\u5927\u8bed\u8a00\u6a21\u578b\u5728\u5f00\u59cb\u201cAttention\u201d\u8ba1\u7b97\u4e4b\u524d\uff0c\u8fd8\u4f1a\u5bf9\u201ctoken embedding\u201d\u8fdb\u884c\u4e00\u4e9b\u9884\u5904\u7406\uff0c\u8fd9\u4e9b\u9884\u5904\u7406\u5305\u62ec\u4e86\u201c\u878d\u5165\u201d\u4f4d\u7f6e\u5411\u91cf\u3001\u5bf9\u5411\u91cf\u8fdb\u884c\u201c\u5f52\u4e00\u5316\u201d\u5904\u7406\uff08\u5c06\u5404\u4e2a\u5411\u91cf\u90fd\u8f6c\u5316\u4e3a\u5747\u503c\u4e3a0\u3001\u65b9\u5dee\u4e3a1\u7684\u5411\u91cf\uff0c\u957f\u5ea6\u8981\u7edf\u4e00\u53d8\u62101\u5417\uff1f\uff09\u3002<\/p>\n\n\n\n
\u4f8b\u5982\uff0c\u5728\u8fd9\u91cc\u7684\u4f8b\u5b50\u4e2d\uff0c\u63d0\u793a\u8bcd \u201cIt\u2019s very hot in summer. Swimming is<\/span>\u201d\uff0c\u5148\u8f6c\u6362\u4e3aembedding\uff0c\u7136\u540e\u52a0\u4e0a\u4f4d\u7f6e\u7f16\u7801\uff08positional encoder)\u3001\u518d\u8fdb\u884c\u6b63\u89c4\u5316\uff0c\u6700\u540e\u53d8\u6362\u4e3a\u5982\u4e0b\u7684\u5411\u91cf \u201c X \u201d \uff1a<\/p>\n\n\n\n \u8fd9\u91cc\u7684 \u201cX\u201d \u662f\u4e00\u4e2a\u753112\u4e2a \u201ctoken embedding\u201d\u7ec4\u6210\u7684\u77e9\u9635\uff0c\u201c\u5f62\u72b6\u201d\u662f 12 x 768 \u3002\u5728\u6570\u5b66\u7b26\u53f7\u4e0a\uff0c\u6709\uff1a<\/p>\n\n\n\n \u5728\u6b63\u5f0f\u4ecb\u7ecd Attention \u4e4b\u524d\uff0c\u4e3a\u4e86\u80fd\u591f\u6bd4\u8f83\u597d\u7684\u7406\u89e3\u201c\u4e3a\u4ec0\u4e48\u201d\u662f\u8fd9\u6837\uff0c\u8fd9\u91cc\u5148\u5f15\u5165\u4e86\u201cSimilarity\u201d\u7684\u6982\u5ff5\uff0c\u6700\u7ec8\u5728\u8be5\u6982\u5ff5\u4e0a\uff0c\u65b0\u589e\u6743\u91cd\u77e9\u9635\uff0c\u5c31\u662f\u6700\u7ec8\u7684 Attention \uff1a<\/p>\n\n\n\n $$ \\text{Similarity} = \\text{softmax}(\\frac{XX^{T}}{\\sqrt{d}})X $$<\/p>\n\n\n\n \u8fd9\u91cc\u5220\u9664\u4e86\u53c2\u6570\u77e9\u9635\uff1a\\(W^Q \\quad W^K \\quad W^V \\)<\/p>\n<\/div>\n\n\n\n $$ \\text{Attention} = \\text{softmax}(\\frac{QK^{T}}{\\sqrt{d}})V $$<\/p>\n\n\n\n \u5176\u4e2d\uff0c \\(Q = XW^Q \\quad K = XW^K \\quad V = XW^V \\)<\/p>\n<\/div>\n<\/div>\n\n\n\n <\/p>\n\n\n\n \u5728\u5411\u91cf\u4e3a\u5355\u4f4d\u957f\u5ea6\u7684\u65f6\u5019\uff0c\u901a\u5e38\u53ef\u4ee5\u76f4\u63a5\u4f7f\u7528\u201c\u5185\u79ef\u201d\u4f5c\u4e3a\u4e24\u4e2a\u5411\u91cf\u7684\u76f8\u4f3c\u5ea6\u5ea6\u91cf\u3002\u4f8b\u5982\uff0c\u8003\u8651\u8bcd\u8bed \u201chot\u201d \u4e0e \u201csummer\u201d \u7684\u76f8\u4f3c\u5ea6\uff0c\u5219\u53ef\u4ee5\u201c\u7b80\u5316\u201d\u7684\u5904\u7406\u8fd9\u4e24\u4e2a\u8bcd\uff08Token\uff09\u7684\u5411\u91cf\u7684\u201c\u5185\u79ef\u201d\u3002<\/p>\n\n\n\n \u5728\u524d\u9762\u7684\u6587\u7ae0\u201c\u5927\u8bed\u8a00\u6a21\u578b\u7684\u8f93\u5165\uff1atokenize<\/a>\u201d\u4e2d\uff0c\u8f83\u4e3a\u8be6\u7ec6\u7684\u4ecb\u7ecd\u4e86\u5927\u8bed\u8a00\u6a21\u578b\u5982\u4f55\u628a\u4e00\u4e2a\u53e5\u5b50\u8f6c\u6362\u6210\u4e00\u4e2a\u7684 Token\uff0c\u7136\u540e\u518d\u8f6c\u6362\u4e3a\u4e00\u4e2a\u4e2a\u201c\u5411\u91cf\u201d\u3002\u90a3\u4e48\uff0c\u6211\u4eec\u901a\u5e38\u4f1a\u901a\u8fc7\u4e24\u4e2a\u5411\u91cf\u7684\u4f59\u5f26\u76f8\u4f3c\u5ea6\u6765\u63cf\u8ff0\u5176\u76f8\u4f3c\u5ea6\uff0c\u5982\u679c\u5411\u91cf\u7684\u201c\u957f\u5ea6\u201d\uff08\\(L_2 \\) \u8303\u6570\uff09\u662f\u5355\u4f4d\u957f\u5ea6\uff0c\u90a3\u4e48\u4e5f\u901a\u5e38\u4f1a\u76f4\u63a5\u4f7f\u7528\u201c\u5185\u79ef\u201d\u63cf\u8ff0\u4e24\u4e2a\u5411\u91cf\u7684\u76f8\u4f3c\u5ea6\uff1a<\/p>\n\n\n\n $$ \\(f(x) = \\cos(x) \\) \u7684\u56fe\u50cf\u5982\u53f3\u56fe\uff0c\u6545\uff1a<\/p>\n\n\n\n \u4f8b\u5982\uff0c<\/p>\n\n\n\n \u56e0\u4e3a\u4e24\u4e2a\u5411\u91cf\u7684\u201c\u5185\u79ef\u201d\u67d0\u79cd\u7a0b\u5ea6\u53ef\u4ee5\u8868\u793a\u4e3a\u76f8\u4f3c\u5ea6\u3002\u90a3\u4e48\uff0c\u5bf9\u4e8e\u53e5\u5b50\u4e2d\u7684\u67d0\u4e2a token \u6765\u8bf4\uff0c\u4e0e\u5176\u4ed6\u6240\u6709\u5411\u91cf\u5404\u81ea\u8ba1\u7b97\u201c\u5185\u79ef\u201d\uff0c\u5c31\u53ef\u4ee5\u83b7\u5f97\u4e00\u4e2a\u4e0e\u5176\u4ed6\u6240\u6709\u5411\u91cf\u201c\u76f8\u4f3c\u7a0b\u5ea6\u201d\u7684\u6570\u7ec4\uff0c\u518d\u5bf9\u8fd9\u4e2a\u6570\u7ec4\u8fdb\u884c softmax \u8ba1\u7b97\u5c31\u53ef\u4ee5\u83b7\u5f97\u4e00\u4e2a\u8be5 token \u4e0e\u5176\u4ed6\u6240\u6709\u5411\u91cf\u201c\u76f8\u4f3c\u7a0b\u5ea6\u201d\u7684\u5f52\u4e00\u5316\u6570\u7ec4\u3002\u8fd9\u4e2a\u5f52\u4e00\u5316\u7684\u6570\u636e\uff0c\u5c31\u53ef\u4ee5\u7406\u89e3\u4e3a\u8fd9\u91cc\u7684\u201cSimilarity Score Matrix\u201d\u3002<\/p>\n\n\n\n \u8fd9\u91cc\u4f9d\u65e7\u4ee5\u201c\u5927\u8bed\u8a00\u6a21\u578b\u7684\u8f93\u5165\uff1atokenize<\/a>\u201d\u793a\u4f8b\u4e2d\u7684\u53e5\u5b50\u4e3a\u6f14\u793a\u793a\u4f8b\u3002<\/p>\n\n\n\n \u66f4\u4e3a\u5177\u4f53\u7684\uff0c\u53ef\u4ee5\u53c2\u8003\u53f3\u56fe\u3002\u8fd9\u91cc\u8003\u8651 Token \u201cIt\u201d \u4e0e\u5176\u4ed6\u6240\u6709\u8bcd\u8bed\u7684\u76f8\u4f3c\u5ea6\u3002\u5373\u8ba1\u7b97 Token \u201cIt\u201d \u7684 Embedding \u5411\u91cf\uff0c\u4e0e\u5176\u4ed6\u6240\u6709\u5411\u91cf\u7684\u201c\u5185\u79ef\u201d\u3002<\/p>\n\n\n\n \u66f4\u8fdb\u4e00\u6b65\uff0c\u5982\u679c\u8ba1\u7b97\u4e24\u4e24\u8bcd\u8bed\u4e4b\u95f4\u7684\u76f8\u4f3c\u5ea6\uff0c\u5e76\u8fdb\u884c\u5f52\u4e00\u5316\uff08softmax \uff09\uff0c\u5219\u6709\u5982\u4e0b\u7684Similarity Matrix\uff1a<\/p>\n\n\n\n \u5728\u8fd9\u4e2a\u793a\u4f8b\u4e2d\uff0c\u5219\u4f1a\u6709\u4e0a\u8ff0 12×12 \u7684\u77e9\u9635\u3002\u8be5\u77e9\u9635\u53cd\u5e94\u4e86\u201c\u8bcd\u201d\u4e0e\u201c\u8bcd\u201d\u4e4b\u95f4\u7684\u76f8\u4f3c\u5ea6\u3002\u5982\u679c\uff0c\u6211\u4eec\u628a\u6bcf\u4e00\u884c\u518d\u8fdb\u884c\u4e00\u4e2a\u201c\u5f52\u4e00\u5316\u201d\uff08\u6ce8\u53f3\u56fe\u5df2\u7ecf\u7ecf\u8fc7\u4e86\u5f52\u4e00\u5316\uff09\uff0c\u90a3\u4e48\u6bcf\u4e00\u884c\uff0c\u5c31\u53cd\u5e94\u4e86\u4e00\u4e2a\u8bcd\u8bed\u4e0e\u5176\u4ed6\u6240\u6709\u8bcd\u8bed\u76f8\u4f3c\u7a0b\u5ea6\u7684\u4e00\u4e2a\u5ea6\u91cf\u3002<\/p>\n\n\n\n \u4f8b\u5982\uff0c\u53f3\u56fe\u4e2d it \u53ef\u80fd\u4e0e very \u6700\u4e3a\u76f8\u4f3c\uff08\u9664\u4e86\u81ea\u8eab\uff09\u3002<\/p>\n<\/div>\n<\/div>\n\n\n\n \u6ce8\u610f\u5230\u6700\u7ec8\u7684 \u201cAttention\u201d \u8ba1\u7b97\u516c\u5f0f\u548c\u4e0a\u8ff0\u7684\u201cSimilarity Score Matrix\u201d\u7684\u5dee\u522b\u5c31\u662f\u53c2\u6570\u77e9\u9635W\uff1a<\/p>\n\n\n\n $$ \\text{Similarity Score} = \\text{softmax}(\\frac{XX^{T}}{\\sqrt{d}})X $$<\/p>\n\n\n\n \u8fd9\u91cc\u6ca1\u6709\u53c2\u6570\u77e9\u9635\uff1a\\(W^Q \\quad W^K \\quad W^V \\)<\/p>\n<\/div>\n\n\n\n $$ \\text{Attention} = \\text{softmax}(\\frac{QK^{T}}{\\sqrt{d}})V $$<\/p>\n\n\n\n \u5176\u4e2d\uff0c \\(Q = XW^Q \\quad K = XW^K \\quad V = XW^V \\)<\/p>\n<\/div>\n<\/div>\n\n\n\n \u90a3\u4e48\uff0c\u4e3a\u4ec0\u4e48\u9700\u8981 \\(W^Q \\,, W^K \\,, W^V \\) \u5462\uff1f\u8fd9\u4e09\u4e2a\u53c2\u6570\u77e9\u9635\u4e58\u6cd5\uff0c\u610f\u5473\u7740\u4ec0\u4e48\u5462\uff1f\u8981\u8bf4\u6e05\u695a\u3001\u8981\u7406\u89e3\u8fd9\u4e2a\u70b9\u5e76\u4e0d\u5bb9\u6613\uff0c\u4e5f\u6ca1\u6709\u4ec0\u4e48\u7b80\u5355\u7684\u63cf\u8ff0\u53ef\u4ee5\u8bf4\u6e05\u695a\u7684\uff0c\u8fd9\u4e5f\u5927\u6982\u662f\u4e3a\u4ec0\u4e48\uff0c\u5bf9\u4e8e\u975e NLP \u4e13\u4e1a\u7684\u4eba\uff0c\u8981\u60f3\u771f\u6b63\u7406\u89e3 Transformer \u6216 Attention \u662f\u6bd4\u8f83\u56f0\u96be\u7684\u3002<\/p>\n\n\n\n \u4f60\u53ef\u80fd\u4f1a\u770b\u5230\u8fc7\u4e00\u79cd\u6bd4\u8f83\u666e\u904d\u7684\u3001\u7b80\u5316\u7248\u672c\uff0c\u5927\u6982\u662f\u8bf4 \\(W^Q \\) \u662f\u4e00\u4e2a Query \u77e9\u9635\uff0c\u8868\u793a\u8981\u67e5\u8be2\u4ec0\u4e48\uff1b\\(W^K \\) \u662f\u4e00\u4e2a Key \u77e9\u9635\uff0c\u8868\u793a\u4e00\u4e2a\u8bcd\u6709\u4ec0\u4e48\u3002\u8fd9\u4e2a\u8bf4\u6cd5\u4f3c\u4e4e\u5e76\u4e0d\u80fd\u589e\u52a0\u5bf9\u4e0a\u8ff0\u516c\u5f0f\u7684\u7406\u89e3\u3002<\/p>\n\n\n\n \u90a3\u4e48\uff0c\u4e00\u4e2a\u5411\u91cf\u4e58\u4ee5\u4e00\u4e2a\u77e9\u9635\u65f6\uff0c\u8fd9\u4e2a\u201c\u77e9\u9635\u201d\u610f\u5473\u7740\u4ec0\u4e48\uff1f\u662f\u7684\uff0c\u5c31\u662f\u201c\u7ebf\u6027\u53d8\u6362\u201d\u3002<\/p>\n\n\n\n \u4e00\u822c\u6765\u8bf4\uff0c\\(W^Q \\,, W^K \\,, W^V \\) \u662f\u4e00\u4e2a \\(d \\times d \\) \u7684\u77e9\u9635[1]<\/sup>\u3002\u5bf9\u4e8e\u7684 Token Embedding \uff08\u4e0a\u8ff0\u7684\u77e9\u9635 X \uff09\u6240\u5728\u7684\u5411\u91cf\u7a7a\u95f4\uff0c\u90a3\u4e48 \\(W^Q \\,, W^K \\,, W^V \\) \u5c31\u662f\u8be5\u5411\u91cf\u7a7a\u95f4\u7684\u4e09\u4e2a\u201c\u7ebf\u6027\u53d8\u6362\u201d\u3002<\/p>\n\n\n\n \u90a3\u4e48\u7ebf\u6027\u53d8\u6362\u5bf9\u4e8e\u5411\u91cf\u7a7a\u95f4\u7684\u4f5c\u7528\u662f\u4ec0\u4e48\u5462\uff1f\u8fd9\u91cc\u6211\u4eec\u4ee5\u201c\u5947\u5f02\u503c\u5206\u89e3\u201d\u7684\u89d2\u5ea6\u6765\u7406\u89e3\u8fd9\u4e2a\u95ee\u9898[2]<\/sup>\uff0c\u5373\u5bf9\u5411\u91cf\u8fdb\u884c\u62c9\u4f38\/\u538b\u7f29\u3001\u65cb\u8f6c\u3001\u955c\u50cf\u53d8\u6362\u3002\\(W^Q \\,, W^K \\,, W^V \\) \u5219\u4f1a\u5206\u522b\u5bf9\u5411\u91cf\u7a7a\u95f4\u7684\u5411\u91cf\uff08\u5373Token Embedding\uff09\u505a\u7c7b\u4f3c\u7684\u53d8\u6362\u3002\u53d8\u6362\u7684\u7ed3\u679c\u5373\u4e3a\uff1a<\/p>\n\n\n $$ Q = XW^Q \\quad \\quad K = XW^K \\quad \\quad V = XW^V $$<\/p>\n\n\n\n \u90a3\u4e48\uff0c\u5982\u679c\u53c2\u6570\u77e9\u9635\u201c\u8bbe\u8ba1\u201d\u5408\u7406\uff0c\u201cToken\u201d\u4e0e\u201cToken\u201d\u4e4b\u95f4\u5c31\u53ef\u4ee5\u5efa\u7acb\u201c\u671f\u671b\u201d\u7684 Attention \u5173\u7cfb\uff0c\u4f8b\u5982\uff1a\u201c\u4ee3\u8bcd\u201d\uff08it\uff09\uff0c\u603b\u662f\u66f4\u591a\u7684\u5173\u6ce8\u4e8e\u201c\u540d\u8bcd\u201d\uff1b\u201c\u540d\u8bcd\u201d\u66f4\u591a\u7684\u5173\u6ce8\u4e0e\u9644\u8fd1\u7684\u201c\u5f62\u5bb9\u8bcd\u201d\uff1b\u518d\u6bd4\u5982\uff0c\u201c\u52a8\u8bcd\u201d\u66f4\u591a\u5173\u6ce8\u524d\u540e\u7684\u201c\u540d\u8bcd\u201d\u7b49\u3002\u9664\u4e86\u8bcd\u6027\uff0c\u7ebf\u6027\u53d8\u6362\u5173\u6ce8\u7684\u201c\u7ef4\u5ea6\u201d\u53ef\u80fd\u6709\u5f88\u591a\uff0c\u4f8b\u5982\u201c\u4f4d\u7f6e\u201d\u3001\u201c\u60c5\u611f\u201d\u3001\u201c\u52a8\u7269\u201d\u3001\u201c\u690d\u7269\u201d\u3001\u201c\u79ef\u6781\/\u6d88\u6781\u201d\u7b49\u3002\u5173\u4e8e\u5982\u4f55\u7406\u89e3 token embedding \u7684\u5404\u4e2a\u201c\u7ef4\u5ea6\u201d\u542b\u4e49\u53ef\u4ee5\u53c2\u8003\uff1aWord Embedding \u7684\u53ef\u89e3\u91ca\u6027\u63a2\u7d22<\/a>\u3002<\/p>\n\n\n\n \u5f53\u7136\uff0c\u8fd9\u4e09\u4e2a\u53c2\u6570\u77e9\u9635\u90fd\u4e0d\u662f\u201c\u8bbe\u8ba1\u201d\u51fa\u6765\u7684\uff0c\u800c\u662f\u201c\u8bad\u7ec3\u201d\u51fa\u6765\u7684\u3002\u6240\u4ee5\uff0c\u8981\u60f3\u5bfb\u627e\u4e0a\u8ff0\u5982\u6b64\u6e05\u6670\u7684\u201c\u53ef\u89e3\u91ca\u6027\u201d\u5e76\u4e0d\u5bb9\u6613\u30022019\u5e74\u7684\u8bba\u6587\u300aWhat Does BERT Look At? An Analysis of BERT’s Attention<\/a>\u300b\u8f83\u4e3a\u7cfb\u7edf\u7684\u8ba8\u8bba\u4e86\u8fd9\u4e2a\u95ee\u9898\uff0c\u611f\u5174\u8da3\u7684\u53ef\u4ee5\u53bb\u770b\u770b\u3002<\/p>\n\n\n\n|-------------------------------------------------------------------------------------------------------------------------------------------
| Token<\/mark> | wte[:3] (word token embedding)<\/mark> | wpe[:3](word positional e..)<\/mark>| wte [:3] + wpe [:3]<\/mark> | X[:3] (Norm)<\/mark> |
|-------------------------------------------------------------------------------------------------------------------------------------------
|It | [0.039, -0.0869, 0.0662 ,...] | [-0.0188, -0.1974, 0.004 ] | [0.0202, -0.2844,0.0702 ] | [ 0.0129, -0.1104, -0.0317,...] |
|\u00e2\u0122 | [-0.075, 0.0948, -0.0034,...] | [0.024, -0.0538, -0.0949] | [-0.051, 0.041, -0.0982] | [-0.0530, 0.0588, -0.1290,...] |
|\u013b | [-0.0223, 0.0182, 0.2631 ,...] | [0.0042, -0.0848, 0.0545 ] | [-0.0181,-0.0666,0.3176 ] | [-0.0170, -0.0242, 0.1639,...] |
|s | [-0.064, -0.0469, 0.2061 ,...] | [-0.0003, -0.0738, 0.1055 ] | [-0.0643,-0.1207,0.3116 ] | [-0.0754, -0.0842, 0.1842,...] |
