\u6709\u53ef\u80fd\u7684 \u8a5e\u5e8f\u5c0d\u505a\u8003\u616e\uff0c\u5982\u6b64\u4e00\u4f86\u53ef\u5c07\u4fc3\u767c\u5e8f\u5c0d\u7684\u7e3d\u500b\u6578\u63a7\u5236\u65bc|V| 2 \u5167\u3002 1 log ( ) log ( ) k MERGED i i i P S a P S \u5176\u4e2d 0 1 i a \u2264 \u2264 \u4e14 k 1 i a = \uff0c\u5728\u9019\u908a\u6211\u5011\u6709\u5169\u500b\u6a21\u578b\u6a5f\u7387\u5b58\u5728 \u2211 = = \u22c5 \u2211 (20) \u8a5e\u5169\u5169\u70ba\u4e00\u7d44\u5efa\u7acb\u5e8f\u5c0d\u96c6\u5408(a</td></tr><tr><td>i=1</td><td/><td/><td/><td/><td/><td/><td/><td/><td/></tr><tr><td colspan=\"7\">1. P 1 (S) = P n-gram (S) \u70ba n-gram \u6a21\u578b\u5c0d\u6587\u53e5 S \u6240\u4f30\u6e2c\u51fa\u4e4b\u6a5f\u7387\u3002</td><td/><td/><td/></tr><tr><td colspan=\"9\">2. P 2 (S) = P MI-Trigger-pair (S) \u70ba\u89f8\u767c\u5e8f\u5c0d\u6a21\u578b\u5c0d\u6587\u53e5 S \u6240\u4f30\u6e2c\u51fa\u4e4b\u6a5f\u7387\u3002</td><td/></tr><tr><td colspan=\"10\">\u900f\u904e(20)\u5f0f\u7684\u8a08\u7b97\uff0c\u6211\u5011\u53ef\u4ee5\u4f7f\u7528\u89f8\u767c\u5e8f\u5c0d\u8a08\u7b97\u51fa\u4e00\u6bb5\u6587\u53e5\u7684\u6a5f\u7387\uff0c\u4e14\u6b64\u6a5f\u7387\u6709</td></tr><tr><td colspan=\"9\">( , ) ( ) ( ) i j i j P W W P W P W \u9577\u8ddd\u96e2\u8cc7\u8a0a\u5b58\u5728\uff0c\u6bd4\u8d77\u50b3\u7d71\u7684 n-gram \u6a21\u578b\u5728\u8cc7\u8a0a\u64f7\u53d6\u4e0a\u70ba\u4f73\u3002 ( , ) ( ; ) ( , )log ( , )log ( ) ( ) i j i j i j i j i j AMI W W P W W P W W P W P W = + P W W</td><td/></tr><tr><td colspan=\"10\">( , ) ( ) ( ) i j i j P W W P W P W 4. \u95dc\u806f\u6cd5\u5247\u8207\u5176\u61c9\u7528 ( , ) i j P W W + \u5728\u9019\u908a\u6211\u5011\u5f15\u5165\u4e86\u4e00\u500b\u5728\u8cc7\u6599\u63a2\u52d8(Data Ming)\u9818\u57df\u53d7\u5230\u5341\u5206\u5ee3\u6cdb\u904b\u7528\u7684 ( , ) ( , ) ( ) ( ) i j i j i j P W W P W W + P W P W</td></tr><tr><td colspan=\"10\">Apriori \u6f14\u7b97\u6cd5[1]\uff0c\u6b64\u6f14\u7b97\u6cd5\u53ef\u4ee5\u7528\u4f86\u5efa\u7acb\u95dc\u9375\u8a5e\u7684\u95dc\u806f\u6cd5\u5247\uff0c\u8209\u4f8b\u800c\u8a00\uff0c\u5047\u8a2d</td></tr><tr><td colspan=\"10\">i P W W \u4ee3\u8868\u5728\u540c\u4e00\u500b\u8996 j \u6709\u4e00\u7d44\u4ea4\u6613\u7d00\u9304\u8cc7\u6599\u5eab\uff0c\u6b64\u8cc7\u6599\u5eab\u8a18\u9304\u8457\u6bcf\u7b46\u4ea4\u6613\u6240\u5305\u542b\u7684\u5546\u54c1\uff0c\u95dc\u806f\u6cd5\u5247\u6240\u8981</td></tr><tr><td colspan=\"10\">W \u7a97\u4e2d\u53ea\u51fa\u73fe W i ,\u800c\u6c92\u51fa\u73fe W j \u7684\u6a5f\u7387\u3002\u900f\u904e AMI \u8a55\u4f30\u6a19\u6e96\uff0c\u6211\u5011\u5c07\u5176\u9078\u70ba\u89f8\u767c\u5e8f C W \u03bb \u2212 \u2212 + \u2212 + \u2212 + \u2212 + \u2212 + \u2212 + \u22c5 \u2212 \u64f7\u53d6\u7684\u5c31\u662f\u6bcf\u500b\u5546\u54c1\u9593\u7684\u76f8\u4e92\u95dc\u4fc2\uff0c\u4e5f\u5c31\u662f\u8aaa\u6211\u5011\u60f3\u77e5\u9053\u4e00\u7b46\u4ea4\u6613\u51fa\u73fe\u4e86\u67d0\u7a2e\u5546 = \u22c5 + \u2211 \u5c0d\uff0c\u4ee5\u7b26\u865f( i j W W \u2192 ) \u8868\u793a\u3002\u7576\u5e8f\u5c0d\u9078\u53d6\u5b8c\u7562\u5f8c\uff0c\u5fc5\u9808\u8981\u5c0d\u6bcf\u500b\u89f8\u767c\u5e8f\u5c0d\u8a08\u7b97\u5176 \u54c1\u5f8c\uff0c\u9084\u6709\u54ea\u4e9b\u5546\u54c1\u662f\u53ef\u80fd\u51fa\u73fe\u5728\u540c\u4e00\u7b46\u4ea4\u6613\u7d00\u9304\u4e4b\u4e2d\uff0c\u5982\u679c\u8aaa\u5546\u5bb6\u5f9e\u95dc\u806f\u6cd5\u5247</td></tr><tr><td colspan=\"10\">\u76f8\u4e92\u8cc7\u8a0a MI (mutual information)\uff0c\u7528\u5c0d\u6578\u8868\u793a\u4e4b\u5982\u4e0b \u4e2d\u77e5\u9053\u9867\u5ba2\u8cb7\u4e86\u5546\u54c1\u7532\u5f8c\uff0c\u9084\u6709\u5f88\u5927\u7684\u6a5f\u7387\u6703\u53bb\u8cb7\u5546\u54c1\u4e59\uff0c\u5247\u53ef\u5c07\u5546\u54c1\u7532\u8207\u5546\u54c1</td></tr><tr><td colspan=\"5\">( , ) ( ) ( ) i j i j P W W P W P W \u4e59\u653e\u5728\u9644\u8fd1\u589e\u52a0\u9867\u5ba2\u7684\u65b9\u4fbf\u6027\u8207\u5546\u5bb6\u7684\u696d\u7e3e\u3002 ( ; ) log i j MI W W =</td><td/><td/><td/><td/><td>(18)</td></tr><tr><td colspan=\"10\">\u5982\u679c W i \u548c W j \u662f\u76f8\u4e92\u7368\u7acb\u7684\u8a71\uff0c\u5247 MI(W i , W j ) = 0\uff0c\u76f8\u4e92\u8cc7\u8a0a\u53cd\u6620\u4e86\u89f8\u767c\u5e8f\u5c0d\u4e2d\u5169 4.1 Apriori \u6f14\u7b97\u6cd5</td></tr><tr><td colspan=\"10\">\u500b\u8a5e\u76f8\u4e92\u9593\u7684\u8cc7\u8a0a\u8b8a\u5316\u3002\u800c\u89f8\u767c\u5e8f\u5c0d\u4e26\u7121\u6cd5\u55ae\u7368\u4f7f\u7528[14]\uff0c\u56e0\u70ba\u5b83\u53ea\u80fd\u53cd\u6620\u51fa\u8a5e \u5047\u8a2d\u6211\u5011\u6709\u4e00\u7d44\u65b0\u805e\u6587\u4ef6\u8cc7\u6599\u5eab D\uff0c\u88e1\u9762\u5305\u542b\u4e86|D|\u7bc7\u6587\u7ae0\uff0c\u6bcf\u7bc7\u6587\u7ae0\u5747\u662f</td></tr><tr><td colspan=\"10\">\u8207\u8a5e\u7684\u8cc7\u8a0a\u8b8a\u5316\uff0c\u6240\u4ee5\u6211\u5011\u5fc5\u9808\u5c07\u5176\u8207 unigram \u505a\u7d50\u5408\uff0c\u5982\u6b64\u4e00\u4f86\u6240\u7372\u5f97\u7684\u8cc7\u8a0a \u8fad\u5178 1 2 { , ,......, } n L w w w = \u7684\u5b50\u96c6\u5408\uff0c\u7528\u4e0a\u9762\u7684\u4f8b\u5b50\u89e3\u91cb\u5c31\u662f\u5404\u7a2e\u5546\u54c1\u7684\u96c6\u5408\u4e4b</td></tr><tr><td colspan=\"10\">\u610f\uff0c\u800c\u95dc\u806f\u6cd5\u5247\u4ee5 X Y \u21d2 \u7684\u578b\u5f0f\u8868\u793a\uff0c\u5176\u4e2d X\u3001Y \u5747\u662f L \u7684\u5b50\u96c6\u5408(subset)\u4e14\u4e92</td></tr></table>",
                "html": null,
                "num": null,
                "type_str": "table"
            }
        }
    }
}