Segmented Discourse Representation Theory (SDRT)
Standard DRT does a solid job of tracking referents and resolving anaphora across sentences, but it treats discourse as a flat sequence. Real conversations and texts have structure: some sentences elaborate on others, some explain why something happened, some contrast two ideas. SDRT was developed to capture exactly this kind of organization.
Introduction to SDRT
SDRT extends DRT by adding rhetorical relations and hierarchical discourse structure. It was developed by Nicholas Asher and Alex Lascarides in the 1990s.
Where standard DRT gives you a Discourse Representation Structure (DRS) that tracks discourse referents and conditions, SDRT introduces Segmented Discourse Representation Structures (SDRSs). An SDRS does everything a DRS does, but it also represents:
- The way discourse breaks into segments (roughly, clause- or sentence-sized units of meaning)
- The rhetorical relations that connect those segments (like Elaboration, Explanation, or Contrast)
- The hierarchical organization of segments, showing which segments are subordinate to others
This extra layer of structure helps with problems DRT already cares about, like pronoun resolution, because knowing how two segments relate tells you which referents are still accessible.
Rhetorical Relations in SDRT
Rhetorical relations are the glue that holds discourse together. They specify the semantic and pragmatic relationship between two discourse segments. In an SDRS, these relations appear as labeled edges connecting segments.
Some key rhetorical relations:
- Elaboration: The second segment provides more detail about the first. ("The party was great. The food was incredible and the music was perfect.")
- Explanation: The second segment gives a reason or cause for the first. ("John fell. The sidewalk was icy.")
- Contrast: The two segments highlight a difference. ("Mary loves hiking. John prefers staying indoors.")
- Narration: The segments describe events in temporal sequence. ("She opened the door. She walked inside.")
These relations matter beyond just labeling connections. They affect how discourse structure is built hierarchically. For instance, an Elaboration segment is subordinate to the segment it elaborates on, while Narration typically connects segments at the same level. This hierarchical organization determines which discourse referents remain accessible for anaphora and how the scope of semantic operators is resolved.
Variations of DRT
Beyond SDRT, other extensions address different limitations of standard DRT:
Underspecified DRT (UDRT) tackles semantic ambiguity. Instead of committing to one interpretation, UDRT allows a single underspecified DRS to represent multiple possible readings at once. This is especially useful for scope ambiguities. Consider "Every student read a book": does every student read the same book, or each a different one? UDRT can represent both readings without choosing between them, deferring disambiguation until more context is available.
Presuppositional DRT (PDRT) brings presuppositions into the framework. A presupposition is something a sentence takes for granted rather than asserting directly. In "John stopped smoking," the presupposition is that John used to smoke. PDRT treats presuppositions as constraints on the context of interpretation and models two key phenomena:
- Projection: how presuppositions survive when embedded under negation, conditionals, or other operators
- Accommodation: how listeners adjust the context to accept a presupposition they didn't previously share
Applications of DRT Extensions
These frameworks have practical uses, particularly in computational linguistics and AI:
- Natural Language Processing (NLP): DRT and its extensions provide formal tools for representing discourse meaning. SDRT is particularly useful for tasks like text summarization and discourse parsing, where understanding how parts of a text relate to each other is essential. Other applications include information extraction and question answering.
- Dialogue Systems: Conversation involves tracking referents across turns, resolving anaphora, and maintaining coherence as topics shift. DRT-based models help manage this conversational context, and when combined with machine learning and natural language generation, they support more interactive dialogue systems.
- Other domains: The dynamic, context-updating nature of DRT makes it relevant to formal verification, knowledge representation, and even the semantics of programming languages, anywhere that meaning unfolds incrementally and depends on prior context.