|\u0120very | [-0.0553, -0.0348, 0.0606 ,...] | [0.0076, -0.0251, 0.127 ] | [-0.0477,-0.0599,0.1876 ] | [-0.0566, -0.0280, 0.0953,...] |
|\u0120hot | [0.0399, -0.0053, 0.0742 ,...] | [0.0096, -0.0339, 0.1312 ] | [0.0495, -0.0392,0.2054 ] | [ 0.0587, -0.0086, 0.1073,...] |
|\u0120in | [-0.0337, 0.0108, 0.0293 ,...] | [0.0027, -0.0205, 0.1196 ] | [-0.031, -0.0098,0.149 ] | [-0.0391, 0.0209, 0.0731,...] |
|\u0120summer | [0.0422, 0.0138, -0.0213,...] | [0.0025, -0.0032, 0.1174 ] | [0.0448, 0.0106, 0.0961 ] | [ 0.0532, 0.0397, 0.0181,...] |
|. | [0.0466, -0.0113, 0.0283 ,...] | [-0.0012, -0.0018, 0.111 ] | [0.0454, -0.0131,0.1394 ] | [ 0.0553, 0.0152, 0.0579,...] |
|\u0120Sw | [0.0617, 0.0373, 0.1018 ,...] | [0.0049, 0.0021, 0.1178 ] | [0.0666, 0.0395, 0.2196 ] | [ 0.0807, 0.0691, 0.1216,...] |
|imming | [-0.1385, -0.1774, -0.0181,...] | [0.0016, 0.0062, 0.1004 ] | [-0.1369,-0.1711,0.0823 ] | [-0.1528, -0.1249, -0.0017,...] |
|\u0120is | [-0.0097, 0.0101, 0.0556 ,...] | [-0.0036, 0.0175, 0.1068 ] | [-0.0133,0.0275, 0.1623 ] | [-0.0175, 0.0605, 0.0880,...] |
|-------------------------------------------------------------------------------------------------------------------------------------------<\/pre>\n\n\n\n --------------------------------------- ------------------------------------------ -------------
| [ 0.0129, -0.1104, -0.0317,...] | | x_1<\/mark> = [ 0.0129, -0.1104, -0.0317,...] | |It |
| [-0.0530, 0.0588, -0.1290,...] | | x_2<\/mark> = [-0.0530, 0.0588, -0.1290,...] | |\u00e2\u0122 |
| [-0.0170, -0.0242, 0.1639,...] | | x_3<\/mark> = [-0.0170, -0.0242, 0.1639,...] | |\u013b |
| [-0.0754, -0.0842, 0.1842,...] | | x_4<\/mark> = [-0.0754, -0.0842, 0.1842,...] | |s |
| [-0.0566, -0.0280, 0.0953,...] | | x_5<\/mark> = [-0.0566, -0.0280, 0.0953,...] | |\u0120very |
| X<\/mark> = [ 0.0587, -0.0086, 0.1073,...] | | x_6<\/mark> = [ 0.0587, -0.0086, 0.1073,...] | |\u0120hot |
| [-0.0391, 0.0209, 0.0731,...] | | x_7<\/mark> = [-0.0391, 0.0209, 0.0731,...] | |\u0120in |
| [ 0.0532, 0.0397, 0.0181,...] | | x_8<\/mark> = [ 0.0532, 0.0397, 0.0181,...] | |\u0120summer |
| [ 0.0553, 0.0152, 0.0579,...] | | x_9<\/mark> = [ 0.0553, 0.0152, 0.0579,...] | |. |
| [ 0.0807, 0.0691, 0.1216,...] | | x_10<\/mark> = [ 0.0807, 0.0691, 0.1216,...] | |\u0120Sw |
| [-0.1528, -0.1249, -0.0017,...] | | x_11<\/mark> = [-0.1528, -0.1249, -0.0017,...] | |imming |
