Aspect-level sentiment analysis method based on LSTM (Long Short Term Memory) and grammar distance

Abstract

The invention discloses an aspect-level sentiment analysis method based on LSTM (Long Short Term Memory) and grammar distance, which comprises the following steps: S1, feature input: mapping word vectors by using a GloVe pre-training model, and respectively enabling aspect words and context words to pass through Bi-LSTM to obtain hidden representations of the aspect words and the context; s2, semantic feature extraction: respectively extracting semantic features of aspect words and contexts by adopting a graph convolutional network and an mLSTM; s3, a semantic interaction stage of the aspect word and the context word: performing a dot product attention operation on the features of the aspect word and the context extracted in the S2; and S4, an emotion prediction stage: performing maximum pooling operation on the features obtained in the S3, and then performing softmax operation to obtain the final predicted emotion polarity. The probability of correct matching of aspect words and important contexts is improved, a grammar distance weight is introduced to replace a relative distance, the context with a large degree of association with an aspect word method is further extracted on the grammar level, and finally the accuracy of aspect-level sentiment analysis is improved.

Description

technical field [0001] The invention relates to an aspect-level sentiment analysis method based on LSTM and grammatical distance, and mainly relates to the field of natural language. Background technique [0002] Different from traditional sentiment analysis, Aspect-Based Sentiment Analysis (ABSA) has been widely studied because it can analyze the sentiment of specific aspects in sentences. In recent years, deep learning methods have been widely used in aspect-level sentiment analysis, and mostly rely on Long-Short TermMemory (LSTM) and attention mechanisms. LSTM is widely used for feature extraction because it can avoid the problem of vanishing or exploding gradients. Tang et al. used the hidden representation of LSTM to predict sentiment. Ruder et al. used a bidirectional long short-term memory network (Bidirectional LSTM, Bi-LSTM) to obtain better classification results. Li et al. proposed Target-Specific Transformation Networks (Tnet) to convolve contextual features a...

AI Technical Summary

Problems solved by technology

[0003] (1) Using relative position weights, lacking consideration of grammatical distance, it is very likely to mistakenly match aspect words with contexts with low grammatical relevance, resulting in wrong sentiment classification

[0004] (2) Most methods lack the interaction between aspect words and context words, and it is easy to make aspect words match context words that have nothing to do with their emotional polarity judgment, which affects the judgment of emotional polarity

[0005] (3) When obtaining the final representation, only the feature representations of aspect words and context words are considered, and other features are ignored

Images