基于深度学习的外语翻译机器人自动问答检索研究

Research on Automatic Question Answering Retrieval of Foreign Language Translation Robot Based on Deep Learning

针对传统日语翻译机器人自动问答检索极易出现语义丢失现象,导致生成答复和检索准确率低的问题,设计一个基于生成式和检索式相结合的外语翻译机器人自动问答检索系统。基于BiLSTM网络中的Seq2Seq模型,通过加入注意力机制、Beam Search算法和TF-IDF算法,分别构建生成式和检索式回复模型;最后将两个模型相结合实现外语翻译机器人自动问答检索。实验结果表明,相较于单向LSTM网络,双向BiLSTM可实现日语语句的双向预测,可有效避免部分语义丢失。且构建的检索式回复模型的回复正确率均保持在90%以上,最高可达95%,生成式模型回复模型在60次测试中,相关回复率最高为96.67%。由此可知,设计的系统可提升外语翻译机器人生成答复和检索准确率,可实现准确的自动问答检索,具备一定的有效性。

The traditional Japanese translation robot is easy to have semantic loss phenomenon, which leads to the low accuracy of reply generation and retrieval. A foreign language translation robot automatic question and answer retrieval system designed based on the combination of generation and retrieval. Based on the Seq2 Seq model in BiLSTM network, generative and retrieval reply models are constructed respectively by adding attention mechanism, Beam Search algorithm and TF-IDF algorithm. Finally, the two models are combined to realize automatic question retrieval of foreign language translation robot. Experimental results show that, compared with unidirectional LSTM networks, bidirectional BiLSTM can realize the bidirectional prediction of Japanese statements, and can effectively avoid partial semantic loss. Moreover, the highest response rate of the constructed retrieval response model was kept above 90% and up to 95%. Among the 60 tests, the highest correlation response rate was 96.67%. It can be seen that the designed system can improve the response generation and retrieval accuracy of the foreign language translation robot, and realize accurate and automatic question and answer retrieval, which has certain effectiveness.