\u5728\u5f53\u4eca\u5927\u6570\u636e\u65f6\u4ee3，\u60c5\u611f\u5206\u6790\u5df2\u6210\u4e3a\u81ea\u7136\u8bed\u8a00\u5904\u7406（NLP）\u9886\u57df\u4e2d\u6700\u91cd\u8981\u7684\u5e94\u7528\u4e4b\u4e00\u3002\u65e0\u8bba\u662f\u7535\u5546\u5e73\u53f0\u7684\u4ea7\u54c1\u8bc4\u8bba\u5206\u6790，\u8fd8\u662f\u793e\u4ea4\u5a92\u4f53\u7684\u8206\u60c5\u76d1\u63a7，\u51c6\u786e\u8bc6\u522b\u7528\u6237\u60c5\u611f\u503e\u5411\u90fd\u5177\u6709\u5de8\u5927\u5546\u4e1a\u4ef7\u503c\u3002\u7136\u800c，\u4f20\u7edf\u7684\u673a\u5668\u5b66\u4e60\u65b9\u6cd5\u5728\u5904\u7406\u5e8f\u5217\u6570\u636e\u65f6\u9762\u4e34\u7740\u5de8\u5927\u6311\u6218，\u7279\u522b\u662f\u5bf9\u4e8e\u5177\u6709\u4e0a\u4e0b\u6587\u4f9d\u8d56\u7684\u6587\u672c\u6570\u636e\u3002<\/p>\n
\u5faa\u73af\u795e\u7ecf\u7f51\u7edc（RNN）\u4f5c\u4e3a\u4e00\u79cd\u4e13\u95e8\u5904\u7406\u5e8f\u5217\u6570\u636e\u7684\u795e\u7ecf\u7f51\u7edc\u67b6\u6784，\u5728\u60c5\u611f\u5206\u6790\u4efb\u52a1\u4e2d\u5c55\u73b0\u51fa\u72ec\u7279\u4f18\u52bf\u3002\u4e0e\u4f20\u7edf\u65b9\u6cd5\u4e0d\u540c，RNN\u80fd\u591f\u6355\u6349\u6587\u672c\u4e2d\u7684\u65f6\u5e8f\u4fe1\u606f\u548c\u957f\u671f\u4f9d\u8d56\u5173\u7cfb，\u8fd9\u5bf9\u4e8e\u7406\u89e3\u81ea\u7136\u8bed\u8a00\u7684\u8bed\u4e49\u81f3\u5173\u91cd\u8981\u3002\u4f8b\u5982，\u5728\u4e2d\u6587\u60c5\u611f\u5206\u6790\u4e2d，"\u8fd9\u4e2a\u4ea7\u54c1\u867d\u7136\u8d35\u4f46\u662f\u8d28\u91cf\u5f88\u597d"\u8fd9\u6837\u7684\u53e5\u5b50，\u53ea\u6709\u7406\u89e3\u6574\u4e2a\u53e5\u5b50\u4e0a\u4e0b\u6587\u624d\u80fd\u51c6\u786e\u5224\u65ad\u5176\u79ef\u6781\u60c5\u611f\u503e\u5411\u3002<\/span><\/p>\n \u672c\u6587\u5c06\u4ece\u96f6\u5f00\u59cb，\u8be6\u7ec6\u8bb2\u89e3\u5982\u4f55\u5b9e\u73b0\u4e00\u4e2a\u5b8c\u6574\u7684RNN\u6a21\u578b，\u5e76\u5c06\u5176\u5e94\u7528\u4e8e\u4e2d\u6587\u60c5\u611f\u5206\u6790\u4efb\u52a1\u3002\u901a\u8fc7\u8fd9\u4e2a\u5b9e\u8df5\u9879\u76ee，\u60a8\u4e0d\u4ec5\u5c06\u6df1\u5165\u7406\u89e3RNN\u7684\u5de5\u4f5c\u539f\u7406，\u8fd8\u80fd\u638c\u63e1\u4ece\u6570\u636e\u5904\u7406\u5230\u6a21\u578b\u8bad\u7ec3\u7684\u5b8c\u6574\u6d41\u7a0b\u3002<\/p>\n \u6784\u5efaRNN\u6a21\u578b\u9700\u8981\u591a\u4e2aPython\u5e93\u7684\u652f\u6301，\u6bcf\u4e2a\u5e93\u90fd\u6709\u5176\u7279\u5b9a\u4f5c\u7528：<\/p>\n # \u5bfc\u5165\u6570\u503c\u8ba1\u7b97\u5e93 NumPy：\u63d0\u4f9b\u9ad8\u6548\u7684\u6570\u7ec4\u64cd\u4f5c，\u662f\u795e\u7ecf\u7f51\u7edc\u8ba1\u7b97\u7684\u57fa\u7840<\/span><\/p>\n<\/li>\n Pandas：\u7b80\u5316\u6570\u636e\u5904\u7406\u6d41\u7a0b，\u65b9\u4fbf\u6570\u636e\u63a2\u7d22\u548c\u5206\u6790<\/span><\/p>\n<\/li>\n Jieba：\u51c6\u786e\u9ad8\u6548\u7684\u4e2d\u6587\u5206\u8bcd\u5de5\u5177<\/span><\/p>\n<\/li>\n scikit-learn：\u63d0\u4f9b\u6570\u636e\u96c6\u5212\u5206\u548c\u8bc4\u4f30\u529f\u80fd<\/span><\/p>\n<\/li>\n warnings：\u63a7\u5236\u8b66\u544a\u4fe1\u606f\u8f93\u51fa，\u4fdd\u6301\u4ee3\u7801\u6574\u6d01<\/span><\/p>\n<\/li>\n<\/ul>\n 1.\u00a0RNN\u7684\u6838\u5fc3\u601d\u60f3<\/span><\/p>\n \u5faa\u73af\u795e\u7ecf\u7f51\u7edc\u901a\u8fc7\u5f15\u5165\u5faa\u73af\u8fde\u63a5，\u4f7f\u7f51\u7edc\u80fd\u591f"\u8bb0\u4f4f"\u4e4b\u524d\u7684\u4fe1\u606f\u3002\u8fd9\u79cd\u8bb0\u5fc6\u80fd\u529b\u4f7fRNN\u7279\u522b\u9002\u5408\u5904\u7406\u5e8f\u5217\u6570\u636e，\u5982\u6587\u672c\u3001\u8bed\u97f3\u3001\u65f6\u95f4\u5e8f\u5217\u7b49\u3002<\/p>\n 2.\u00a0RNN\u7684\u57fa\u672c\u7ed3\u6784<\/span><\/p>\n RNN\u5728\u65f6\u95f4\u6b65t\u7684\u8ba1\u7b97\u516c\u5f0f\u5982\u4e0b：<\/p>\n \u9690\u85cf\u72b6\u6001：h_t = tanh(W_xh\u00b7x_t + W_hh\u00b7h_{t-1} + b_h)<\/span><\/p>\n<\/li>\n \u8f93\u51fa：y_t = softmax(W_hy\u00b7h_t + b_y)<\/span><\/p>\n<\/li>\n<\/ul>\n \u5176\u4e2d：<\/p>\n x_t：\u65f6\u95f4\u6b65t\u7684\u8f93\u5165<\/span><\/p>\n<\/li>\n h_t：\u65f6\u95f4\u6b65t\u7684\u9690\u85cf\u72b6\u6001<\/span><\/p>\n<\/li>\n y_t：\u65f6\u95f4\u6b65t\u7684\u8f93\u51fa<\/span><\/p>\n<\/li>\n W_xh, W_hh, W_hy：\u6743\u91cd\u77e9\u9635<\/span><\/p>\n<\/li>\n b_h, b_y：\u504f\u7f6e\u9879<\/span><\/p>\n<\/li>\n<\/ul>\n RNN\u6a21\u578b\u7684\u521d\u59cb\u5316\u6d89\u53ca\u591a\u4e2a\u91cd\u8981\u53c2\u6570\u7684\u8bbe\u7f6e：<\/p>\n class SimpleRNN: # \u521d\u59cb\u5316\u6743\u91cd\u77e9\u9635 # \u521d\u59cb\u5316\u504f\u7f6e\u9879 \u6743\u91cd\u521d\u59cb\u5316\u91c7\u7528\u5c0f\u968f\u673a\u6570\u7b56\u7565（\u4e58\u4ee50.01），\u8fd9\u79cd\u7b56\u7565\u6709\u52a9\u4e8e\u907f\u514d\u68af\u5ea6\u7206\u70b8\u6216\u6d88\u5931\u95ee\u9898\u3002\u504f\u7f6e\u9879\u521d\u59cb\u5316\u4e3a\u96f6，\u8fd9\u662f\u795e\u7ecf\u7f51\u7edc\u4e2d\u7684\u5e38\u89c1\u505a\u6cd5\u3002<\/p>\n \u6fc0\u6d3b\u51fd\u6570\u662f\u795e\u7ecf\u7f51\u7edc\u7684\u975e\u7ebf\u6027\u6765\u6e90，\u4f7f\u7f51\u7edc\u80fd\u591f\u5b66\u4e60\u590d\u6742\u7684\u6a21\u5f0f：<\/p>\n """Tanh\u6fc0\u6d3b\u51fd\u6570""" """Softmax\u6fc0\u6d3b\u51fd\u6570""" \u6fc0\u6d3b\u51fd\u6570\u9009\u62e9\u4f9d\u636e：<\/p>\n Tanh：\u8f93\u51fa\u8303\u56f4\u5728[-1, 1]，\u9002\u5408\u4f5c\u4e3a\u9690\u85cf\u5c42\u6fc0\u6d3b\u51fd\u6570<\/span><\/p>\n<\/li>\n