Aspect-Context Level Information Extraction via Transformer Based Interactive Attention Mechanism for Sentiment Classification

Aspect-context sentiment classification aims to classify the sentiments about an aspect that corresponds its context. Typically, machine learning models considers and context separately. They do not execute in parallel. To model contexts aspects separately, most of methods with attention mechanisms typically employ Long Short Term Memory network approach. Attention mechanisms, on other hand, take this into account compute parallel sequencing aspects-context. The interactive mechanism extracts features a specific regarding sequence, which means are considered when generating sequence representations. However, determining relationship between words sentence, does consider semantic dependency information. Moreover, did capture polysemous words. Normally conventional embedding models, such as GloVe word vectors, have been used. In study, transformers embedded approaches overcome problem. For reason, BERT pre-train language is used among sentence. then applied model’s distribution word. final sequence-to-sequence representation terms general classifiers for aspect-level classification. proposed was evaluated two datasets, i.e., Restaurant Laptop review. approach has state-of-the-art results all attained significantly better performance than existing ones.