Traditional sentiment analysis treats opinion mining as a coarse classification task, determining whether a sentence or document is generally positive or negative. However, real-world language is rarely so uniform. A single sentence can express multiple distinct opinions about different features of a product or service—for instance, "The food was delicious, but the service was incredibly slow."
To capture this nuance, the field has shifted toward fine-grained, structured sentiment analysis. As formalized by Haiyun Peng et al. (2020), this requires extracting unified sentiment triplets that identify the What (the aspect term), the Why (the explicit opinion descriptor), and the How (the sentiment polarity). Recent work has extended this to sentiment quadruplets, adding broader aspect categories or identifying explicit opinion holders (An Wang et al. (2023); Franck Dernoncourt et al. (2024)).
The core challenge of structured sentiment extraction is not simply identifying these individual components, but correctly mapping the complex, overlapping relationships between them.
The Bottleneck: Overlapping and Many-to-Many Relations
Early approaches attempted to solve this problem using standard sequence labeling pipelines (e.g., extracting aspects and opinions independently via Named Entity Recognition and then pairing them). However, Jie Zhou et al. (2022) note that because these components are functionally and structurally linked, extracting them independently frequently results in mismatched or orphaned sentiment tuples.
This issue becomes acute when sentences exhibit one-to-many or many-to-many relationships, often referred to as the "multi-overlap" problem (Yanghui Rao et al. (2021); Shiman Zhao et al. (2022)). Consider two common linguistic patterns:
- One Aspect, Multiple Opinions: "The crust is thin, crispy, and perfect." (One target aspect maps to three distinct opinion words).
- Multiple Aspects, One Opinion: "The chicken and the beef were both overcooked." (Two target aspects share a single opinion word).
If a model extracts these spans in isolation, it struggles to determine which opinion belongs to which target. In conversational settings, these relations are further complicated because the target, aspect, and opinion phrases may be scattered across different speakers and dialogue turns (Bobo Li 0001 et al. (2023)).
To solve this, modern architectures focus on joint extraction frameworks that bind boundaries and relations simultaneously. Three primary paradigms have emerged to handle this relational mapping: grid tagging, dual syntactic-semantic graphs, and query-based generation.
Paradigm 1: 2D Grid Tagging Schemes
Instead of treating the sentence as a 1D sequence of tokens, grid-based methods project the sentence onto a 2D matrix where both the rows and columns represent the tokens of the sentence.
The crust is thin and crispy
The [ ] [ A ] [ ] [ ] [ ] [ ]
crust [ ] [ A ] [ ] [ ] [ ] [ ]
is [ ] [ ] [ ] [ ] [ ] [ ]
thin [ ] [ POS ] [ ] [ O ] [ ] [ ]
and [ ] [ ] [ ] [ ] [ ] [ ]
crispy [ ] [ POS ] [ ] [ ] [ ] [ ]
In this formulation, the diagonal and near-diagonal elements are used to detect entity boundaries (e.g., marking "crust" as an Aspect span A, and "thin" or "crispy" as Opinion spans O). Crucially, the intersection cells between row and column spans explicitly encode the relationship and its polarity (e.g., the intersection of "thin" and "crust" is tagged as POS for positive sentiment).
Qiao Sun et al. (2024) demonstrate that this minimalist 2D grid representation, when paired with token-level contrastive learning, allows specialized models to outperform massive large language models. Because every pair of words has an explicit coordinate in the matrix, the model can naturally capture multiple overlapping relations without running into alignment or collision errors.
Paradigm 2: Dual Syntactic-Semantic Graphs
Another intuition relies on the fact that span boundaries and semantic relationships are governed by different linguistic rules. Span boundaries are largely determined by local constituency syntax (phrases), while sentiment links are governed by long-range dependency syntax (which words modify which).
To exploit this, the authors of several recent papers have proposed dual-tree or graph-sequence architectures. For example, Kui Meng et al. (2022) designed a Multi-Task Dual-Tree Network that routes information through two distinct channels:
- Constituency Trees are leveraged to optimize boundary detection, ensuring that multi-word aspects (like "unbelievably fresh bluefin tuna") are cleanly isolated.
- Dependency Trees are modified and leveraged to map the relational paths between those isolated boundaries, tracking how modifiers propagate sentiment across the sentence.
By combining these structural views, models like the one proposed by Zhexue Chen et al. (2021) can simultaneously learn word order and complex linguistic dependencies, preventing the model from breaking down when sentences deviate from simple subject-verb-object structures.
Paradigm 3: Query-Based and Generative Binding
A third paradigm avoids explicit structural graphs or matrices by framing extraction as an iterative or generative process.
Yi Shen et al. (2021) introduced a Dual-Machine Reading Comprehension (MRC) framework that conceptualizes triplet extraction as a multi-turn dialogue. The model first issues a query to find all aspects in a sentence (e.g., "Which targets are mentioned?"). For every aspect it discovers, it generates a new, context-dependent query to locate the corresponding opinions and polarities (e.g., "What is the opinion and sentiment toward [crust]?"). This multi-turn querying naturally binds the components together: an opinion is never extracted in a vacuum; it is always extracted in response to a specific aspect anchor.
Alternatively, this can be handled via sequence-to-sequence generation. By using instruction tuning, Siddharth Varia et al. (2023) showed that complex relational structures can be flattened into natural language target strings (e.g., generating text like "The aspect is crust, the opinion is thin, the sentiment is positive"). While Gaku Morio et al. (2022) note that generative models can occasionally suffer from alignment errors when mapping generated tokens back to the exact source text anchors, they offer high precision and adapt exceptionally well to low-data, few-shot scenarios.
Key Foundational Work
Multilingual Connotation Frames: A Case Study on Social Media for Targeted Sentiment Analysis and Forecast (2017) — contributed 1 concept elements to this direction.
SentiHood: Targeted Aspect Based Sentiment Analysis Dataset for Urban Neighbourhoods (2016) — contributed 1 concept elements to this direction.