Source-linked AI summary
Better Document-level Sentiment Analysis from RST Discourse Parsing
Parminder Bhatia, Yangfeng Ji, Jacob Eisenstein
TL;DR
Document-level sentiment analysis has often relied on word counts despite the importance of multi-sentence discourse structure. This paper combines RST parses with sentiment analysis through discourse-depth reweighting and recursive composition, finding robust lexicon-based gains and substantial classifier improvements. The main evaluation caveat is that imperfect discourse-relation detection limits the available structural signal.
Problem
Document-level sentiment analysis is dominated by bag-of-words methods, while linguistic context is typically modeled only at the sentence level despite multi-sentence texts.
Method
The paper combines off-the-shelf RST parses with sentiment analysis through dependency-tree discourse-depth reweighting and recursive sentiment propagation.
Results
The reweighting approach improves lexicon-based sentiment analysis by 4–5%, while the recursive model improves accuracy by more than 3% on Stanford data and 0.5% on Pang and Lee data.
Takeaways & Limitations
RST structure offers robust advantages for lexicon-based document sentiment analysis, and recursive modeling can substantially outperform a bag-of-words classifier.
Takeaways & Limitations
Discourse relation detection reaches only 60% accuracy for even the best systems, and evaluating gold parses would require a large corpus with both RST trees and sentiment labels.
Abstract
from arXiv · showhide
Discourse structure is the hidden link between surface features and document-level properties, such as sentiment polarity. We show that the discourse analyses produced by Rhetorical Structure Theory (RST) parsers can improve document-level sentiment analysis, via composition of local information up the discourse tree. First, we show that reweighting discourse units according to their position in a dependency representation of the rhetorical structure can yield substantial improvements on lexicon-based sentiment analysis. Next, we present a recursive neural network over the RST structure, which offers significant improvements over classification-based methods.
1 Introduction
Document-level sentiment analysis needs discourse structure because bag-of-words methods and added linguistic context typically remain below the document level. The paper introduces two RST-based architectures that combine discourse parses with sentiment analysis.
- Document-level sentiment analysis remains dominated by bag-of-words approaches, while added linguistic context typically stops at the sentence level.
- RST structure can favor the final sentence’s negative polarity even when positive sentiment words are more numerous overall.
- Modern RST parsers motivate reconsidering discourse structure for document-level sentiment analysis.The paper notes that contemporary systems improve raw parsing accuracy over SPADE by 5–10%, depending on the metric.
- The paper evaluates discourse-unit reweighting based on position in a dependency-like discourse representation.Weights may be defined with a simple function or learned from a small amount of data.
- The paper also recursively propagates sentiment through the RST parse using an architecture inspired by recursive neural networks.
- Both architectures work with lexicon-based analyzers or trained classifiers; reweighting helps lexicon users, while the recursive model yields further classifier gains.
2 Background
RST models discourse composition from elementary units to whole documents, offering structural information for sentiment analysis beyond word counts. This background motivates combining discourse-aware reweighting or composition with sentiment methods across domains and review datasets.
- 2.1 Rhetorical Structure Theory: RST combines elementary discourse units into progressively larger units that ultimately cover the document.Relations may involve nuclei and satellites or multiple nuclei.
- 2.1 Rhetorical Structure Theory: RST nuclearity motivates weighting text segments by importance, with satellites generally less important than nuclei.Prior work applied this intuition to summarization and intra-sentential sentiment analysis.
- 2.1 Rhetorical Structure Theory: The paper compares discourse-unit reweighting with a compositional method that propagates sentiment polarity up the discourse tree.
- 2.1 Rhetorical Structure Theory: The study uses the publicly available DPLP RST parser, described as having the best relation-identification F-measure among considered systems and linear document-time complexity.
- 2.2 Sentiment analysis: Sentiment analysis often targets overall positive-versus-negative document polarity, but supervised performance depends on same-domain labeled data.
- 2.2 Sentiment analysis: The proposed approach can combine with either sentiment-analysis method and could in principle extend to document-level categories such as politeness.
- 2.2 Sentiment analysis: Evaluation uses 2,000 movie reviews with ten-fold cross-validation and 50,000 reviews with a predefined 50/50 training-test split.
