Humor Computation Based On Linguistic Features

Humor is a common and important phenomenon in daily life.It can adjust the mood,active atmosphere and promote communication.It can be said that humor is the crystallization of human wisdom.Humor is a kind of sentiment in nature,but its causes and expressions are more complicated than the general explicit sentiment,which belongs to the implicit sentiment.In recent years,with the rapid development of the Internet,people share a large number of humorous texts on social media such as Weibo and BBS.At the same time,machine learning and deep learning have achieved remarkable results in the field of explicit sentiment analysis.These two factors promote the development of humor computation,which helps computers to understand humor better from two aspects: humor recognition and humor generation.Humor computation is of great value and significance in both academic and industrial fields.This paper focuses on the recognition of humor,and explores the generation of humor.Firstly,this paper proposes a new method of humor level recognition.Traditional humor recognition tasks are designed to distinguish between humorous and non-humorous texts and treat humor as a binary feature.Recent research suggests that humor is continuous,and different humorous texts have different levels of humor,which can be classified and identified.For this task,this paper proposes a multi-granularity semantic interactive understanding network,which interacts and understands the semantic information of "setup" and "punchline" in humorous text from the two levels of words and clauses,so as to predict the humor level of the text.In this paper,we conduct experiments on a public humor dataset and prove that modeling the semantic relationship between the setup and the punchline can effectively improve the performance of the model in recognizing humor levels.Secondly,this paper proposes a feature integrated classification method for Chinese humor.This paper studies the classification of Chinese humor,including the recognition of Chinese humor level and the classification of Chinese humor type.The former is based on the humor degree of the text,and the latter is based on the causes of the humor contained in the text.This paper explores four special features of Chinese humor,including phonetic,semantic,grammar and style,and proposes an attention network integrating linguistic features.Experimental results show that the proposed method achieves the current optimal results on both tasks,which proves that the addition of linguistic features can significantly improve the performance of the model on the Chinese humor classification task.Finally,this paper proposes a framework for humor text generation.The existing research on humor generation is mainly based on two ideas: style transfer and text ending.This paper proposes a framework of humor generation combining the two ideas.The framework consists of three modules: local modification,punchline generation and punchline selection.The local modification module generates the setup through word substitution.The punchline generation module uses two ways: conditional generation and heuristic retrieval to get the candidate set of punchline corresponding to the setup.The punchline selection module proposes two schemes of humor level recognition and semantic relation detection to select the best punchline from the candidate set.In addition,three artificial evaluation metrics,humor degree,novelty degree and fluency degree,were proposed for the task of humor generation.Experiments show that the humor generation framework proposed in this paper can achieve close to the effect of artificial humor generation on three scoring metrics.