| With the increase of netizens' participation,the number of user-generated texts on the network has also increased.By analyzing the emotional tendencies in text comments,we can know the quality of products and can also obtain the public's opinion on a hot event...So,the task of Chinese-oriented sentiment polarity classification is crucial.This paper will build corpus in the fields of hotels and mobile phones to explore the classification task of Chinese sentiment polarity.This paper will work from the following aspects:Sentiment calculation of implicit characteristics: we proposes CNN-Bi LSTM joint model and using the dynamic sentiment calculation method to calculate the semantic similarity between the candidate sentiment words,and the sentences without obvious characteristics were calculated separately.The results show that the performance is greatly improved by dealing with implicit characteristics separately.In addition,we also compare difference between Word2 vec and Elmo training word vectors.Multi-feature fusion method: aspect sentiment and chapter sentiment were taken as characteristics respectively and incorporated into the classification task of sentiment polarity,to explore the influence of different feature qualities and different feature fusion methods on the experiment,and to verify the differences between early fusion and late fusion.The experiment shows that the performances are improved to a certain extent by the integrate of aspect sentiment and chapter sentiment.Multi-granularity joint model: the sentiment intensity is calculated from the three granularity of words,phrases and sentences,and then the sentiment intensity of each granularity is combined.This paper explores two different methods of combining sentiment intensity.Compared with the single-granularity classification of sentiment polarity,the multi-granularity joint model can better consider the relationship between different granularity and solve the problem of inconsistent contextual sentiment. |
PDF Full Text Request