Research On Emotional Intensity Response Generation Methods For Multi-Turn Conversations

As a key part of dialogue systems,dialogue generation research has received a great deal of attention as it generates responses that are thematically consistent,linguistically diverse,and personalized.Emotion is ubiquitous in people’s daily communication,and it is important to give emotional messages to dialogue systems to meet practical needs such as psychological guidance and emotional comfort,so many researchers have conducted a lot of research on emotional response generation.Emotion response generation is a method of generating emotionally charged replies to a user’s question,either by specifying an emotion category or by randomly generating emotionally charged replies,which mainly addresses the issues of consistency of emotion expression,appropriateness of emotion expression,and quality of emotion response content.The current research on emotion response generation mainly focuses on the expression of a single emotion or a specific intensity of emotion,without considering the impact of changes in the intensity of emotion on the interaction between the two parties in the dialogue process,so in practice,the difference in the intensity of emotion may cause the user to have a poor emotional experience.To introduce emotional intensity and thus improve the appropriateness of emotional response generation in multi-round dialogues,this paper will conduct a study on emotional intensity response generation in multi-round dialogues for the task of generating emotional intensity responses in multi-round dialogues,with the following specific research content and methods.Firstly,the lack of sentiment intensity labels in the current dataset leads to low accuracy of emotional intensity annotation models,therefore,the corpus annotation method is investigated.A rule-based sentiment intensity annotation model with a voting mechanism is proposed.We use multiple rules to obtain different labels for a given utterance,then use the voting mechanism to decide the most accurate label for the utterance,and finally use the decided labels to complete the training of the sentiment intensity annotation model.The results of the emotion intensity annotation experiments show that the rule-based and votingbased emotion intensity annotation model can complete the training and improve 11.9%,11.4%,12.5%,and 15% in terms of Precision,Recall,F1,and kappa index respectively compared with the baseline model,which can be used for emotion intensity annotation of large-scale multi-turn dialogue datasets.Secondly,the lack of prediction of emotional category and intensity in the multi-round dialogue model leads to inappropriate emotional category and intensity generated by the responses,therefore,the multi-round dialogue emotion and the intensity prediction model is investigated and a multi-round dialogue sentiment and intensity prediction unit is proposed.The unit takes the multi-turn dialogue interaction scenario as the starting point,combines the characteristics of human-computer interaction in which the emotion and intensity sequences present a time series,and uses historical information to complete the prediction of emotion and intensity in multi-turn dialogue responses.The results of the emotion and intensity prediction experiments show that the emotion intensity prediction unit improves5.5%,9.1%,and 8.3% in the Precision,Recall,and F1 metrics respectively compared with the baseline model,while the emotion category prediction unit improves 5%,6.1% and 5.1%in the Precision,Recall and F1 metrics respectively compared with the other models.and5.1%,demonstrating the effectiveness of the emotion and intensity prediction units in the prediction task.Finally,the lack of expression of emotional intensity information in the multi-round dialogue response generation model leads to the problem of inconsistency between emotion and intensity in the responses.Therefore,the sentiment intensity dialogue response generation model is investigated and the sentiment intensity response utterance generation unit is proposed.The unit is based on a neural network model,and a gated decoding module is added to realize the expression of emotion and intensity in the decoding process,while a word type selection module is added to realize the expression of emotion and intensity in the text,to control the generation of different types of words explicitly.The experimental results of emotion intensity response generation show that the emotion intensity response utterance generation unit can achieve both emotion consistency expressions and emotion intensity consistency expressions compared to multiple baseline models.This paper investigates the task of generating emotion intensity responses for multiturn conversations and completes automatic and manual evaluation experiments on a multiturn conversation emotion intensity response generation dataset.The results of the automatic evaluation experiments show that the proposed emotion intensity dialogue response generation model achieves optimal results in terms of language quality and authentic responses compared with the baseline model.The results of the manual evaluation experiments show that the emotion intensity dialogue response generation model is better than the baseline model in terms of appropriateness of emotion expression and appropriateness of emotion intensity expression,and can generate responses with appropriate emotion intensity.This effectively solves the problem of poor user experience due to the lack of consideration of emotion intensity in current research and provides a new method for emotion response generation research.