Softmax：\u5c06\u8f93\u51fa\u8f6c\u6362\u4e3a\u6982\u7387\u5206\u5e03，\u9002\u5408\u591a\u5206\u7c7b\u95ee\u9898<\/span><\/p>\n<\/li>\n<\/ul>\n \u524d\u5411\u4f20\u64ad\u662fRNN\u7684\u6838\u5fc3，\u8d1f\u8d23\u8ba1\u7b97\u7f51\u7edc\u8f93\u51fa：<\/p>\n """\u524d\u5411\u4f20\u64ad""" for t in range(seq_len): # \u8ba1\u7b97\u8f93\u51fa\u5c42\u72b6\u6001 return self.y<\/p>\n \u65f6\u95f4\u6b65\u8ba1\u7b97\u6d41\u7a0b：<\/span><\/p>\n \u8f93\u5165\u51c6\u5907：\u63a5\u6536\u4e09\u7ef4\u5f20\u91cf（\u5e8f\u5217\u957f\u5ea6\u00d7\u6279\u6b21\u5927\u5c0f\u00d7\u7279\u5f81\u7ef4\u5ea6）<\/span><\/p>\n<\/li>\n \u72b6\u6001\u521d\u59cb\u5316：\u9690\u85cf\u72b6\u6001\u521d\u59cb\u5316\u4e3a\u96f6\u5411\u91cf<\/span><\/p>\n<\/li>\n \u5faa\u73af\u8ba1\u7b97：\u6309\u65f6\u95f4\u6b65\u987a\u5e8f\u8ba1\u7b97\u6bcf\u4e2a\u65f6\u95f4\u6b65\u7684\u9690\u85cf\u72b6\u6001\u548c\u8f93\u51fa<\/span><\/p>\n<\/li>\n<\/ul>\n \u53cd\u5411\u4f20\u64ad\u901a\u8fc7\u65f6\u95f4（BPTT）\u662fRNN\u8bad\u7ec3\u7684\u5173\u952e：<\/p>\n """\u53cd\u5411\u4f20\u64ad""" # \u521d\u59cb\u5316\u68af\u5ea6 # \u4ece\u6700\u540e\u4e00\u4e2a\u65f6\u95f4\u6b65\u5f00\u59cb\u53cd\u5411\u4f20\u64ad # \u8ba1\u7b97\u9690\u85cf\u5c42\u5230\u8f93\u51fa\u5c42\u7684\u6743\u91cd\u77e9\u9635\u68af\u5ea6 # \u8ba1\u7b97\u8f93\u51fa\u5c42\u504f\u7f6e\u9879\u68af\u5ea6 # \u8ba1\u7b97\u5f53\u524d\u8f93\u51fa\u8bef\u5dee\u5f15\u8d77\u7684\u9690\u85cf\u5c42\u68af\u5ea6 # \u8ba1\u7b97\u4e0b\u4e00\u65f6\u95f4\u6b65\u7684\u9690\u85cf\u5c42\u68af\u5ea6 # \u5408\u5e76\u68af\u5ea6\u5e76\u5e94\u7528tanh\u7684\u5bfc\u6570 # \u8ba1\u7b97\u6743\u91cd\u68af\u5ea6 # \u8ba1\u7b97\u504f\u7f6e\u68af\u5ea6 # \u8ba1\u7b97\u8f93\u5165\u68af\u5ea6 return self.dx<\/p>\n BPTT\u7b97\u6cd5\u8981\u70b9：<\/p>\n \u65f6\u95f4\u53cd\u5411\u4f20\u64ad：\u4ece\u6700\u540e\u4e00\u4e2a\u65f6\u95f4\u6b65\u5f00\u59cb，\u5411\u524d\u4f20\u64ad\u68af\u5ea6<\/span><\/p>\n<\/li>\n \u68af\u5ea6\u94fe\u5f0f\u6cd5\u5219：\u901a\u8fc7\u94fe\u5f0f\u6cd5\u5219\u8ba1\u7b97\u6bcf\u4e2a\u53c2\u6570\u7684\u68af\u5ea6<\/span><\/p>\n<\/li>\n \u68af\u5ea6\u7d2f\u79ef：RNN\u4e2d\u540c\u4e00\u53c2\u6570\u5728\u4e0d\u540c\u65f6\u95f4\u6b65\u5171\u4eab，\u68af\u5ea6\u9700\u8981\u7d2f\u52a0<\/span><\/p>\n<\/li>\n<\/ul>\n """\u6743\u91cd\u66f4\u65b0""" """\u6a21\u578b\u8bad\u7ec3""" # \u8ba1\u7b97\u635f\u5931 def create_dataset(): # \u8f6c\u6362\u4e3aDataFrame \u6570\u636e\u96c6\u8bbe\u8ba1\u539f\u5219：<\/span><\/p>\n \u7c7b\u522b\u5e73\u8861：\u786e\u4fdd\u6b63\u9762\u548c\u8d1f\u9762\u6837\u672c\u6570\u91cf\u63a5\u8fd1<\/span><\/p>\n<\/li>\n \u591a\u6837\u6027：\u8986\u76d6\u4e0d\u540c\u4e3b\u9898\u548c\u8868\u8fbe\u65b9\u5f0f<\/span><\/p>\n<\/li>\n \u771f\u5b9e\u6027：\u6a21\u62df\u771f\u5b9e\u573a\u666f\u4e2d\u7684\u8bc4\u8bba\u5185\u5bb9<\/span><\/p>\n<\/li>\n<\/ul>\n \u9884\u5904\u7406\u6b65\u9aa4\u76f4\u63a5\u5f71\u54cd\u6a21\u578b\u6027\u80fd：<\/p>\n # \u6587\u672c\u9884\u5904\u7406 \u4e2d\u6587\u5206\u8bcd\u7684\u91cd\u8981\u6027，\u4e2d\u6587\u6ca1\u6709\u81ea\u7136\u7684\u5206\u9694\u7b26，\u5206\u8bcd\u662f\u5c06\u8fde\u7eed\u5b57\u7b26\u8f6c\u6362\u4e3a\u6709\u610f\u4e49\u7684\u8bcd\u8bed\u7684\u5173\u952e\u6b65\u9aa4\u3002\u51c6\u786e\u7684\u5206\u8bcd\u6709\u52a9\u4e8e\u6a21\u578b\u7406\u89e3\u8bed\u4e49\u3002<\/p>\n \u5c06\u6587\u672c\u8f6c\u6362\u4e3a\u6570\u503c\u8868\u793a\u662fNLP\u4efb\u52a1\u7684\u6838\u5fc3：<\/p>\n # \u5c06\u6587\u672c\u8f6c\u6362\u4e3a\u5e8f\u5217 # \u586b\u5145\u5e8f\u5217\u5230\u56fa\u5b9a\u957f\u5ea6(\u4e0d\u8db3\u7684\u90e8\u5206\u75280\u586b\u5145) return sequence<\/p>\n One-hot\u7f16\u7801\u539f\u7406：One-hot\u7f16\u7801\u5c06\u6bcf\u4e2a\u8bcd\u8868\u793a\u4e3a\u957f\u5ea6\u4e3a\u8bcd\u6c47\u8868\u5927\u5c0f\u7684\u5411\u91cf，\u5176\u4e2d\u53ea\u6709\u4e00\u4e2a\u4f4d\u7f6e\u4e3a1，\u5176\u4f59\u4e3a0\u3002\u867d\u7136\u7b80\u5355\u76f4\u89c2，\u4f46\u5728\u8bcd\u6c47\u8868\u8f83\u5927\u65f6\u4f1a\u4ea7\u751f\u9ad8\u7ef4\u7a00\u758f\u95ee\u9898\u3002<\/p>\n \u4e3b\u51fd\u6570\u6574\u5408\u4e86\u4ece\u6570\u636e\u51c6\u5907\u5230\u6a21\u578b\u8bc4\u4f30\u7684\u5b8c\u6574\u6d41\u7a0b： def main(): # 2.\u6587\u672c\u9884\u5904\u7406 # 3.\u6784\u5efa\u8bcd\u6c47\u8868 vocab_size = len(vocab) # 4.\u8f6c\u6362\u4e3a\u5e8f\u5217 # \u8f6c\u6362\u4e3a\u72ec\u70ed\u7f16\u7801 # \u6807\u7b7e\u7f16\u7801 # 5.\u6570\u636e\u96c6\u5206\u5272 # 6.\u8bad\u7ec3RNN\u6a21\u578b model = SimpleRNN(vocab_size, hidden_size, output_size, learning_rate=0.01) # 7.\u6a21\u578b\u8bc4\u4f30 # 8.\u9884\u6d4b\u65b0\u6587\u672c for text in new_texts: x = np.zeros((max_len, 1, vocab_size)) result = model.predict(x)[0] 1.\u00a0\u8bad\u7ec3\u8fc7\u7a0b\u89c2\u5bdf<\/span><\/p>\n \u5728\u8bad\u7ec3\u8fc7\u7a0b\u4e2d，\u635f\u5931\u51fd\u6570\u5e94\u5448\u73b0\u4e0b\u964d\u8d8b\u52bf\u3002\u5982\u679c\u635f\u5931\u6ce2\u52a8\u8f83\u5927\u6216\u4e0b\u964d\u7f13\u6162，\u53ef\u80fd\u9700\u8981\u8c03\u6574\u5b66\u4e60\u7387\u6216\u68c0\u67e5\u6570\u636e\u9884\u5904\u7406\u8fc7\u7a0b\u3002<\/p>\n 2.