| [-0.0175, 0.0605, 0.0880,...] | | x_12<\/mark> = [-0.0175, 0.0605, 0.0880,...] | |\u0120is |
--------------------------------------- ------------------------------------------ -------------<\/pre>\n\n\n\n3. Similarity Score Matrix<\/h3>\n\n\n\n
3.1 \u4e24\u4e2a\u201c\u8bcd\u8bed\u201d\u7684\u76f8\u4f3c\u5ea6<\/h4>\n\n\n\n
\n\\cos \\theta = \\frac{\\alpha \\cdot \\beta}{\\|\\alpha\\| \\|\\beta\\| }
\n$$<\/p>\n\n\n\n\n
![\"\"](\"https:\/\/www.orczhou.com\/wp-content\/uploads\/2025\/11\/image-8-1024x613.png\")
<\/figure>\n<\/div>\n<\/div>\n\n\n\n3.2 \u201cSimilarity Score Matrix\u201d<\/h4>\n\n\n\n
![\"\"](\"https:\/\/www.orczhou.com\/wp-content\/uploads\/2025\/11\/image-9-1024x830.png\")
<\/figure>\n\n\n\n[0.6975, 0.0347, 0.0236, 0.0298, 0.0386, 0.0282, 0.0272, 0.0270, 0.0216, 0.0244, 0.0219, 0.0254]
[0.0000, 0.9994, 0.0002, 0.0000, 0.0000, 0.0001, 0.0000, 0.0000, 0.0000, 0.0001, 0.0001, 0.0000]
[0.0000, 0.0004, 0.9987, 0.0001, 0.0001, 0.0001, 0.0000, 0.0002, 0.0000, 0.0001, 0.0002, 0.0000]
[0.0002, 0.0003, 0.0004, 0.9945, 0.0008, 0.0004, 0.0011, 0.0003, 0.0007, 0.0005, 0.0002, 0.0007]
[0.0013, 0.0012, 0.0013, 0.0042, 0.9724, 0.0044, 0.0047, 0.0021, 0.0030, 0.0013, 0.0012, 0.0028]
[0.0002, 0.0003, 0.0003, 0.0004, 0.0009, 0.9932, 0.0005, 0.0020, 0.0004, 0.0009, 0.0006, 0.0003]
[0.0049, 0.0048, 0.0038, 0.0299, 0.0258, 0.0125, 0.7728, 0.0086, 0.0779, 0.0095, 0.0025, 0.0471]
[0.0001, 0.0001, 0.0005, 0.0002, 0.0003, 0.0015, 0.0002, 0.9959, 0.0002, 0.0003, 0.0003, 0.0002]
[0.0049, 0.0048, 0.0045, 0.0255, 0.0203, 0.0126, 0.0974, 0.0082, 0.7698, 0.0094, 0.0024, 0.0401]
[0.0001, 0.0001, 0.0001, 0.0002, 0.0001, 0.0003, 0.0001, 0.0002, 0.0001, 0.9983, 0.0002, 0.0002]
[0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.9998, 0.0000]
[0.0033, 0.0027, 0.0028, 0.0129, 0.0110, 0.0063, 0.0334, 0.0055, 0.0228, 0.0075, 0.0025, 0.8893]<\/pre>\n\n\n\n4. Self-Attention<\/h3>\n\n\n\n
4.1 \u5bf9\u6bd4<\/h4>\n\n\n\n
4.2 \u4e3a\u4ec0\u4e48\u9700\u8981\u53c2\u6570\u77e9\u9635 W<\/h4>\n\n\n\n
4.3 \u7ebf\u6027\u53d8\u6362 Linear Transformations<\/h4>\n\n\n\n
![\"\"](\"https:\/\/www.orczhou.com\/wp-content\/uploads\/2025\/12\/image-1024x389.png\")
<\/figure>\n\n\n\n![\"\"](\"https:\/\/www.orczhou.com\/wp-content\/uploads\/2025\/11\/image-11.png\")
<\/figure>\n<\/div>\n\n\n\n![\"\"](\"https:\/\/www.orczhou.com\/wp-content\/uploads\/2025\/12\/self-attention-1.svg\")