Research On Chat Robot Dialogue Based On Seq2seq Model

With deep breakthroughs in deep learning research,deep learning techniques are more widely used in natural language understanding,word vector technology,Chinese word segmentation,and sentiment analysis.Compared with chat robots based on retrieval technology,chat robots based on deep learning are more scalable and have higher system development efficiency,and have a very broad development prospect.Currently,the seq2 seq model is the most commonly used deep learning model for open-field chat bots,but the traditional seq2 seq model is directly applied to chat bots with problems such as long-distance dependence and security response.Aiming at the problems existing in the traditional seq2 seq model,this paper proposes a chat robot dialogue model,namely the attention mechanism,the beam search algorithm,the chat robot dialogue model obtained by combining BiLSTM and the traditional seq2 seq model.The specific research work is as follows:(1)This paper studies the technology of word embedding.The traditional text representation method cannot express the semantic similarity between words,and when the number of dictionaries is extremely large,the problem of large vector dimension will appear.Word2 vec makes the word vector with semantic information by mapping semantically similar words to similar positions in the vector space,and reduces the amount of computation by "dimension reduction".(2)This paper analyzes the long-distance dependence problem of the traditional seq2 seq model in detail,and proposes the use of BiLSTM and attention nodel.The LSTM used in the traditional seq2 seq model cannot encode back-to-front information.BiLSTM can better capture the semantic dependency of two-way and solve the problem of losing part of the semantic information.The traditional seq2 seq model uses only one fixed-length vector for encoding and decoding.The attention nodel selectively learns these inputs by retaining the intermediate output of the encoder input sequence,and outputs the sequence when the model is output.Associated with it,thus solving the problem of loss of long text sequence information.(3)This paper studies the security response of the traditional seq2 seq model.The output principle of the decoder in the traditional seq2 seq model is to directly output the most probable statement in the candidate result set,but the most probable statement is often the most common statement in the corpus,such as "I don't know","Hello" and other security responses.In this paper,the cluster search algorithm is adopted.Through the sorting and pruning,the chat robot dialogue model in this paper is more responsive than the traditional seq2 seq model,which improves the security response problem.The TensorFlow framework is used to implement the chat robot dialogue model proposed in this paper,and the chat dialogue model proposed in this paper is trained in Chinese dialogue.Through the comparison of the final experimental results,it is verified that the chat robot dialogue model proposed in this paper is effective and feasible.