TL;DR
This paper introduces SLSeg, a novel discourse segmentation system that uses syntactic and lexical information to identify elementary discourse units (EDUs) in text. By adopting a more conservative approach than existing methods, SLSeg achieves higher precision, reducing over-segmentation while maintaining a strong F-score, making it a robust tool for downstream discourse parsing tasks.
What Problem Does This Paper Address?
Discourse parsing aims to understand the structure and relationships between different parts of a text, much like how we understand sentences. A crucial first step in this process is discourse segmentation, which involves breaking down a text into its smallest meaningful units, called elementary discourse units (EDUs). These EDUs can be entire sentences or even clauses within a sentence. The challenge is that existing segmentation methods often "over-segment" text, creating too many small units that don't accurately reflect how humans naturally group ideas. This over-segmentation can lead to errors in the subsequent discourse parsing stage. The authors tackle this by developing a segmenter that prioritizes accuracy and avoids creating unnecessary boundaries, especially in informal texts where segmentation can be more complex.
Methods
The core of SLSeg lies in its syntactic and lexical-based approach. It starts by leveraging the output of a standard syntactic parser to identify potential segment boundaries based on sentence structure. Specifically, it focuses on identifying adjunct clauses (clauses that add extra information but are not essential to the main clause's meaning) as potential EDUs, while generally excluding complement clauses (like subject or object clauses) which are often too fine-grained for discourse purposes.
Key aspects of their method include:
- Syntactic Rules: A set of 12 rules are applied based on the syntactic parse tree. These rules help identify and preserve the integrity of certain clause structures.
- Lexical Refinement: After syntactic segmentation, lexical information is used to further refine boundaries. This includes considering multi-word expressions and specific "discourse cue phrases" (like "in order to") that signal relationships between segments.
- Handling Idiomatic Structures: The system is designed to avoid breaking segments at idiomatic phrases that might appear to be independent clauses but function as a single unit (e.g., "as it stands").
- Conservative Boundary Insertion: A key principle is to insert boundaries only when a discourse relation could plausibly hold between the resulting segments. This conservative approach aims to minimize incorrect boundaries.
Datasets
The paper evaluates SLSeg on a gold standard test set consisting of 9 human-annotated texts. This set is diverse, including:
- 3 texts from the RST literature.
- 3 online product reviews from Epinions.com (representing informal text).
- 3 Wall Street Journal articles from the Penn Treebank (representing formal text).
The total dataset comprises 191 sentences and 340 discourse segments. The authors also mention using the RST Discourse Treebank (RST-DT) for comparison with other systems.
Key Insights and Ideas
- Precision over Recall for Robustness: The paper strongly advocates for a conservative segmentation strategy that prioritizes precision (correctly identified boundaries) over recall (finding all possible boundaries). This approach, while potentially missing some valid boundaries, significantly reduces errors and leads to more reliable input for discourse parsers.
- Syntactic Structure is Key, but Lexical Cues Matter: While syntactic parsing provides the foundational structure, lexical information (like discourse markers and idiomatic phrases) is crucial for fine-tuning segment boundaries and ensuring they align with semantic coherence.
- Informal Text Segmentation is Challenging but Achievable: The system demonstrates comparable performance on both formal and informal texts, suggesting that its principled approach can handle the variability of language found in online reviews.
- Existing Parsers are Too Fine-Grained: The authors highlight that standard syntactic parsers, while useful, produce segments that are too granular for discourse purposes, necessitating a post-processing step like SLSeg.
Key Results
The paper presents a comparison of SLSeg against several baseline and state-of-the-art segmenters in Table 1.
- High Precision: SLSeg consistently achieves high precision across different datasets and parser combinations. For instance, using the Charniak parser on combined formal and informal texts, SLSeg achieves high precision.
- Competitive F-scores: While recall is sometimes lower than other systems, SLSeg maintains a high F-score, indicating a good balance between precision and recall. For example, on the combined dataset, SLSeg (Stanford) achieves an F-score of .77, and SLSeg (Charniak) achieves .80.
- Parser Independence: The performance of SLSeg is shown to be relatively independent of the underlying syntactic parser used (comparing Stanford and Charniak parsers), suggesting the robustness of their segmentation rules.
- Consistency with RST: SLSeg's approach is shown to be more consistent with the original RST guidelines compared to other systems, particularly when evaluated on RST literature texts.
Limitations and Future Work
- Recall Trade-off: The authors acknowledge that their conservative approach leads to lower recall, meaning some valid discourse boundaries might be missed.
- Future Work:
- Broader Corpora: Investigating the robustness of SLSeg on more diverse corpora, including formal texts from specialized domains (like medical texts) and other types of informal texts.
- Quantitative Study of Segmentation Strategies: Further research into the impact of high-precision/low-recall versus low-precision/high-recall segmenters on the downstream task of discourse tree construction.
- Assisting Manual Annotation: The system could be adapted to assist human annotators by providing an initial set of discourse segments for manual annotation of discourse relations.
Summary updated