【摘要】 在传统的seq2seq模型的基础上,加入时间卷积网络TCN设计了一种自动数学应用题解算器模型seq2seq+TCN.在数据预处理部分采用FoolNLTK分词工具和ELMo词嵌入方法,以保证较高质量的数据输入;利用LSTM和GRU的不同组合构成seq2seq模型的编码器和解码器;在seq2seq模型中加入TCN,以扩大解码器的感受野,保证应用题文本的时序性.在数据集Math23K和AI2上进行实验,结果表明,由BiLSTM-LSTM构成的seq2seq+TCN模型性能最优,答案准确率分别达到70.3%和87.5%,高于现有的2个集成seq2seq的解算器.更多还原

【Abstract】 An automatic math word solver is designed on the basis of the traditional seq2 seq model by adding temporal convolutional network TCN. In the data preprocessing section,word segmentation tool FoolNLTK and word embedding method ELMo are adopted to ensure high-quality data input. Different combinations of LSTM and GRU are used to construct the encoder and decoder of seq2 seq model. TCN in the seq2 seq model can expand receptive field of the decoder and guarantee the timing of the text of the math word problems. The model is trained on databases Math23 K and AI2. The results show that the performance of seq2 seq + TCN model composed of Bi LSTM-LSTM is best,whose accuracies of answers reach 70.3% and 87.5%,respectively,higher than those of two existing solvers integrated with seq2 seq.更多还原