3 Discourse depth reweighting
The approach converts RST structure into EDU depth and uses that depth to reweight local sentiment contributions before producing a document-level prediction. It supports lexicon-based and classification-based sentiment analysis, with larger gains reported for the lexicon-based setting.
- Discourse representation: DEP-DT converts the constituent-like RST tree into a directed graph over elementary discourse units, enabling discourse-depth measurements.Edges propagate head information from nuclei to satellite EDUs; an EDU’s depth counts its embeddings as a satellite.
- Depth weighting: Each EDU’s sentiment contribution is weighted by a linear function of its discourse depth, with weights learned or specified by a sentiment lexicon.The prediction for each EDU is θ⊤w_i, where w_i is its bag-of-words vector and θ contains learned or lexicon-specified word weights.
- Results: 4−5% raw improvements over bag-of-words were obtained for lexicon-based sentiment analysis with discourse-depth weighting.The approach was evaluated with both lexicon-based and classification-based sentiment analysis.
- Results: Less than 1% improvement was obtained for classification-based models on both datasets.The passage contrasts this smaller gain with the larger lexicon-based improvement.
4 Rhetorical Recursive Neural Networks
R2N2 recursively composes sentiment over RST discourse trees, using scalar representations and relation-sensitive parameters. It improves classification accuracy, although imperfect discourse-relation detection limits the gains.
- Model: R2N2 recursively propagates sentiment scores from elementary discourse units to the root of the RST tree.The model uses a recursive neural-network architecture for combining discourse units.
- Model: Each elementary discourse unit starts with a bag-of-words sentiment score, while parameters are reused across documents with different tree structures.The architecture varies by document structure, but its learned parameters are shared across instances.
- Model: R2N2 uses scalar discourse-unit scores and composition parameters rather than latent distributed representations, because discourse units typically contain multiple words.This distinguishes it from syntactic recursive models that compose sentence constituents.
- Results: More than 3% accuracy improvement occurs on the Stanford data and 0.5% on the Pang & Lee dataset without distinguishing discourse relations.Distinguishing contrastive from non-contrastive relations further improves Pang & Lee performance, exceeding baseline classifier D2 by 1.3%.
- Limitations: Only 60% accuracy in discourse-relation detection may help explain why relation-sensitive modeling provides a limited boost.The paper notes that evaluating with gold RST trees would require a larger corpus containing both gold parses and sentiment labels.
5 Related Work
Related work has incorporated discourse structure into sentiment analysis through intra-sentential RST relations, explicit connectives, and shallow discourse relations. Direct comparison remains limited because prior studies use different datasets, filtering criteria, or evaluation units.
- Discourse-based sentiment analysis: Earlier RST-based sentiment studies often focus on intra-sentential relations rather than discourse relations spanning entire documents.Other work addresses Chinese sentiment or uses explicit connectives when a discourse parser is unavailable.
- Alternative discourse representations: PDTB relations provide an alternative shallow discourse representation and have been used as features, posterior constraints, or components of sentiment-flow models.The passage contrasts these connective-oriented relations with the paper’s RST-based document structure.
- Comparability: Direct comparison with prior work is difficult because studies differ in excluded documents, sentence-versus-document evaluation, and whether they use lexicon-based sentiment analysis.The only identified direct comparison evaluates a similar weighting scheme on Pang and Lee data using lexicon-based analysis.
6 Conclusion
The paper argues that RST structure can improve document-level sentiment analysis beyond preponderance-of-evidence methods. Simple reweighting helps lexicon-based analysis, while a recursive neural network can outperform a bag-of-words classifier.
- Conclusion: RST discourse parsers make it practical to incorporate document-level structure into sentiment polarity analysis.The paper frames this as an alternative to relying on words or sentences expressing overall polarity to outweigh counterpoints and concessions.
- Conclusion: Simple discourse reweighting offers robust advantages for lexicon-based sentiment analysis.The conclusion presents this as a practical benefit of combining discourse structure with sentiment analysis.
- Conclusion: A recursive neural network can substantially outperform a bag-of-words classifier.The paper identifies richer sentence-level models combined with discourse structure as future work.