Improving Accuracy of Key Information Acquisition for Social Media Text Summarization

The attention-based sequence-to-sequence (Seq2Seq) model for abstractive summarization has achieved good performance, but it still suffers from generating inaccurate information and repetitive content when applied to noisy social media text. To address these issues, we propose a novel model that includes a feature extractor and a semantic enhancer. In our model, the feature extractor produces feature vectors that contain key information, filtering out noise in the source text. The semantic enhancer takes the concatenation of the feature vectors and the encoder hidden states as input to learn long-range dependencies, further improving semantic relevance and reducing repetition. Experimental results on a Chinese social media dataset show that our model can capture the key information from noisy text and achieve improvement as compared to state-of-the-art baseline models.