| As the "Pearl" in the crown of artificial intelligence,language is becoming more and more important in the study of natural language processing.Communication between natural language and computer is what people have been pursuing for a long time,because it has obvious practical and theoretical significance.Dialogue evaluation is a kind of evaluation of human-computer dialogue system,the key of which is natural language understanding(NLU),and the purpose of textual entailment(TE)is to identify the potential logical reasoning relationship between two texts.Although textual entailment and dialogue evaluation have achieved good results,there are still some problems,for example: only the interaction information between texts is considered in textual entailment,not much attention is paid to the context logic relationship of the text itself,and the computer cannot use the current general knowledge to deal with the implication relationship between specific texts.In conversation evaluation,the logical reasoning relationship between the sentences around the sentence representation has not been fully considered,which sometimes bring nice scores to the conversations that does not answer the question positively by mistake.In view of the above problems,this thesis proposes a dialogue evaluation method combined with textual entailment.A textual entailment recognition method based on multi-level attention mechanism is proposed.Existing mainstream Siamese model and ESIM model cannot use the global information of the text itself for auxiliary reasoning.To solve this problem,this thesis proposes a multi-level attention mechanism to obtain the global information and the interactive information between texts,which improves the accuracy of textual entailment recognition by 4.90% and 1.66% compared with Siamese model and ESIM model.At present,the existing model cannot use the general knowledge and the data set size is too small.This thesis fine tunes the pre-training model,introduces the general knowledge,and the data set size of the pre-training model is large enough.Experimental results show that the textual entailment recognition method based on pre-training model gets 6.07% improvement over based on multi-level attention mechanism.The experimental results show that the accuracy of the pre-training model is greatly improved.This thesis introduces the method of text entailment in dialogue evaluation,the current dialogue evaluation system only obtains sentence features through BiLSTM,without considering the logical reasoning relationship between the dialogue contexts.In order to alleviate this problem,this thesis uses the reasoning advantage of textual entailment method to strengthen the logical connection between dialogue contexts.Finally,the experimental results show that the proposed method is better than the current method based on multi-layer BiLSTM model,which improves scoring the indicators of conversation accomplishment(A-Score),conversation satisfaction(S-score)and conversation effectiveness(E-score).The specific examples of the experiment show that the model designed in this thesis is effective and can review the dialogue history many times,making full use of the logical reasoning relationship of context. |
