Sentic Computing for Aspect-Based Opinion Summarization Using Multi-Head Attention with Feature Pooled Pointer Generator Network

Kumar, A ORCID: https://orcid.org/0000-0003-4263-7168, Seth, S, Gupta, S and Maini, S (2022) Sentic Computing for Aspect-Based Opinion Summarization Using Multi-Head Attention with Feature Pooled Pointer Generator Network. Cognitive Computation, 14 (1). pp. 130-148. ISSN 1866-9956

Preview

Accepted Version Available under License In Copyright. Download (2MB) | Preview

Abstract

Neural sequence to sequence models have achieved superlative performance in summarizing text. But they tend to generate generic summaries that under-represent the opinion-sensitive aspects of the document. Additionally, the sequence to sequence models are prone to test-train discrepancy (exposure-bias) arising from the differential summary decoding processes in the training and testing phases. The models use ground truth summary words in the decoder training phase and predicted outputs in the testing phase. This inconsistency leads to error accumulation and substandard performance. To address these gaps, a cognitive aspect-based opinion summarizer, Feature Pooled Pointer Generator Network (FP2GN), is proposed which selectively attends to thematic and contextual cues to generate sentiment-aware review summaries. This study augments the pointer generator framework with opinion feature extraction, feature pooling, and mutual attention mechanism for opinion summarization. The proposed model FP2GN identifies the aspect terms in review text using sentic computing (SenticNet 5 and concept frequency-inverse opinion frequency) and statistical feature engineering. These aspect terms are encoded into context embeddings using weighted average feature pooling, which is processed in a pointer-generator framework inspired stacked Bi-LSTM encoder–decoder model with multi-head self-attention. The decoder system uses temporal and mutual attention mechanisms to ensure the appropriate representation of input-sequence. The study also proffers the use of teacher forcing ratio to curtail the exposure-bias-related error-accumulation. The model achieves ROUGE-1 score of 86.04% and ROUGE-L score of 88.51% on the Amazon Fine Foods dataset. An average gain of 2% over other methods is observed. The proposed model reinforces pointer generator network architecture with opinion feature extraction, feature pooling, and mutual attention mechanism to generate human-readable opinion summaries. Empirical analysis substantiates that the proposed model is better than the baseline opinion summarizers.