Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Revisiting Structured Sentiment Analysis as Latent Dependency Graph Parsing

Chengjie Zhou, Bobo Li, Hao Fei, Fei Li, Chong Teng, Donghong Ji

TL;DR

This paper reframes Structured Sentiment Analysis (SSA) as latent dependency graph parsing by treating flat sentiment spans as latent subtrees and solving them with a constrained TreeCRF in a two-stage parsing framework. The encoder–scorer architecture combines rich token representations with higher-order and headed-span scoring to jointly predict expression, holder, and target spans while marginalizing latent internal structures. The approach yields state-of-the-art results across five benchmarks, particularly improving long-span boundary detection and overall sentiment graph accuracy, and maintains end-to-end optimization. This latent-tree formulation advances SSA by explicitly modeling internal span structures, offering improved interpretability and robustness for complex sentiment structures with potential practical impact on negation and hedge detection in real-world text.

Abstract

Structured Sentiment Analysis (SSA) was cast as a problem of bi-lexical dependency graph parsing by prior studies. Multiple formulations have been proposed to construct the graph, which share several intrinsic drawbacks: (1) The internal structures of spans are neglected, thus only the boundary tokens of spans are used for relation prediction and span recognition, thus hindering the model's expressiveness; (2) Long spans occupy a significant proportion in the SSA datasets, which further exacerbates the problem of internal structure neglect. In this paper, we treat the SSA task as a dependency parsing task on partially-observed dependency trees, regarding flat spans without determined tree annotations as latent subtrees to consider internal structures of spans. We propose a two-stage parsing method and leverage TreeCRFs with a novel constrained inside algorithm to model latent structures explicitly, which also takes advantages of joint scoring graph arcs and headed spans for global optimization and inference. Results of extensive experiments on five benchmark datasets reveal that our method performs significantly better than all previous bi-lexical methods, achieving new state-of-the-art.

Revisiting Structured Sentiment Analysis as Latent Dependency Graph Parsing

TL;DR

This paper reframes Structured Sentiment Analysis (SSA) as latent dependency graph parsing by treating flat sentiment spans as latent subtrees and solving them with a constrained TreeCRF in a two-stage parsing framework. The encoder–scorer architecture combines rich token representations with higher-order and headed-span scoring to jointly predict expression, holder, and target spans while marginalizing latent internal structures. The approach yields state-of-the-art results across five benchmarks, particularly improving long-span boundary detection and overall sentiment graph accuracy, and maintains end-to-end optimization. This latent-tree formulation advances SSA by explicitly modeling internal span structures, offering improved interpretability and robustness for complex sentiment structures with potential practical impact on negation and hedge detection in real-world text.

Abstract

Structured Sentiment Analysis (SSA) was cast as a problem of bi-lexical dependency graph parsing by prior studies. Multiple formulations have been proposed to construct the graph, which share several intrinsic drawbacks: (1) The internal structures of spans are neglected, thus only the boundary tokens of spans are used for relation prediction and span recognition, thus hindering the model's expressiveness; (2) Long spans occupy a significant proportion in the SSA datasets, which further exacerbates the problem of internal structure neglect. In this paper, we treat the SSA task as a dependency parsing task on partially-observed dependency trees, regarding flat spans without determined tree annotations as latent subtrees to consider internal structures of spans. We propose a two-stage parsing method and leverage TreeCRFs with a novel constrained inside algorithm to model latent structures explicitly, which also takes advantages of joint scoring graph arcs and headed spans for global optimization and inference. Results of extensive experiments on five benchmark datasets reveal that our method performs significantly better than all previous bi-lexical methods, achieving new state-of-the-art.
Paper Structure (26 sections, 13 equations, 4 figures, 5 tables, 1 algorithm)

This paper contains 26 sections, 13 equations, 4 figures, 5 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. OM
  4. ABSA
  5. SSA
  6. SSA as Latent Graph Parsing Scheme
  7. Conversion and Training on Latent Tree
  8. Decoding and Recovery on Latent Tree
  9. Methodology
  10. Encoder
  11. Tree Scoring
  12. Training Objective
  13. Inside Algorithm
  14. Experiments
  15. Datasets
  16. ...and 11 more sections

Figures (4)

  • Figure 1: (a) An example of original structure sentiment analysis. (b) The head-first parsing graph proposed by barnes_structured_2021. (c) The label strategy proposed by shi_effective_2022. The label formulation proposed by zhai_ussa_2023 is similar to this.
  • Figure 2: Illustration of our SSA as Latent Dependency Graph and Two-stage Parsing. (a) The conversion and training of Stage I-expression extraction: convert the ROOT-expression structure to a dependency tree with (dotted) latent annotations; non-expression spans are assigned a label of $\varnothing$ for clarity. We omit the $\varnothing$ edge from ROOT to "to import ..." for brevity. (b) The conversion and training of corresponding Stage-II-holder/target extraction: with the given expression, convert the expression-holder/target structure to a dependency tree with (dotted) latent annotations and irrelevant spans assigned with $\varnothing$ as well. (c) The decoding and recovery of the converted latent dependency tree: Parsing the best dependency tree (above the sentence) and recovery the tree to SSA structure (below). We combine the separate two-stage decoding and recovery in one figure for brevity.
  • Figure 3: Analysis on long span identification in NoReC Test Dataset. The dataset contains 8448 expressions with up to 30 token and average length of 4.9 tokens. The Head-1 represents head-first method. (a) Expression F$_1$ scores breakdown by span length (b) SF$_1$ scores breakdown by tuple length.
  • Figure 4: Deduction rules for our constrained Inside algorithm based on Eisner algorithm eisner_bilexical_1997, we extend it to second-order and introduce span scoring yang_combining_2022 for unify the deduction framework to expression identification and expression-target/holder relation prediction. (R-Comb and R-Link), combination with span scoring (R-Finish) and its second-order extension (Comb and R-Link2). Our modified rule constraints are highlighted in blue color. $r^h_{*,i}$ denotes a span that takes $h$ as the headword and ends with $i$, $h \in \mathcal{E}$ denotes h is a word within expression span. We show only R-rules, omitting the symmetric L-rules as well as initial conditions for brevity.