Multi–scale heterogeneous aspect–based sentiment analysis (MH–ABSA)

Abstract

Aspect-based sentiment analysis (ABSA) demands more than a single polarity label per document — it requires assigning a distinct sentiment to each named attribute a text discusses, even when the relevant opinion words are syntactically distant, negated, or figurative. Existing graph convolutional approaches to ABSA rely on three design decisions that limit how far they can go: sentiment lexicon weights computed once before training that cannot adapt to how words are used in context; a single deterministic dependency parser whose errors embed directly into the adjacency matrix with no correction path; and graphs built exclusively from token-level nodes, leaving clause-level and phrase-level sentiment units without any dedicated representation. MH-ABSA addresses each of these gaps through five independently ablatable modules. LLM-Guided Dynamic Edge Reweighting (LDEM) substitutes static lexicon scores with context-sensitive weights produced by a frozen instruction-tuned language model, blended into the adjacency matrix through a learnable coefficient. Ensemble-Uncertainty Dependency Fusion (EUDF) combines three structurally diverse parsers through a sentence-conditioned MLP, allowing the model’s trust in each parser to adapt with the input register. The Multi-Scale Heterogeneous Graph (MHG) extends the token-level graph to three tiers — token, span, and discourse — so that clause-level sentiment signals can propagate to individual aspect tokens in a single message-passing hop rather than through a long token chain. Two further modules round out the design. Contrastive Sentiment Alignment Module (CSAM) introduces a contrastive training objective built around LLM-generated paraphrases and domain-internal aspect swaps, giving the model a second gradient pathway that operates on meaning rather than surface vocabulary — useful precisely because sarcasm and implicit sentiment defeat both the lexicon and the parser. Domain-Adaptive Edge Masking (DAEM) addresses a subtler problem: polarity scores that are globally reasonable but locally wrong for a given product category, which it handles through per-edge learned gates that suppress unreliable weights at training time. The proposed framework addresses several limitations of existing graph-based ABSA systems, including static lexicon weighting, parser uncertainty, and lack of multi-scale linguistic representations.