\u00a0\u51c6\u786e\u7387\u8bc4\u4f30<\/span><\/p>\n \u6a21\u578b\u5728\u6d4b\u8bd5\u96c6\u4e0a\u7684\u51c6\u786e\u7387\u53cd\u6620\u4e86\u5176\u6cdb\u5316\u80fd\u529b\u3002\u5bf9\u4e8e\u4e8c\u5206\u7c7b\u95ee\u9898，\u51c6\u786e\u7387\u8d85\u8fc7\u968f\u673a\u731c\u6d4b（50%）\u5373\u8868\u793a\u6a21\u578b\u5b66\u5230\u4e86\u4e00\u4e9b\u6a21\u5f0f\u3002<\/p>\n \u5f53\u524d\u7684SimpleRNN\u5b9e\u73b0\u5b58\u5728\u4e00\u4e9b\u5c40\u9650\u6027：<\/p>\n \u68af\u5ea6\u6d88\u5931\/\u7206\u70b8：\u5728\u5904\u7406\u957f\u5e8f\u5217\u65f6\u5bb9\u6613\u51fa\u73b0<\/span><\/p>\n<\/li>\n \u8ba1\u7b97\u6548\u7387：One-hot\u7f16\u7801\u5bfc\u81f4\u9ad8\u7ef4\u7a00\u758f\u95ee\u9898<\/span><\/p>\n<\/li>\n \u6a21\u578b\u5bb9\u91cf：\u7b80\u5355RNN\u96be\u4ee5\u6355\u6349\u590d\u6742\u6a21\u5f0f<\/span><\/p>\n<\/li>\n<\/ul>\n 1.\u00a0\u4f7f\u7528\u9ad8\u7ea7RNN\u53d8\u4f53<\/span><\/p>\n LSTM：\u901a\u8fc7\u95e8\u63a7\u673a\u5236\u89e3\u51b3\u957f\u4f9d\u8d56\u95ee\u9898<\/span><\/p>\n<\/li>\n GRU：\u7b80\u5316\u7248LSTM，\u8ba1\u7b97\u6548\u7387\u66f4\u9ad8<\/span><\/p>\n<\/li>\n \u53cc\u5411RNN：\u540c\u65f6\u8003\u8651\u524d\u540e\u6587\u4fe1\u606f<\/span><\/p>\n<\/li>\n<\/ul>\n 2.\u00a0\u00a0\u8bcd\u5d4c\u5165\u6280\u672f<\/span><\/p>\n \u7528\u8bcd\u5411\u91cf（Word2Vec, GloVe, FastText）\u4ee3\u66ffOne-hot\u7f16\u7801：<\/p>\n \u964d\u4f4e\u7ef4\u5ea6<\/span><\/p>\n<\/li>\n \u6355\u6349\u8bed\u4e49\u76f8\u4f3c\u6027<\/span><\/p>\n<\/li>\n \u5904\u7406\u672a\u767b\u5f55\u8bcd<\/span><\/p>\n<\/li>\n<\/ul>\n 3.\u00a0\u6ce8\u610f\u529b\u673a\u5236<\/span><\/p>\n \u5f15\u5165\u6ce8\u610f\u529b\u673a\u5236\u4f7f\u6a21\u578b\u80fd\u591f\u5173\u6ce8\u8f93\u5165\u5e8f\u5217\u4e2d\u7684\u91cd\u8981\u90e8\u5206：<\/p>\n \u63d0\u9ad8\u6a21\u578b\u89e3\u91ca\u6027<\/span><\/p>\n<\/li>\n \u6539\u5584\u957f\u5e8f\u5217\u5904\u7406\u80fd\u529b<\/span><\/p>\n<\/li>\n<\/ul>\n 4.\u00a0\u6570\u636e\u589e\u5f3a<\/span><\/p>\n \u540c\u4e49\u8bcd\u66ff\u6362<\/span><\/p>\n<\/li>\n \u56de\u8bd1\u6280\u672f<\/span><\/p>\n<\/li>\n \u968f\u673a\u63d2\u5165\/\u5220\u9664<\/span><\/p>\n<\/li>\n<\/ul>\n \u6027\u80fd\u4f18\u5316：\u8003\u8651\u4f7f\u7528\u6279\u91cf\u63a8\u7406\u548c\u6a21\u578b\u91cf\u5316<\/span><\/p>\n<\/li>\n \u5b9e\u65f6\u6027\u8981\u6c42：\u4f18\u5316\u9884\u6d4b\u5ef6\u8fdf，\u652f\u6301\u9ad8\u5e76\u53d1<\/span><\/p>\n<\/li>\n \u6a21\u578b\u66f4\u65b0：\u5efa\u7acb\u5b9a\u671f\u66f4\u65b0\u673a\u5236，\u9002\u5e94\u8bed\u8a00\u53d8\u5316<\/span><\/p>\n<\/li>\n<\/ul>\n \u5f53\u5c06\u6a21\u578b\u5e94\u7528\u4e8e\u7279\u5b9a\u9886\u57df\u65f6：<\/p>\n \u9886\u57df\u8bcd\u5178：\u6784\u5efa\u9886\u57df\u4e13\u4e1a\u8bcd\u6c47\u8868<\/span><\/p>\n<\/li>\n \u8fc1\u79fb\u5b66\u4e60：\u4f7f\u7528\u9884\u8bad\u7ec3\u6a21\u578b\u8fdb\u884c\u5fae\u8c03<\/span><\/p>\n<\/li>\n \u9886\u57df\u6570\u636e：\u6536\u96c6\u9886\u57df\u7279\u5b9a\u8bad\u7ec3\u6570\u636e<\/span><\/p>\n<\/li>\n<\/ul>\n \u4ee5\u4e0b\u662f\u5b8c\u6574\u7684RNN\u4e2d\u6587\u60c5\u611f\u5206\u6790\u5b9e\u73b0\u4ee3\u7801：<\/p>\n # \u5173\u952e\u70b9\u8bf4\u660e： # \u8fd9\u4e2a\u4ee3\u7801\u4e3b\u8981\u76ee\u7684\u662f\u5c55\u793aRNN\u7684\u57fa\u672c\u539f\u7406\u548c\u5de5\u4f5c\u6d41\u7a0b，\u5b9e\u9645\u5e94\u7528\u4e2d\u53ef\u80fd\u9700\u8981： # \u5bfc\u5165\u6570\u503c\u8ba1\u7b97\u5e93 # \u5b9e\u73b0\u7b80\u5355\u7684RNN\u6a21\u578b """\u7b80\u5355RNN\u6a21\u578b\u521d\u59cb\u5316""" # \u521d\u59cb\u5316\u8f93\u5165\u5c42\u5230\u9690\u85cf\u5c42\u7684\u6743\u91cd\u77e9\u9635 # \u521d\u59cb\u5316\u9690\u85cf\u5c42\u504f\u7f6e\u9879 """Tanh\u6fc0\u6d3b\u51fd\u6570""" """Softmax\u6fc0\u6d3b\u51fd\u6570""" """\u524d\u5411\u4f20\u64ad""" for t in range(seq_len): # \u8ba1\u7b97\u8f93\u51fa\u5c42\u72b6\u6001 return self.y<\/p>\n """\u53cd\u5411\u4f20\u64ad""" # \u4ece\u6700\u540e\u4e00\u4e2a\u65f6\u95f4\u6b65\u5f00\u59cb\u53cd\u5411\u4f20\u64ad（BPTT\u6838\u5fc3） # \u8ba1\u7b97\u9690\u85cf\u5c42\u5230\u8f93\u51fa\u5c42\u7684\u6743\u91cd\u77e9\u9635\u68af\u5ea6 # \u8ba1\u7b97\u8f93\u51fa\u5c42\u504f\u7f6e\u9879\u68af\u5ea6 # \u8ba1\u7b97\u5f53\u524d\u8f93\u51fa\u8bef\u5dee\u5f15\u8d77\u7684\u9690\u85cf\u5c42\u68af\u5ea6 # \u8ba1\u7b97\u4e0b\u4e00\u65f6\u95f4\u6b65\u7684\u9690\u85cf\u5c42\u68af\u5ea6\u901a\u8fc7W_hh\u4f20\u9012\u7684\u90e8\u5206 # \u5408\u5e76\u68af\u5ea6\u5e76\u5e94\u7528tanh\u7684\u5bfc\u6570 # \u8ba1\u7b97\u9690\u85cf\u5c42\u5230\u9690\u85cf\u5c42\u7684\u6743\u91cd\u77e9\u9635\u68af\u5ea6 # \u8ba1\u7b97\u8f93\u5165\u5c42\u5230\u9690\u85cf\u5c42\u7684\u6743\u91cd\u77e9\u9635\u68af\u5ea6 # \u8ba1\u7b97\u9690\u85cf\u5c42\u504f\u7f6e\u9879\u68af\u5ea6 # \u8ba1\u7b97\u8f93\u5165\u5c42\u72b6\u6001\u68af\u5ea6 return self.dx<\/p>\n """\u6743\u91cd\u66f4\u65b0""" # \u91cd\u7f6e\u68af\u5ea6（\u53ef\u9009，\u907f\u514d\u68af\u5ea6\u7d2f\u79ef） """\u6a21\u578b\u8bad\u7ec3""" # \u8ba1\u7b97\u635f\u5931 """\u9884\u6d4b\u6a21\u578b""" # \u6570\u636e\u5904\u7406\u51fd\u6570 # \u8f6c\u6362\u4e3aDataFrame # \u6587\u672c\u9884\u5904\u7406 # \u5c06\u6587\u672c\u8f6c\u6362\u4e3a\u5e8f\u5217 # \u586b\u5145\u5e8f\u5217\u5230\u56fa\u5b9a\u957f\u5ea6(\u4e0d\u8db3\u7684\u90e8\u5206\u75280\u586b\u5145) return sequence<\/p>\n # \u4e3b\u51fd\u6570 # 2.