Research On Conversational Emotion Detection Based On Deep Learning

Emotion detection in the conversation has gradually been widely used in different fields such as opinion mining,human resources,medical care and so on.It is an indispensable part of many intelligent applications.Emotion detection aims to analyze the text in the conversation to identify the emotion of each sentence in the conversation.By dong so,machine intelligence can make appropriate responses according to different emotional states for better human-machine communication and interaction,which has a profound impact and positive significance on the development of information technology and intelligence in various fields.Although the current conversational emotion detection methods have achieved good results,the existing methods are not able to integrate textual sentiment information well and are slightly inadequate in terms of feature extraction to truly understand the sentiment in the conversation.Furthermore,existing methods also ignore the intrinsic connection of conversational emotional sequences when delivering output on sentiment labels.In this paper,we address above issues by building appropriate contextual features for conversation emotion detection model.In order to integrate more conversational contextual information,we propose a fusion model of constructing contextual features to better analyze the emotional features in the conversation.Specifically,we first encode the text using the Ro BERTa model and perform initial feature extraction using a global GRU model to build a contextual feature representation.Then we construct an overall feature representation module by concatenating the contextual feature vectors,while using an attention mechanism to construct deeper contextual features.Finally,the two features are stitched together and fed into a classification neural network to produce a final sentiment category.We evaluated the prediction effect of the model on the IEMOCAP and MELD datasets,the model successfully improved performance by constructing diverse contextual features for fusion.In the field of conversational emotion detection,the current research work rarely considers the intrinsic links between emotion sequences,which are also important for detecting emotions.Base on the observation that the polarity of emotion categories of adjacent sessions tends to be the same,we then explore the relationship between conversational emotion.As a second contribution,this paper attempts to explore and verify two different methods.The first method is to model through a conditional random field model,which uses a transfer matrix of conditional random fields to model the transfer of conversational emotions.The second method is to use a multi-task learning approach to model,where the main task of the model is set to emotion detection and the related task is used to discriminate the consistency of emotional tendency.Using the above two ways to make a good attempt on the consistency of emotion transfer,which also provides a certain research basis for future research.This research aims at the needs of conversational emotion detection in artificial intelligence applications.By constructing a richer model for contextual feature extraction and then modelling the relationships between emotion sequences.The performance of the models was evaluated on two publicly available datasets.The experimental results show that compared with the state-of-the-art research,the F1 values of our models on the IEMOCAP and MELD datasets improved by 2.71% and1.68% respectively.