Research On Humor Recognition And Sentiment Analysis Based On Multi-task Learning

Humor is a unique and implicit sentiment of human beings.Humor plays an important role in relieving the tension,activating the atmosphere,and promoting harmonious communication in people's daily life.Humor computation aims to make computers have the ability to recognize and generate humor,which can improve the intelligence of human-computer dialogue systems and has broad application prospects.A large number of user remarks in the social networks provide data support for humor computation,while the rapid development of big data cloud computing and natural language processing technologies provide technical support for humor computation.Humor computation covers humor recognition and humor generation.Since humor recognition is an important basis for humor generation,the focus of this thesis is humor recognition.Firstly,this thesis proposes a new method of humor text judgement.A humorous punchline is often built around background context and external commonsense knowledge.However,most humor recognition models are built based on data-driven strategies,which can't learn prior knowledge beyond the corpus.Therefore,this thesis combines prior knowledge and the highlevel abstraction capabilities of deep learning to propose a humor recognition method that fuses external knowledge sources and distributed semantic representations.It's basic idea is to obtain text embeddings based on external commonsense knowledge and the large-scale pre-trained language model,introducing prior knowledge and general representation of text;using Bidirectional Long and Short-Term Memory Network to learn contextual semantic interaction and long-term dependencies;using self-attention Network to learn the contributions of different words to mine the key elements and causes of humor;combining the attention-weighted features with shallow humor features,and feed them to the humor recognition prediction layer for classification.The experimental results on public datasets show that the combination of commonsense knowledge and pre-trained vector embedding representation and the combination of recurrent neural network and self-attention network can effectively improve the effect of humor recognition.Secondly,this thesis proposes a multi-task learning-based method for humor recognition and sentiment analysis.Traditional researches ignore the correlation between humor and sentiment,modeling humor recognition and sentiment analysis as two separate tasks.Therefore,this thesis treats humor recognition and sentiment analysis as two related tasks for joint modeling.Each task uses its private networks for private features modeling,that is,learning humorous features and sentiment features separately;the hierarchical attention network is used to learn the internal relationship of sentences in a single task and the feature interaction relationship between different tasks;the parameter sharing of multi-task learning is adopted to fuse humor and sentiment features.Finally,the proposed method combines single-task private features and multi-task fusion features for humor level classification and sentiment polarity classification.The experimental results of this method on the humor-sentiment dataset constructed in this thesis show that adding sentiment features to humor recognition by multitask learning way can effectively improve the effect of humor recognition.