\u6587\u672c\u9884\u5904\u7406 # 3.\u6784\u5efa\u8bcd\u6c47\u8868 # \u521b\u5efa\u8bcd\u6c47\u8868，<PAD>\u8868\u793a\u586b\u5145\u8bcd vocab_size = len(vocab) # 4.\u8f6c\u6362\u4e3a\u5e8f\u5217 for text in df['processed']: # \u904d\u5386\u6240\u6709\u6587\u672c # \u8f6c\u6362\u4e3a\u72ec\u70ed\u7f16\u7801 # \u5c06\u6807\u7b7e\u8f6c\u6362\u4e3a\u72ec\u70ed\u7f16\u7801\u683c\u5f0f # 5.\u6570\u636e\u96c6\u5206\u5272 # 6.\u8bad\u7ec3RNN\u6a21\u578b model = SimpleRNN(vocab_size, hidden_size, output_size, learning_rate=0.01) # 7.\u6a21\u578b\u8bc4\u4f30 # 8.\u9884\u6d4b\u65b0\u6587\u672c new_texts = [ for text in new_texts: # \u8f6c\u6362\u4e3a\u72ec\u70ed\u7f16\u7801 # \u9884\u6d4b if __name__ == "__main__": \u901a\u8fc7\u672c\u6587\u7684\u5b8c\u6574\u5b9e\u8df5，\u6211\u4eec\u4ece\u96f6\u5f00\u59cb\u5b9e\u73b0\u4e86\u4e00\u4e2a\u57fa\u4e8eRNN\u7684\u4e2d\u6587\u60c5\u611f\u5206\u6790\u7cfb\u7edf\u3002\u8fd9\u4e2a\u7cfb\u7edf\u867d\u7136\u7b80\u5355，\u4f46\u6db5\u76d6\u4e86RNN\u7684\u6838\u5fc3\u6982\u5ff5\u3001\u6570\u636e\u9884\u5904\u7406\u3001\u6a21\u578b\u8bad\u7ec3\u548c\u8bc4\u4f30\u7684\u5b8c\u6574\u6d41\u7a0b\u3002<\/p>\n \u672a\u6765\u53d1\u5c55\u65b9\u5411\u5305\u62ec：<\/p>\n \u6a21\u578b\u4f18\u5316：\u4f7f\u7528LSTM\u3001GRU\u7b49\u9ad8\u7ea7RNN\u53d8\u4f53<\/span><\/p>\n<\/li>\n \u8bcd\u5d4c\u5165：\u91c7\u7528Word2Vec\u3001BERT\u7b49\u9884\u8bad\u7ec3\u8bcd\u5411\u91cf<\/span><\/p>\n<\/li>\n \u591a\u4efb\u52a1\u5b66\u4e60：\u540c\u65f6\u8fdb\u884c\u60c5\u611f\u5206\u6790\u548c\u4e3b\u9898\u5206\u7c7b<\/span><\/p>\n<\/li>\n \u5b9e\u65f6\u7cfb\u7edf：\u6784\u5efa\u4f4e\u5ef6\u8fdf\u7684\u5728\u7ebf\u60c5\u611f\u5206\u6790\u670d\u52a1<\/span><\/p>\n<\/li>\n<\/ul>\n RNN\u4f5c\u4e3a\u5e8f\u5217\u5efa\u6a21\u7684\u57fa\u7840\u67b6\u6784，\u867d\u7136\u5728\u67d0\u4e9b\u4efb\u52a1\u4e0a\u5df2\u88abTransformer\u7b49\u65b0\u67b6\u6784\u8d85\u8d8a，\u4f46\u5176\u6838\u5fc3\u601d\u60f3\u4ecd\u7136\u6df1\u523b\u5f71\u54cd\u4e86\u73b0\u4ee3\u6df1\u5ea6\u5b66\u4e60\u7684\u53d1\u5c55\u3002\u7406\u89e3RNN\u7684\u5de5\u4f5c\u539f\u7406，\u5bf9\u4e8e\u638c\u63e1\u66f4\u590d\u6742\u7684\u5e8f\u5217\u6a21\u578b\u5177\u6709\u91cd\u8981\u610f\u4e49\u3002<\/p>\n","protected":false},"excerpt":{"rendered":" 1. \u5f15\u8a00：\u5faa\u73af\u795e\u7ecf\u7f51\u7edc\u5728\u60c5\u611f\u5206\u6790\u4e2d\u7684\u5e94\u7528\u4ef7\u503c\u5728\u5f53\u4eca\u5927\u6570\u636e\u65f6\u4ee3，\u60c5\u611f\u5206\u6790\u5df2\u6210\u4e3a\u81ea\u7136\u8bed\u8a00\u5904\u7406（NLP）\u9886\u57df\u4e2d\u6700\u91cd\u8981\u7684\u5e94\u7528\u4e4b\u4e00\u3002\u65e0\u8bba\u662f\u7535\u5546\u5e73\u53f0\u7684\u4ea7\u54c1\u8bc4\u8bba\u5206\u6790，\u8fd8\u662f\u793e\u4ea4\u5a92\u4f53\u7684\u8206\u60c5\u76d1\u63a7，\u51c6\u786e\u8bc6\u522b\u7528\u6237\u60c5\u611f\u503e\u5411\u90fd\u5177\u6709\u5de8\u5927\u5546\u4e1a\u4ef7\u503c\u3002\u7136\u800c，\u4f20\u7edf\u7684\u673a\u5668\u5b66\u4e60\u65b9\u6cd5\u5728\u5904\u7406\u5e8f\u5217\u6570\u636e\u65f6\u9762\u4e34\u7740\u5de8\u5927\u6311\u6218，\u7279\u522b\u662f\u5bf9\u4e8e\u5177\u6709\u4e0a\u4e0b\u6587\u4f9d\u8d56\u7684\u6587\u672c\u6570\u636e\u3002\u5faa\u73af\u795e\u7ecf\u7f51\u7edc（RNN）\u4f5c\u4e3a\u4e00\u79cd<\/p>\n","protected":false},"author":2,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[2068,81,50],"topic":[],"class_list":["post-65874","post","type-post","status-publish","format-standard","hentry","category-server","tag-nlp","tag-python","tag-50"],"yoast_head":"\n2. \u73af\u5883\u51c6\u5907\u4e0e\u7406\u8bba\u57fa\u7840<\/h2>\n
2.1. \u6838\u5fc3\u5e93\u4ecb\u7ecd<\/h3>\n
\nimport numpy as np
\n# \u5bfc\u5165\u6570\u636e\u5904\u7406\u5e93
\nimport pandas as pd
\n# \u5bfc\u5165\u4e2d\u6587\u5206\u8bcd\u5e93
\nimport jieba as jb
\n# \u5bfc\u5165\u6570\u636e\u96c6\u5212\u5206\u5de5\u5177
\nfrom sklearn.model_selection import train_test_split
\n# \u5bfc\u5165\u6a21\u578b\u8bc4\u4f30\u5de5\u5177，\u7528\u4e8e\u8ba1\u7b97\u6a21\u578b\u7684\u51c6\u786e\u7387
\nfrom sklearn.metrics import accuracy_score
\n# \u5bfc\u5165\u8b66\u544a\u8fc7\u6ee4\u5de5\u5177
\nimport warnings
\n# \u5ffd\u7565\u8b66\u544a\u4fe1\u606f
\nwarnings.filterwarnings("ignore")<\/p>\n\n
2.2. RNN\u7406\u8bba\u57fa\u7840<\/h3>\n
\n
\n
3. SimpleRNN\u7c7b\u7684\u8bbe\u8ba1\u4e0e\u5b9e\u73b0<\/h2>\n
3.1. \u7c7b\u521d\u59cb\u5316<\/h3>\n
\n """\u7b80\u5355RNN\u6a21\u578b\u521d\u59cb\u5316"""
\n def __init__(self, input_size, hidden_size, output_size, learning_rate=0.01):
\n # \u53c2\u6570\u8bf4\u660e：
\n # input_size: \u8f93\u5165\u7279\u5f81\u5927\u5c0f
\n # hidden_size: \u9690\u85cf\u5c42\u5927\u5c0f
\n # output_size: \u8f93\u51fa\u5c42\u5927\u5c0f
\n # learning_rate: \u5b66\u4e60\u7387，\u9ed8\u8ba4\u503c\u4e3a0.01
\n self.input_size = input_size
\n self.hidden_size = hidden_size
\n self.output_size = output_size
\n self.learning_rate = learning_rate<\/p>\n
\n self.W_xh = np.random.randn(input_size, hidden_size) * 0.01
\n self.W_hh = np.random.randn(hidden_size, hidden_size) * 0.01
\n self.W_hy = np.random.randn(hidden_size, output_size) * 0.01<\/p>\n
