TL;DR
Discourse analysis, crucial for understanding text beyond sentence boundaries, is hampered by a Babel of annotation frameworks (RST, PDTB, SDRT). This paper proposes a unified framework to bridge these gaps, enabling better data utilization and model synergy. The core contribution lies in a systematic approach to unifying both discourse structure and relations, aiming for a more comprehensive and computationally tractable discourse representation.
What Problem Does This Paper Address?
Natural language processing (NLP) models are increasingly moving beyond sentence-level analysis to capture the richer meaning conveyed by discourse. However, the field is fragmented by numerous discourse annotation frameworks, each with its own theoretical underpinnings and annotation conventions (e.g., Rhetorical Structure Theory (RST), Penn Discourse Treebank (PDTB), Segmented Discourse Representation Theory (SDRT)). This diversity leads to considerable variation in annotated corpora, making it difficult to combine data from different sources or leverage insights across frameworks. While previous work has attempted to map or unify specific aspects like discourse relations or structures, no comprehensive approach has addressed both simultaneously and evaluated it rigorously. This paper aims to fill that gap by proposing a unified framework that can integrate these disparate annotation schemes, paving the way for more robust discourse understanding in NLP.
Methods
The paper outlines a research plan to develop a unified discourse annotation framework. The core methodological approach involves:
- Choosing a Unified Structure: Investigating dependency structures as a common ground for representing discourse, drawing inspiration from syntactic parsing and previous work that converts RST and SDRT to dependency formats.
- Identifying Key Properties: Determining which properties from existing frameworks (e.g., nuclearity in RST, sentence-boundedness in PDTB, multi-sense annotation in SDRT) are essential for unification and which can be omitted.
- Incorporating Entity and Lexical Models: Exploring how to integrate entity-based and lexical-based discourse information, which are handled differently across frameworks.
- Automatic Unification: Primarily relying on automatic means for unification, though semi-automatic methods are not precluded.
- Evaluation: Employing both intrinsic evaluation (assessing structural complexity) and extrinsic evaluation (measuring performance on downstream tasks like text summarization and document discrimination).
Datasets
The paper mentions several datasets that will be used or are relevant to the research:
- RST-DT (Rhetorical Structure Theory Discourse Treebank): Annotated texts from the Penn Treebank, formally written Wall Street Journal articles.
- PDTB 3.0 (Penn Discourse Treebank): Also annotated with texts from the Penn Treebank.
- ANNODIS corpus: Contains formally written French texts, which will be translated to English.
- STAC corpus and Molweni corpus: These are mentioned as corpora annotated under the SDRT framework, used for analyzing multi-party dialogues, but are noted as difficult to use with the other two corpora.
- CNN and Dailymail dataset: To be used for text summarization tasks.
- SciDTB corpus: Annotated under the dependency framework, to be used for empirical investigation.
The paper does not introduce new datasets but aims to unify existing ones.
Key Insights and Ideas
- The "Babel" of Discourse Annotation: The fundamental problem is the lack of a common language for discourse, hindering cross-corpus analysis and model development.
- Dependency Structure as a Unifier: Dependency structures offer a promising common ground for representing discourse, bridging the gap between different formalisms.
- Balancing Properties: Unification requires careful selection of which properties from existing frameworks to retain, balancing expressiveness with tractability.
- Multi-task Learning for Synergy: The proposed framework could enable multi-task learning approaches to leverage information across different discourse corpora, boosting performance.
- Bridging Structure and Relations: A key challenge and goal is to unify both discourse structure and discourse relations, a gap not fully addressed by prior work.
Key Results
This paper is a research proposal and does not present experimental results. The "Key Results" section is not applicable.
Limitations and Future Work
The paper acknowledges several limitations and outlines future work:
- Complexity of Unification: Determining the optimal unified structure, selecting which properties to keep, and incorporating diverse models (entity-based, lexical-based) are significant challenges.
- Ambiguity in Transformation: Transforming between structures can introduce ambiguity, particularly when converting back to original frameworks.
- Sentence-Boundedness: The varying treatment of sentence boundaries across frameworks complicates unification.
- Multi-sense Annotation: While SDRT and PDTB allow multi-sense annotation, RST does not, posing a challenge for a unified approach.
- Future Work: The authors plan to:
- Investigate dependency structures and constrained DAGs as unified representations.
- Explore methods for maintaining one-to-one correspondence during structure transformation.
- Empirically evaluate the suitability of the dependency structure.
- Address the challenges of incorporating entity-based and lexical-based models.
- Conduct intrinsic and extrinsic evaluations of the unified framework.
- Re-investigate previous multi-task learning and GNN studies with the unified framework.
- Examine what might be lost by ignoring certain properties during unification.
Summary updated