Multi-SVM For Aspect Based Sentiment Analysis

Sentiment Analysis is a research field that analyzes people's opinions,evaluations,attitudes and emotions from the text.Aspect based sentiment analysis is the latest task in the field of sentiment analysis,which focuses on identifying the sentiment polarity of different aspects in a document.Compared to common document based sentiment analysis techniques,we not only need to identify and extract descriptions with emotional colors,but also accurately assign those descriptions to corresponding aspects using correctly captured grammatical information.In a word,aspect based sentiment analysis has a considerable research difficulty and practice value.Currently,there are four kinds of methods proposed to solve the problem and all of them have particular shortcomings.The rule-based methods rely heavily on the participation of experts,and they are difficult to migrate in different fields.Traditional machine learning methods are usually limited by the amounts of parameters of the classifier.Deep learning methods are limited by the amounts and diversity of data sets.Large scale pre-trained models based methods need large computation and storage costs,and the models are slow in inference.In view of the above situations,we propose an aspect clustering based multi-support vector machine model to tackle the aspect based sentiment analysis.We consider that the samples having similar aspects may correspond to similar features,so we represent aspect by word embeddings and then perform K-Means clustering method to divide the original data set into several data subsets.The samples in one data subset may have some similarities.By extracting aspect-related features and aspect-independent features on a specific data subset,potential aspect category information and general semantic information can be effectively captured.We also use sequential model-based automatic machine learning techniques to optimize the support vector machine.Extensive experiments show that our proposed model has low calculation and storage costs,accurate determination of sentiment polarity,and fast determination speed.