\n self.b_h = np.zeros((1, hidden_size))
\n self.b_y = np.zeros((1, output_size))<\/p>\n3.2. \u6fc0\u6d3b\u51fd\u6570\u5b9e\u73b0<\/h3>\n
\ndef tanh(self, x):
\n return np.tanh(x) # \u9690\u85cf\u5c42\u6fc0\u6d3b\u51fd\u6570<\/p>\n
\ndef softmax(self, x):
\n exp_x = np.exp(x – np.max(x, axis=1, keepdims=True)) # \u9632\u6b62\u6ea2\u51fa
\n return exp_x \/ np.sum(exp_x, axis=1, keepdims=True) # \u8f93\u51fa\u5c42\u6fc0\u6d3b\u51fd\u6570<\/p>\n\n
3.3. \u524d\u5411\u4f20\u64ad\u5b9e\u73b0<\/h3>\n
\ndef forward(self, X):
\n # \u83b7\u53d6\u8f93\u5165\u5e8f\u5217\u957f\u5ea6\u3001\u6279\u6b21\u5927\u5c0f
\n seq_len, batch_size, _ = X.shape
\n # \u521d\u59cb\u5316\u9690\u85cf\u5c42\u72b6\u6001
\n self.h = np.zeros((seq_len + 1, batch_size, self.hidden_size))
\n # \u521d\u59cb\u5316\u8f93\u51fa\u5c42\u72b6\u6001
\n self.y = np.zeros((seq_len, batch_size, self.output_size))<\/p>\n
\n # \u8ba1\u7b97\u9690\u85cf\u5c42\u72b6\u6001
\n # h_t = tanh(W_xh * x_t + W_hh * h_t-1 + b_h)
\n self.h[t + 1] = self.tanh(np.dot(X[t], self.W_xh) +
\n np.dot(self.h[t], self.W_hh) + self.b_h)<\/p>\n
\n # y_t = softmax(W_hy * h_t + b_y)
\n self.y[t] = self.softmax(np.dot(self.h[t + 1], self.W_hy) + self.b_y)<\/p>\n\n
3.4. \u53cd\u5411\u4f20\u64ad\u5b9e\u73b0<\/h3>\n
\ndef backward(self, X, y):
\n seq_len, batch_size, _ = X.shape
\n self.dh = np.zeros((seq_len + 2, batch_size, self.hidden_size))
\n self.dy = np.zeros((seq_len, batch_size, self.output_size))
\n self.dx = np.zeros((seq_len, batch_size, self.input_size))<\/p>\n
\n self.dW_hy = np.zeros_like(self.W_hy)
\n self.dW_hh = np.zeros_like(self.W_hh)
\n self.dW_xh = np.zeros_like(self.W_xh)
\n self.db_h = np.zeros_like(self.b_h)
\n self.db_y = np.zeros_like(self.b_y)<\/p>\n
\n for t in reversed(range(seq_len)):
\n # \u8ba1\u7b97\u8f93\u51fa\u5c42\u72b6\u6001\u68af\u5ea6
\n self.dy[t] = self.y[t] – y[t]<\/p>\n
\n self.dW_hy += np.dot(self.h[t + 1].T, self.dy[t])<\/p>\n
\n self.db_y += np.sum(self.dy[t], axis=0, keepdims=True)<\/p>\n
\n dh_from_output = np.dot(self.dy[t], self.W_hy.T)<\/p>\n
\n dh_from_next = np.dot(self.dh[t + 2], self.W_hh.T) if (t + 1) < seq_len else 0<\/p>\n
\n self.dh[t + 1] = (dh_from_output + dh_from_next) * (1 – self.h[t + 1] ** 2)<\/p>\n
\n self.dW_hh += np.dot(self.h[t].T, self.dh[t + 1])
\n self.dW_xh += np.dot(X[t].T, self.dh[t + 1])<\/p>\n
\n self.db_h += np.sum(self.dh[t + 1], axis=0, keepdims=True)<\/p>\n
\n self.dx[t] = np.dot(self.dh[t + 1], self.W_xh.T)<\/p>\n\n
3.5. \u6743\u91cd\u66f4\u65b0\u4e0e\u8bad\u7ec3<\/h3>\n
\ndef update_weights(self):
\n # \u4f7f\u7528\u68af\u5ea6\u4e0b\u964d\u6cd5\u66f4\u65b0\u6240\u6709\u6743\u91cd\u548c\u504f\u7f6e
\n self.W_xh -= self.learning_rate * self.dW_xh
\n self.W_hh -= self.learning_rate * self.dW_hh
\n self.W_hy -= self.learning_rate * self.dW_hy
\n self.b_h -= self.learning_rate * self.db_h
\n self.b_y -= self.learning_rate * self.db_y<\/p>\n
\ndef train(self, X, y, epochs=100):
\n # \u8bad\u7ec3\u6a21\u578b
\n for epoch in range(epochs):
\n # \u524d\u5411\u4f20\u64ad
\n y_pred = self.forward(X)
\n # \u53cd\u5411\u4f20\u64ad
\n self.backward(X, y)
\n # \u66f4\u65b0\u6743\u91cd
\n self.update_weights()<\/p>\n
\n loss = -np.mean(np.sum(y * np.log(y_pred + 1e-8), axis=1))
\n # \u6253\u5370\u635f\u5931
\n if epoch % 10 == 0:
\n print(f"Epoch {epoch}, Loss: {loss:.4f}")<\/p>\n4. \u6570\u636e\u5904\u7406\u4e0e\u9884\u5904\u7406<\/h2>\n
4.1. \u6570\u636e\u96c6\u521b\u5efa<\/h3>\n
\n """\u521b\u5efa\u7b80\u5355\u7684\u60c5\u611f\u5206\u6790\u6570\u636e\u96c6"""
\n # \u6269\u5c55\u8bad\u7ec3\u6570\u636e\u96c6
\n data = {
\n 'text': [
\n '\u8fd9\u4e2a\u4ea7\u54c1\u5f88\u597d', '\u8d28\u91cf\u592a\u5dee\u4e86', '\u670d\u52a1\u4e0d\u9519', '\u4ef7\u683c\u592a\u8d35',
\n '\u7269\u6d41\u5f88\u5feb', '\u5305\u88c5\u7cbe\u7f8e', '\u5ba2\u670d\u8010\u5fc3', '\u5546\u54c1\u6709\u7455\u75b5',
\n '\u6027\u4ef7\u6bd4\u9ad8', '\u4f7f\u7528\u4e0d\u65b9\u4fbf', '\u989c\u8272\u5f88\u597d\u770b', '\u5c3a\u5bf8\u4e0d\u5408\u9002',
\n '\u6750\u8d28\u4e0d\u9519', '\u5473\u9053\u4e0d\u597d\u95fb', '\u4ef7\u683c\u5408\u7406', '\u7269\u6d41\u592a\u6162',
\n '\u5305\u88c5\u7834\u635f', '\u5ba2\u670d\u6001\u5ea6\u597d', '\u8fd9\u4e2a\u4e1c\u897f\u5f88\u597d\u7528', '\u8d28\u91cf\u975e\u5e38\u5dee',
\n '\u670d\u52a1\u6001\u5ea6\u5f88\u597d', '\u4ef7\u683c\u771f\u7684\u8d35', '\u7269\u6d41\u901f\u5ea6\u5f88\u5feb', '\u5305\u88c5\u5f88\u7528\u5fc3',
\n '\u5ba2\u670d\u975e\u5e38\u8010\u5fc3', '\u5546\u54c1\u8d28\u91cf\u6709\u95ee\u9898', '\u6027\u4ef7\u6bd4\u975e\u5e38\u9ad8', '\u4f7f\u7528\u8d77\u6765\u4e0d\u65b9\u4fbf'
\n ],
\n 'label': [1, 0, 1, 0, 1, 1, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 0, 1, 1, 0, 1, 0, 1, 1, 1, 0, 1, 0]
\n }<\/p>\n
\n df = pd.DataFrame(data)
\n return df<\/p>\n\n
4.2. \u6587\u672c\u9884\u5904\u7406<\/h3>\n
\ndef preprocess_text(text):
\n # \u5206\u8bcd
\n seg_list = jb.cut(text)
\n # \u5408\u5e76\u5206\u8bcd\u7ed3\u679c
\n return " ".join(seg_list)<\/p>\n4.3. \u5e8f\u5217\u5316\u4e0e\u7f16\u7801<\/h3>\n
\ndef text_to_sequence(text, vocab, max_len=5):
