Research On Sentiment Analysis Of Multilingual Internet User Generated Content

Sentiment analysis(also named opinion mining),refers to analyze the opinions,evaluations,attitudes,or emotions of user generated texts using nature language processing techniques,text mining techniques,and computer linguistics.Sentiment analysis can be used in several applications,e.g.,stocks movements prediction,public opinion analysis,customer reviews sentiment analysis and so on.According to the length of text,sentiment analysis can be calssified into wordlevel,phrase level,sentence level and document level.The first work of this paper is to construct a multi-dimensional sentiment lexicon.We use pretrained sentiment word vectors and semantic word vectors combined with machine learning algorithms to perform the word-level sentimental intensity prediction task.The corresponding research has been published in the International Conference on Asian Language Processing 2016(IALP 2016).The second work is to classify the sentiment stance of tweet.We treat it as a sentence-level sentiment classification task and adopt welldesigned features combined with machine learning algorithms to tackle the task.The corresponding research has been published in the International Workshop on Semantic Evaluation 2017(SemEval 2017).User generated content often contains multiple emotions which are attached to different entities or aspects.Therefore,the follow-up work of this paper will further explore the fine-grained aspect-based sentiment analysis.In order to reduce the dependence on external resources and error propagation of step-by-step solutions,we propose two ideas and solutions based on deep learning techniques.Specifically,the first idea is to propose a deep learning framework of aspect-based sentiment word embedding(ASWV).This work has already obtained the invention patent.The second idea is to propose a joint end-to-end solution for ABSA.We firstly transform the aspect term extraction and sentiment analysis subtasks into a joint sequence tagging task,then propose two kinds of joint neural network models to solve the task.To verify the effectiveness and robustness of the joint learning method,we perform experiments on standard English datasets and home-made Chinese datasets respectively.The corresponding paper has been accepted by International Joint Conference on Neural Network 2018(IJCNN 2018).