TL;DR
This paper introduces DiMLex, a novel lexicon designed to capture the rich semantic, syntactic, and pragmatic nuances of discourse markers. By providing a structured representation of these "glue words," DiMLex aims to significantly improve both the generation of more natural and coherent text and the understanding of complex textual relationships.
What Problem Does This Paper Address?
Discourse markers, like "but," "however," and "because," are crucial for signaling the relationships between different parts of a text. While they are essential for human communication, current Natural Language Processing (NLP) systems often struggle to use them effectively. This is partly because discourse markers are a heterogeneous group of words, traditionally treated as grammatical function words rather than having distinct lexical meanings. Consequently, NLP systems tend to select them as an afterthought or based on simplistic rules, leading to unnatural or even incorrect text generation. Furthermore, understanding systems often fail to fully leverage the subtle implications these markers convey. The paper addresses this gap by proposing a dedicated, structured lexicon (DiMLex) to systematically represent the multifaceted nature of discourse markers.
Methods
The authors propose building DiMLex by drawing upon linguistic theories and existing research. Their methodology involves:
- Identifying Discourse Markers: They acknowledge that traditional part-of-speech tagging is insufficient and propose a more nuanced approach, considering "relational phrases" and consulting grammars and linguistic literature.
- Classifying Lexical Features: DiMLex aims to capture a range of features, including:
- Syntactic: Part of speech, positioning flexibility, and constraints on linear order.
- Semantic: The core discourse relation expressed (e.g., contrast, concession, causality) and polarity (e.g., "if" vs. "unless").
- Pragmatic: The speaker's intention, presuppositions associated with the marker, and illocutionary force.
- Stylistic: Differences in formality and emphasis.
- Representing Relationships: DiMLex will use an inheritance-based formalism to compactly represent individual marker descriptions and their hyponymic relationships, allowing for the representation of polysemy (a single marker having multiple meanings).
- Top-Down and Bottom-Up Integration: The lexicon integrates insights from text linguistics (top-down, focusing on coherence relations) and grammar/linguistic research (bottom-up, detailing individual marker properties).
Datasets
Not discussed in the paper. The paper focuses on the conceptualization and methodology for building the lexicon, rather than its application to specific datasets for training or evaluation.
Key Insights and Ideas
- Discourse Markers are Lexical, Not Just Grammatical: The paper argues strongly that discourse markers possess rich semantic and pragmatic meanings that warrant their treatment as lexical items, not mere grammatical glue.
- Nuance is Key: The choice of a discourse marker is often meaning-driven and context-dependent, with subtle differences between seemingly similar markers (e.g., "aber" vs. "sondern" in German) carrying significant pragmatic weight.
- A Dedicated Lexicon is Necessary: A unified, structured lexicon like DiMLex is essential for NLP systems to move beyond simplistic selection rules and generate truly natural text.
- Dual Application: DiMLex is designed to benefit both text generation (by providing options for appropriate marker selection) and text understanding (by helping to infer discourse relations and pragmatic implications).
Key Results
The paper does not present quantitative results from a deployed system. Instead, it outlines the conceptual framework and methodology for DiMLex. The "results" are the proposed structure and classification scheme for discourse markers, which the authors believe will enable more sophisticated NLP applications. They describe a generation testbed that allows for exploring the role of DiMLex in producing paragraph-size text, with a goal of producing German and English text by mid-1999.
Limitations and Future Work
- Initial Focus: The current theoretical focus is on German CONTRAST and CONCESSION markers.
- Scope Expansion: Future work includes enlarging DiMLex to cover other groups of markers, developing systematic lexical representations, and giving more attention to text understanding requirements.
- Formalism Development: The authors are still developing the precise formalism for representing the information within DiMLex.
- Text Understanding Integration: While the paper outlines potential applications for text understanding, the specific implementation details and evaluation are left for future work.
Summary updated