\n words = text.split() # \u6309\u7a7a\u683c\u5206\u8bcd
\n sequence = [] # \u521d\u59cb\u5316\u5e8f\u5217
\n for word in words[:max_len]: # \u53ea\u53d6\u524dmax_len\u4e2a\u5355\u8bcd
\n if word in vocab:
\n sequence.append(vocab[word]) # \u83b7\u53d6\u8bcd\u5bf9\u5e94\u7684ID
\n else:
\n sequence.append(0) # \u672a\u77e5\u8bcd\u75280\u8868\u793a<\/p>\n
\n while len(sequence) < max_len:
\n sequence.append(0)<\/p>\n5. \u5b9e\u9a8c\u6d41\u7a0b\u4e0e\u7ed3\u679c\u5206\u6790<\/h2>\n
5.1. \u5b8c\u6574\u5b9e\u9a8c\u6d41\u7a0b<\/h3>\n
\n\u00a0<\/p>\n
\n # 1.\u6570\u636e\u51c6\u5907
\n print("===\u6570\u636e\u51c6\u5907===")
\n df = create_dataset()
\n print(f"\u6570\u636e\u96c6\u5927\u5c0f：{len(df)}")<\/p>\n
\n print("\\\\n===\u6587\u672c\u9884\u5904\u7406===")
\n df['processed'] = df['text'].apply(preprocess_text)<\/p>\n
\n print("\\\\n===\u6784\u5efa\u8bcd\u6c47\u8868===")
\n vocab = {'<PAD>': 0}
\n for text in df['processed']:
\n for word in text.split():
\n if word not in vocab:
\n vocab[word] = len(vocab)<\/p>\n
\n print(f"\u8bcd\u6c47\u8868\u5927\u5c0f：{vocab_size}")<\/p>\n
\n print("\\\\n===\u5e8f\u5217\u8f6c\u6362===")
\n max_len = 5
\n sequences = []
\n for text in df['processed']:
\n seq = text_to_sequence(text, vocab, max_len)
\n sequences.append(seq)<\/p>\n
\n X = np.zeros((len(sequences), max_len, vocab_size))
\n for i, seq in enumerate(sequences):
\n for j, word_index in enumerate(seq):
\n X[i, j, word_index] = 1<\/p>\n
\n y = df['label'].values
\n y_onehot = np.zeros((len(y), 2))
\n y_onehot[np.arange(len(y)), y] = 1<\/p>\n
\n X_train, X_test, y_train, y_test = train_test_split(
\n X, y_onehot, test_size=0.25, random_state=42
\n )
\n # \u8c03\u6574\u7ef4\u5ea6\u4e3a(seq_len, batch_size, input_size)
\n X_train = X_train.transpose(1, 0, 2)
\n X_test = X_test.transpose(1, 0, 2)<\/p>\n
\n print("\\\\n===\u8bad\u7ec3RNN\u6a21\u578b===")
\n hidden_size = 32
\n output_size = 2<\/p>\n
\n model.train(X_train, y_train, epochs=200)<\/p>\n
\n print("\\\\n===\u6a21\u578b\u8bc4\u4f30===")
\n y_pred = model.predict(X_test)
\n y_test_labels = np.argmax(y_test, axis=1)
\n accuracy = accuracy_score(y_test_labels, y_pred)
\n print(f"\u6d4b\u8bd5\u96c6\u51c6\u786e\u7387：{accuracy:.4f}")<\/p>\n
\n print("\\\\n===\u9884\u6d4b\u65b0\u6587\u672c===")
\n new_texts = [
\n '\u8fd9\u4e2a\u5546\u54c1\u5f88\u6709\u4ef7\u503c',
\n '\u670d\u52a1\u7cdf\u7cd5',
\n '\u7269\u6d41\u5f88\u53ca\u65f6',
\n '\u5305\u88c5\u5f88\u7cbe\u7f8e',
\n '\u5ba2\u670d\u5f88\u8010\u5fc3'
\n ]<\/p>\n
\n processed = preprocess_text(text)
\n seq = text_to_sequence(processed, vocab, max_len)<\/p>\n
\n for j, word_id in enumerate(seq):
\n x[j, 0, word_id] = 1<\/p>\n
\n sentiment = "\u6b63\u9762" if result == 1 else "\u8d1f\u9762"
\n print(f"\u6587\u672c: {text}")
\n print(f"\u9884\u6d4b: {sentiment}")
\n print()<\/p>\n5.2. \u5b9e\u9a8c\u7ed3\u679c\u5206\u6790<\/h3>\n
6. \u6a21\u578b\u4f18\u5316\u4e0e\u6539\u8fdb\u65b9\u5411<\/h2>\n
6.1. \u5f53\u524d\u5b9e\u73b0\u5c40\u9650\u6027<\/h3>\n
\n
6.2. \u6539\u8fdb\u65b9\u6848<\/h3>\n
\n
\n
\n
\n
7. \u5b9e\u9645\u5e94\u7528\u5efa\u8bae<\/h2>\n
7.1. \u90e8\u7f72\u6ce8\u610f\u4e8b\u9879<\/h3>\n
\n
7.2. \u9886\u57df\u9002\u5e94<\/h3>\n
\n
8. \u5b8c\u6574\u4ee3\u7801\u793a\u4f8b<\/h2>\n
\n# \u8fd9\u662f\u4e00\u4e2a\u7b80\u5355\u7684RNN\u5b9e\u73b0，\u7528\u4e8e\u4e2d\u6587\u6587\u672c\u60c5\u611f\u5206\u6790（\u6b63\u9762\/\u8d1f\u9762）
\n# \u6570\u636e\u9884\u5904\u7406\u5305\u62ec：\u4e2d\u6587\u5206\u8bcd\u3001\u6784\u5efa\u8bcd\u6c47\u8868\u3001\u5e8f\u5217\u5316\u548cone-hot\u7f16\u7801
\n# RNN\u7684\u6838\u5fc3\u662f\u524d\u5411\u4f20\u64ad\u90e8\u5206，\u8ba1\u7b97\u9690\u85cf\u72b6\u6001\u548c\u8f93\u51fa
\n# \u6700\u540e\u5c55\u793a\u4e86\u5982\u4f55\u7528\u8bad\u7ec3\u597d\u7684\u6a21\u578b\u9884\u6d4b\u65b0\u6587\u672c\u7684\u60c5\u611f\u503e\u5411<\/p>\n
\n# \u5b9e\u73b0\u5b8c\u6574\u7684\u8bad\u7ec3\u8fc7\u7a0b
\n# \u4f7f\u7528\u66f4\u590d\u6742\u7684RNN\u53d8\u4f53（\u5982LSTM\u3001GRU）
\n# \u4f7f\u7528\u8bcd\u5d4c\u5165\u4ee3\u66ffone-hot\u7f16\u7801
\n# \u589e\u52a0\u66f4\u591a\u7684\u8bad\u7ec3\u6570\u636e<\/p>\n
\nimport numpy as np
\n# \u5bfc\u5165\u6570\u636e\u5904\u7406\u5e93
\nimport pandas as pd
\n# \u5bfc\u5165\u4e2d\u6587\u5206\u8bcd\u5e93
\nimport jieba as jb
\n# \u5bfc\u5165\u6570\u636e\u96c6\u5212\u5206\u5de5\u5177
\nfrom sklearn.model_selection import train_test_split
\n# \u5bfc\u5165\u6a21\u578b\u8bc4\u4f30\u5de5\u5177，\u7528\u4e8e\u8ba1\u7b97\u6a21\u578b\u7684\u51c6\u786e\u7387
\nfrom sklearn.metrics import accuracy_score
\n# \u5bfc\u5165\u8b66\u544a\u8fc7\u6ee4\u5de5\u5177
\nimport warnings
\n# \u5ffd\u7565\u8b66\u544a\u4fe1\u606f
\nwarnings.filterwarnings("ignore")<\/p>\n
\nclass SimpleRNN:<\/p>\n
\n def __init__(self, input_size, hidden_size, output_size, learning_rate=0.01):
\n # \u53c2\u6570\u8bf4\u660e：
\n # input_size: \u8f93\u5165\u7279\u5f81\u5927\u5c0f
\n # hidden_size: \u9690\u85cf\u5c42\u5927\u5c0f
\n # output_size: \u8f93\u51fa\u5c42\u5927\u5c0f
\n # learning_rate: \u5b66\u4e60\u7387，\u9ed8\u8ba4\u503c\u4e3a0.01
