Research On Causal Relation Identification Of Emotional Events In Texts

With the rapid development of the Internet,users have changed from simple infor-mation acquirers in the early days to information publishers now.The huge user group constantly communicates and interacts,generating a massive amount of emotional infor-mation.These emotional information are fed back to the real world through dissemina-tion,collision,and aggregation in the virtual network,which has a huge impact on the real society and triggers an urgent need for efficient processing of emotional information.Textual affective computing has made great progress,but there are still two obvious prob-lems.First,most of the existing sentiment analysis methods remain in the identification and classification of emotional expression,and the extraction of emotional expression el-ements,but the lack of systematic research on the causes(emotional events)of emotion restricts the depth of emotional understanding research.Second,the evolution of emotion on Internet is also driven by emotional events.The emergence and development of any emotional event often has potential causal relations with other emotional events.However,current research generally lacks the joint modeling of textual emotion cause analysis and event causality identification.Therefore,this thesis conducts an in-depth and systematic research on the causal relation identification between emotional events in texts from the perspectives of emotional text causality identification and event causality identification.The main work includes:Aiming at the problem of poor interpretability due to the lack of fusion domain knowl-edge in existing methods based on deep neural network in text emotion cause recognition,this thesis proposes a text emotion cause recognition method based on multi-level knowl-edge enhanced network.The method first uses tensor product transformation to achieve deep semantic fusion between emotion expressions and candidate causes,and then con-verts key domain knowledge such as emotion dictionaries,cause labels,and relative dis-tances into explicit supervision signals in the training phase through regularization.The relevant parameters are optimized at three levels:word-attention level,sentence-attention level,and prediction level.This method is the first to realize emotion cause recognition by combining domain knowledge and deep neural network without introducing additional pa-rameters.Experimental analysis shows that the performance of text emotion cause recog-nition based on the multi-level knowledge enhanced network is significantly better than the current mainstream methods.Aiming at the problem of lack of pre-sentiment annotations in practical application scenarios,this thesis studies emotion-cause pair extraction methods for emotion expres-sion recognition and its causes discovery.Existing works often combine emotion expres-sions and candidate causes in a Cartesian product manner,and then use a classifier to perform causal discrimination.This kind of method is simple and straightforward,but the time complexity is O(n~2)(n is the clause number of text).In this thesis,a transition-based method is proposed to represent the different states between emotions and candi-date causes through six targeted actions,and transforms the task into a series of action prediction problems.During the transition process,the system removes some options ir-relevant to the result according to the predicted action,the result can be output in a nearly linear time complexity.In addition,this thesis proposes a emotion-cause pair extraction method based on multi-task sequence tagging.This method designs a novel tag system that introduces relative distance and a tag distribution refinement strategy that combines emotion detection and cause detection.It can synchronously extract emotions and the cor-responding causes through sequence tagging,which theoretically guarantees linear time complexity.Results show that the proposed method based on transition and multi-task sequence tagging achieves the best performance on related datasets,while increasing the computational speed of the model by 50-86%.In order to establish the macro-awareness of emotion situation triggered by emo-tional events,it is first necessary to link the same emotional events in different contexts together.Aiming at the shortcomings of existing methods that they often ignore the dis-tributed structural information between events,this thesis proposes an event coreference resolution method based on graph neural network and adaptive constraints.The method retains the features of text sequences,performs multiple associations of events and related entities based on linguistic features,and conducts higher-level modeling of complex con-textual structures through graph convolution operations.In the graph encoding stage,the method designs similarity constraints oriented to representation learning and consistency constraints oriented to the prediction process according to the prediction probability dis-tribution of the model,thereby providing a better decision boundary for subsequent classi-fiers and enhancing the model training stage stability.Experimental results show that the proposed method based on graph neural network and adaptive constraints achieves better performance than existing methods under different evaluation metrics.Aiming at the problem that the existing event causal identification methods ignore the mutual influence of causal relations between different events in the prediction stage,resulting in conflicting prediction results,this thesis proposes an event causality identifica-tion method based on graph neural network and global optimization.This method realizes complex semantic interaction between event instances and event ontology through graph convolution operations,with the use of co-occurrence and co-referential relations.At the same time,the idea of integer linear programming is introduced into the inference stage of the model,and the characteristics of causality(uniqueness,reflexivity,etc.)impose multiple constraints on its feasible region to ensure that the final prediction results of the model are logically self-consistent.Experimental results show that the event causality identification method based on graph neural network and global optimization achieves the best performance under different test scenarios,especially its capability of effectively identifying the causal relation between long-distance events.On the basis of the above research,this thesis further studies the social media-oriented emotional event causality identification method.First,the emotion-cause pair extraction method proposed in this thesis is used to extract emotions and the corresponding emo-tional events in social media on a large scale.Considering the high computational cost of directly performing cross-text coreference resolution on the extracted emotional events,this thesis introduces a sub-topic detection algorithm combined with hierarchical cluster-ing,and divides emotional texts into different subsets according to their topic relevance.Then,coreference resolution of emotional events is performed for each subset.Based on this,combined with the event causality identification method proposed in this thesis,the causal relations between emotional events can be established.Finally,in response to the lack of labeled corpus,this thesis constructs a causality corpus of emotional events for the topic of”Covid-19”to make up for the shortage of data in related research.Experimental results show that the above method can not only obtain the macro emotional distribution caused by an emotional event,but also can better reveal the promotion effect of the devel-opment of emotional events in social media on user emotional changes according to the causal relations between emotional events.