Abstract
Aspect Sentiment Triplet Extraction (ASTE) is one of the compound tasks of fine-grained aspect-based sentiment analysis (ABSA), aiming at extracting the triplets of aspect terms, corresponding opinion terms and the associated sentiment orientation. Recent efforts in exploiting span-level semantic interaction shown superior performance on ASTE task. However, most of the existing span-based approaches suffer from enumerating all possible spans, since it can introduce too much noise in sentiment triplet extraction. To ease this burden, we propose a dual-channel span generation method to coherently constrain the search space of span candidates. Specifically, we leverage the syntactic relations among aspect/opinion terms and the associated part-of-speech characteristics in those terms to generate span candidates, which reduces span enumeration by nearly half. Besides, feature representations are learned from syntactic and part-of-speech correlation among terms, which renders span representation fruitful linguistic information. Extensive experiments on two versions of public datasets demonstrate both the effectiveness of our design and the superiority on ASTE/ATE/OTE tasks.
Introduction
In fine-grained sentiment analysis, the goal is often to extract triplets: (Aspect, Opinion, Sentiment). For example, in the sentence "My vegetable risotto was burnt," the objective is to extract the triplet ("vegetable risotto", "burnt", Negative).
Recent high-performing models utilize a "span-based" approach, treating sequences of words as single units. However, these models typically suffer from two major issues: computational explosion and high noise. Enumerating every possible span in a sentence of length $n$ results in $O(n^2)$ candidates, leading to $O(n^4)$ possible interactions between aspect and opinion candidates. Furthermore, most of these $O(n^2)$ spans are linguistically nonsensical (e.g., "was burnt and"), which introduces significant noise that the model must learn to ignore. The Dual-Span model aims to constrain this search space by only considering spans that make linguistic sense.
Methodology
The Dual-Span architecture replaces the "brute-force" enumeration of all possible text spans with a linguistically-informed dual-channel generation process.
Dual-Channel Span Generation
Instead of taking every possible sub-string, the model generates candidates through two filters:
- Syntactic Channel: This channel extracts spans that are connected via dependency relations in a syntactic tree. For instance, "vegetable" and "risotto" are linked by a "compound" relation, marking them as a valid candidate span.
- Part-of-Speech (POS) Channel: This channel identifies spans centered around nouns (NN) or adjectives (JJ) within a small window, capturing common patterns for aspects and opinions.
Feature Enhancing Module (RGAT)
The model uses two Relational Graph Attention Networks—SynGAT and PosGAT. These learn high-order interactions (e.g., words that are two hops away in a dependency tree) to enrich the representation of each word before span classification. A learnable gating mechanism then fuses the syntactic and lexical information, allowing the model to decide which linguistic view is more relevant for a specific span.
Triplet Module
Finally, the model pairs the most likely aspect and opinion candidates, classifies their sentiment polarity, and prunes invalid combinations to produce the final triplet set.
Experimental Evaluation
The model was evaluated on four widely used public benchmarks from the SemEval challenges: Lap14 (laptop reviews) and Res14, Res15, Res16 (restaurant reviews). The authors tested on two versions of these datasets: D1 (the original ASTE-Data-v1) and D2 (a refined version, ASTE-Data-v2).
Key Results
Dual-Span consistently outperformed previous state-of-the-art (SOTA) models, such as Span-ASTE and SBN, on both BiLSTM and BERT encoders.
On the D2 dataset, it achieved an F1 score of 75.47 on Res14, representing a significant improvement over previous benchmarks. Beyond accuracy, the dual-channel strategy cut the time required for span generation by roughly 50% compared to the vanilla Span-ASTE model (e.g., 0.44s vs 0.85s on Lap14).
Discussion and Insights
- Linguistic Constraints as Efficiency: Linguistic structure (syntax and POS) is a powerful constraint to reduce input size. Moving from $O(n^2)$ spans to roughly $O(n)$ spans significantly cleans up the data and reduces noise.
- Complementary Views: Syntax helps identify multi-word terms (like "wine list"), while POS tags help identify the type of term (nouns for aspects, adjectives for opinions).
- Higher-Order Relations: By using 2-layer RGATs, the model captures "indirect" relations. For example, an aspect might be linked to a verb, which is then linked to an opinion. Standard span models often miss these multi-hop connections.
Conclusion
The Dual-Span model demonstrates that linguistic information can effectively prune the search space for ASTE tasks without sacrificing accuracy. While the model currently faces minor limitations—such as missing opinion terms that are past participles (VBN) and a reliance on the accuracy of the underlying dependency parser—it sets a new standard for efficiency and performance in fine-grained sentiment analysis. Future work will focus on expanding the POS channel to include more tags and exploring more complex relational structures.
Summary updated