\n self.input_size = input_size
\n self.hidden_size = hidden_size
\n self.output_size = output_size
\n self.learning_rate = learning_rate<\/p>\n
\n self.W_xh = np.random.randn(input_size, hidden_size) * 0.01
\n # \u521d\u59cb\u5316\u9690\u85cf\u5c42\u5230\u9690\u85cf\u5c42\u7684\u6743\u91cd\u77e9\u9635
\n self.W_hh = np.random.randn(hidden_size, hidden_size) * 0.01
\n # \u521d\u59cb\u5316\u9690\u85cf\u5c42\u5230\u8f93\u51fa\u5c42\u7684\u6743\u91cd\u77e9\u9635
\n self.W_hy = np.random.randn(hidden_size, output_size) * 0.01<\/p>\n
\n self.b_h = np.zeros((1, hidden_size))
\n # \u521d\u59cb\u5316\u8f93\u51fa\u5c42\u504f\u7f6e\u9879
\n self.b_y = np.zeros((1, output_size))<\/p>\n
\n def tanh(self, x):
\n return np.tanh(x) # \u9690\u85cf\u5c42\u6fc0\u6d3b\u51fd\u6570<\/p>\n
\n def softmax(self, x):
\n exp_x = np.exp(x – np.max(x, axis=1, keepdims=True)) # \u9632\u6b62\u6ea2\u51fa
\n return exp_x \/ np.sum(exp_x, axis=1, keepdims=True) # \u8f93\u51fa\u5c42\u6fc0\u6d3b\u51fd\u6570<\/p>\n
\n def forward(self, X):
\n # \u83b7\u53d6\u8f93\u5165\u5e8f\u5217\u957f\u5ea6\u3001\u6279\u6b21\u5927\u5c0f
\n seq_len, batch_size, _ = X.shape
\n # \u521d\u59cb\u5316\u9690\u85cf\u5c42\u72b6\u6001
\n self.h = np.zeros((seq_len + 1, batch_size, self.hidden_size))
\n # \u521d\u59cb\u5316\u8f93\u51fa\u5c42\u72b6\u6001
\n self.y = np.zeros((seq_len, batch_size, self.output_size))<\/p>\n
\n # \u8ba1\u7b97\u9690\u85cf\u5c42\u72b6\u6001
\n # h_t = tanh(W_xh * x_t + W_hh * h_t-1 + b_h)
\n self.h[t + 1] = self.tanh(np.dot(X[t], self.W_xh) +
\n np.dot(self.h[t], self.W_hh) + self.b_h)<\/p>\n
\n # y_t = softmax(W_hy * h_t + b_y)
\n self.y[t] = self.softmax(np.dot(self.h[t + 1], self.W_hy) + self.b_y)<\/p>\n
\n def backward(self, X, y):
\n # \u83b7\u53d6\u8f93\u5165\u5e8f\u5217\u957f\u5ea6\u3001\u6279\u6b21\u5927\u5c0f
\n seq_len, batch_size, _ = X.shape
\n # \u521d\u59cb\u5316\u9690\u85cf\u5c42\u72b6\u6001\u68af\u5ea6
\n self.dh = np.zeros((seq_len + 2, batch_size, self.hidden_size)) # \u589e\u52a0\u4e00\u4e2a\u4f4d\u7f6e\u7528\u4e8e\u6700\u540e\u4e00\u4e2a\u9690\u85cf\u72b6\u6001\u7684\u68af\u5ea6
\n # \u521d\u59cb\u5316\u8f93\u51fa\u5c42\u72b6\u6001\u68af\u5ea6
\n self.dy = np.zeros((seq_len, batch_size, self.output_size))
\n # \u521d\u59cb\u5316\u8f93\u5165\u5c42\u72b6\u6001\u68af\u5ea6
\n self.dx = np.zeros((seq_len, batch_size, self.input_size))
\n # \u521d\u59cb\u5316\u9690\u85cf\u5c42\u5230\u8f93\u51fa\u5c42\u7684\u6743\u91cd\u77e9\u9635\u68af\u5ea6
\n self.dW_hy = np.zeros_like(self.W_hy)
\n # \u521d\u59cb\u5316\u9690\u85cf\u5c42\u5230\u9690\u85cf\u5c42\u7684\u6743\u91cd\u77e9\u9635\u68af\u5ea6
\n self.dW_hh = np.zeros_like(self.W_hh)
\n # \u521d\u59cb\u5316\u8f93\u5165\u5c42\u5230\u9690\u85cf\u5c42\u7684\u6743\u91cd\u77e9\u9635\u68af\u5ea6
\n self.dW_xh = np.zeros_like(self.W_xh)
\n # \u521d\u59cb\u5316\u9690\u85cf\u5c42\u504f\u7f6e\u9879\u68af\u5ea6
\n self.db_h = np.zeros_like(self.b_h)
\n # \u521d\u59cb\u5316\u8f93\u51fa\u5c42\u504f\u7f6e\u9879\u68af\u5ea6
\n self.db_y = np.zeros_like(self.b_y)<\/p>\n
\n for t in reversed(range(seq_len)):
\n # \u8ba1\u7b97\u8f93\u51fa\u5c42\u72b6\u6001\u68af\u5ea6 (softmax+\u4ea4\u53c9\u71b5\u7684\u8054\u5408\u68af\u5ea6)
\n self.dy[t] = self.y[t] – y[t]<\/p>\n
\n self.dW_hy += np.dot(self.h[t + 1].T, self.dy[t])<\/p>\n
\n self.db_y += np.sum(self.dy[t], axis=0, keepdims=True)<\/p>\n
\n dh_from_output = np.dot(self.dy[t], self.W_hy.T)<\/p>\n
\n # \u6ce8\u610f：t+1\u662f\u5f53\u524d\u9690\u85cf\u5c42\u7684\u7d22\u5f15，\u4e0b\u4e00\u4e2a\u9690\u85cf\u5c42\u662ft+2
\n dh_from_next = np.dot(self.dh[t + 2], self.W_hh.T) if (t + 1) < seq_len else 0<\/p>\n
\n self.dh[t + 1] = (dh_from_output + dh_from_next) * (1 – self.h[t + 1] ** 2)<\/p>\n
\n self.dW_hh += np.dot(self.h[t].T, self.dh[t + 1])<\/p>\n
\n self.dW_xh += np.dot(X[t].T, self.dh[t + 1])<\/p>\n
\n self.db_h += np.sum(self.dh[t + 1], axis=0, keepdims=True)<\/p>\n
\n self.dx[t] = np.dot(self.dh[t + 1], self.W_xh.T)<\/p>\n
\n def update_weights(self):
\n # \u4f7f\u7528\u68af\u5ea6\u4e0b\u964d\u6cd5\u66f4\u65b0\u6240\u6709\u6743\u91cd\u548c\u504f\u7f6e
\n self.W_xh -= self.learning_rate * self.dW_xh
\n self.W_hh -= self.learning_rate * self.dW_hh
\n self.W_hy -= self.learning_rate * self.dW_hy
\n self.b_h -= self.learning_rate * self.db_h
\n self.b_y -= self.learning_rate * self.db_y<\/p>\n
\n self.dW_xh *= 0
\n self.dW_hh *= 0
\n self.dW_hy *= 0
\n self.db_h *= 0
\n self.db_y *= 0<\/p>\n
\n def train(self, X, y, epochs=100):
\n # \u8bad\u7ec3\u6a21\u578b
\n for epoch in range(epochs):
\n # \u524d\u5411\u4f20\u64ad
\n y_pred = self.forward(X)
\n # \u53cd\u5411\u4f20\u64ad
\n self.backward(X, y)
\n # \u66f4\u65b0\u6743\u91cd
\n self.update_weights()<\/p>\n
\n loss = -np.mean(np.sum(y * np.log(y_pred + 1e-8), axis=1))
\n # \u6253\u5370\u635f\u5931
\n if epoch % 10 == 0:
\n print(f"Epoch {epoch}, Loss: {loss:.4f}")<\/p>\n
\n def predict(self, X):
\n y_pred = self.forward(X)
\n # \u53d6\u6700\u540e\u4e00\u4e2a\u65f6\u95f4\u6b65\u7684\u8f93\u51fa，\u5e76\u8fd4\u56de\u6982\u7387\u6700\u5927\u7684\u7c7b\u522b
