Dependency Grammar Model For Aspect-level Sentiment Classification And Its Application To Open Source Comment Analysis

Sentiment analysis and opinion mining have become a hot research topic in natural language processing which can extract valuable sentiment information.With the rapid development of social media communities,most people express their standpoints on the internet and generate a large number of text comments with subjective affection every day.The sentiment analysis of comment texts is of great significance for public opinion control,business marketing,social governance,etc.Aspect-level sentiment classification is an entity level sentiment classification task,which can identify the sentiment polarity of multiple aspects in a sentence and can extract more valuable information than sentence level.Recent researches on aspect level sentiment classification integrate sentiment knowledge into the pre-training model to enhance the impact of sentiment words and sentiment polarity.This method ignores the dependency between aspect words and sentiment words in the text.It is important to accurately capture the sentiment information.Most of sentiment classification models are mainly applied to the restaurant or e-commerce field,while there are few researches on the aspect-level sentiment analysis of comments in the open source field.In order to solve the above problems,this paper proposes a pre-training model of aspect-level sentiment classification based on dependency grammar(ASK-Ro BERTa).The main work is as follows:(1)How Net and Sentic Net are integrated as a general sentiment dictionary.SO-PMI algorithm is used to expand the comment sentiment dictionary to build the field sentiment dictionary.This paper develop a series of aspect word mining rules based on part of speech tagging and grammar rules,which consider word dependencies,conjunctions and compounding.The main rules mine a single word of aspect.When a sentiment word modifies more than one aspect,this can be captured by conjunction rules.In most cases,the aspect is more than one word,and is composed of multiple words.Multiword rules can combine compound nouns.(2)The mask mechanism of the pretraining model is optimized by masking the mined sentiment words and aspect words.Word polarity prediction and aspect-sentiment pair prediction pretraining objectives are jointly optimized to better capture the dependency between aspect words and sentiment words.(3)The pre-training objectives are analyzed and verified on four Sem Eval datasets experimentally.ACC and Macro-F1 are used to evaluate the performance of the model.This paper conducts experiments with baseline models.Experimental results show the effectiveness of the proposed model.Finally,the ASK-Ro BERTa model is applied to the open source comment text to mine the aspect terms and conduct sentiment analysis.