\n return np.argmax(y_pred[-1], axis=1)<\/p>\n
\ndef create_dataset():
\n """\u521b\u5efa\u7b80\u5355\u7684\u60c5\u611f\u5206\u6790\u6570\u636e\u96c6"""
\n # \u6269\u5c55\u8bad\u7ec3\u6570\u636e\u96c6
\n data = {
\n 'text':[
\n '\u8fd9\u4e2a\u4ea7\u54c1\u5f88\u597d', '\u8d28\u91cf\u592a\u5dee\u4e86', '\u670d\u52a1\u4e0d\u9519', '\u4ef7\u683c\u592a\u8d35',
\n '\u7269\u6d41\u5f88\u5feb', '\u5305\u88c5\u7cbe\u7f8e', '\u5ba2\u670d\u8010\u5fc3', '\u5546\u54c1\u6709\u7455\u75b5',
\n '\u6027\u4ef7\u6bd4\u9ad8', '\u4f7f\u7528\u4e0d\u65b9\u4fbf', '\u989c\u8272\u5f88\u597d\u770b', '\u5c3a\u5bf8\u4e0d\u5408\u9002',
\n '\u6750\u8d28\u4e0d\u9519', '\u5473\u9053\u4e0d\u597d\u95fb', '\u4ef7\u683c\u5408\u7406', '\u7269\u6d41\u592a\u6162',
\n '\u5305\u88c5\u7834\u635f', '\u5ba2\u670d\u6001\u5ea6\u597d', '\u8fd9\u4e2a\u4e1c\u897f\u5f88\u597d\u7528', '\u8d28\u91cf\u975e\u5e38\u5dee',
\n '\u670d\u52a1\u6001\u5ea6\u5f88\u597d', '\u4ef7\u683c\u771f\u7684\u8d35', '\u7269\u6d41\u901f\u5ea6\u5f88\u5feb', '\u5305\u88c5\u5f88\u7528\u5fc3',
\n '\u5ba2\u670d\u975e\u5e38\u8010\u5fc3', '\u5546\u54c1\u8d28\u91cf\u6709\u95ee\u9898', '\u6027\u4ef7\u6bd4\u975e\u5e38\u9ad8', '\u4f7f\u7528\u8d77\u6765\u4e0d\u65b9\u4fbf'
\n ],
\n 'label':[1, 0, 1, 0, 1, 1, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 0, 1, 1, 0, 1, 0, 1, 1, 1, 0, 1, 0]
\n }<\/p>\n
\n df = pd.DataFrame(data)
\n return df<\/p>\n
\ndef preprocess_text(text):
\n # \u5206\u8bcd
\n seg_list = jb.cut(text)
\n # \u5408\u5e76\u5206\u8bcd\u7ed3\u679c
\n return " ".join(seg_list)<\/p>\n
\ndef text_to_sequence(text, vocab, max_len=5):
\n words = text.split() # \u6309\u7a7a\u683c\u5206\u8bcd
\n sequence = [] # \u521d\u59cb\u5316\u5e8f\u5217
\n for word in words[:max_len]: # \u53ea\u53d6\u524dmax_len\u4e2a\u5355\u8bcd
\n if word in vocab:
\n sequence.append(vocab[word]) # \u83b7\u53d6\u8bcd\u5bf9\u5e94\u7684ID
\n else:
\n sequence.append(0) # \u672a\u77e5\u8bcd\u75280\u8868\u793a<\/p>\n
\n while len(sequence) < max_len:
\n sequence.append(0)<\/p>\n
\ndef main():
\n # 1.\u6570\u636e\u51c6\u5907
\n print("===\u6570\u636e\u51c6\u5907===")
\n df = create_dataset()
\n print(f"\u6570\u636e\u96c6\u5927\u5c0f：{len(df)}")<\/p>\n
\n print("\\\\n===\u6587\u672c\u9884\u5904\u7406===")
\n df['processed'] = df['text'].apply(preprocess_text)<\/p>\n
\n print("\\\\n===\u6784\u5efa\u8bcd\u6c47\u8868===")
\n all_words = []
\n for text in df['processed']:
\n all_words.extend(text.split()) # \u6536\u96c6\u6240\u6709\u5206\u8bcd\u540e\u7684\u8bcd<\/p>\n
\n vocab = {'<PAD>': 0}
\n for word in set(all_words): # \u53bb\u91cd\u5e76\u904d\u5386\u6240\u6709\u8bcd
\n if word not in vocab: # \u5982\u679c\u8bcd\u4e0d\u5728\u8bcd\u6c47\u8868\u4e2d
\n vocab[word] = len(vocab) # \u52a0\u5165\u8bcd\u6c47\u8868<\/p>\n
\n print(f"\u8bcd\u6c47\u8868\u5927\u5c0f：{vocab_size}")<\/p>\n
\n print("\\\\n===\u5e8f\u5217\u8f6c\u6362===")
\n max_len = 5 # \u6700\u5927\u5e8f\u5217\u957f\u5ea6
\n sequences = [] # \u521d\u59cb\u5316\u5e8f\u5217\u5217\u8868<\/p>\n
\n seq = text_to_sequence(text, vocab, max_len) # \u5c06\u6587\u672c\u8f6c\u6362\u4e3a\u5e8f\u5217
\n sequences.append(seq) # \u52a0\u5165\u5e8f\u5217\u5217\u8868<\/p>\n
\n X = np.zeros((len(sequences), max_len, vocab_size))
\n for i, seq in enumerate(sequences):
\n for j, word_index in enumerate(seq):
\n X[i, j, word_index] = 1<\/p>\n
\n y = df['label'].values
\n y_onehot = np.zeros((len(y), 2))
\n y_onehot[np.arange(len(y)), y] = 1<\/p>\n
\n X_train, X_test, y_train, y_test = train_test_split(
\n X, y_onehot, test_size=0.25, random_state=42
\n )
\n # \u8c03\u6574\u7ef4\u5ea6\u4e3a(seq_len, batch_size, input_size)
\n X_train = X_train.transpose(1, 0, 2)
\n X_test = X_test.transpose(1, 0, 2)<\/p>\n
\n print("\\\\n===\u8bad\u7ec3RNN\u6a21\u578b===")
\n hidden_size = 32 # \u9690\u85cf\u5c42\u5927\u5c0f
\n output_size = 2 # \u8f93\u51fa\u5c42\u5927\u5c0f(\u4e8c\u5206\u7c7b)<\/p>\n
\n model.train(X_train, y_train, epochs=200)<\/p>\n
\n print("\\\\n===\u6a21\u578b\u8bc4\u4f30===")
\n y_pred = model.predict(X_test)
\n y_test_labels = np.argmax(y_test, axis=1)
\n accuracy = accuracy_score(y_test_labels, y_pred)
\n print(f"\u6d4b\u8bd5\u96c6\u51c6\u786e\u7387：{accuracy:.4f}")<\/p>\n
\n print("\\\\n===\u9884\u6d4b\u65b0\u6587\u672c===")<\/p>\n
\n '\u8fd9\u4e2a\u5546\u54c1\u5f88\u6709\u4ef7\u503c',
\n '\u670d\u52a1\u7cdf\u7cd5',
\n '\u7269\u6d41\u5f88\u53ca\u65f6',
\n '\u5305\u88c5\u5f88\u7cbe\u7f8e',
\n '\u5ba2\u670d\u5f88\u8010\u5fc3'
\n ]<\/p>\n
\n processed = preprocess_text(text) # \u9884\u5904\u7406
\n seq = text_to_sequence(processed, vocab, max_len) # \u8f6c\u6362\u4e3a\u5e8f\u5217<\/p>\n
\n x = np.zeros((max_len, 1, vocab_size))
\n for j, word_id in enumerate(seq):
\n x[j, 0, word_id] = 1<\/p>\n
\n result = model.predict(x)[0]
\n sentiment = "\u6b63\u9762" if result == 1 else "\u8d1f\u9762"
\n print(f"\u6587\u672c: {text}")
\n print(f"\u9884\u6d4b: {sentiment}")
\n print()<\/p>\n
\n main()<\/p>\n9. \u603b\u7ed3\u4e0e\u5c55\u671b<\/h2>\n
\n