Contextual Inference in Computational Semantics

Christof Monz

Introduction

In natural language processing, a presupposition is an implicit assumption required for a sentence to be coherent. For instance, the sentence "Hank likes his wife" presupposes that Hank is married. In computational semantics, resolving these presuppositions is crucial for accurate discourse understanding. However, current methods often suffer from significant computational overhead due to redundant inference tasks. This paper presents a novel approach to streamline this process by treating context as a nested logical structure.

The Redundancy Problem in DRT

Discourse Representation Theory (DRT) is a standard framework for representing meaning in context. When a system encounters a presupposition trigger, it must decide whether to "accommodate" the information globally, locally, or at an intermediate level. This involves checking for "local informativity" (ensuring the information is not already known) and "consistency" (ensuring it does not contradict existing knowledge).

Presupposition representation in DRT: (a) unresolved presupposition $\alpha$ for 'his wife', (b) the resolved structure.

The primary bottleneck is redundancy. Traditional approaches treat global and local contexts as independent proof tasks. In a long discourse, a system might repeatedly re-verify the same global facts to resolve a single local pronoun, leading to exponential slowdowns as the discourse grows.

Contextual Inference and Accommodation

To address this, the author introduces the $\mathcal{L}^{con}$ language, which incorporates an in(K, \phi) predicate. This operator explicitly denotes that a formula $\phi$ holds within a specific context $K$. The $\tau$ function is used to transform standard linguistic representations into these context-aware formulas.

Global accommodation: the presupposed entity 'u' (wife) is moved to the outermost context.

By nesting these contexts, the system can represent the entire hierarchy of discourse information in a single logical expression. This structure allows for different levels of accommodation to be evaluated efficiently. For example, in global accommodation (shown above), the presupposed entity is promoted to the outermost context, making it accessible to all subsequent discourse.

Labeled Tableau Calculus

The core technical innovation is a labeled tableau calculus. In this system, each node in the deduction tree is tagged with a label $(i, \sigma)$, where $i$ identifies the current context and $\sigma$ tracks the set of accessible parent contexts. This mechanism enables "proof sharing": truths established in a parent context are automatically inherited by nested sub-contexts.

By compressing multiple informativity checks into a single $\mathcal{L}^{con}$ formula, the tableau prover avoids restarting the inference engine for each level of accommodation. The system processes the shared global information once, significantly reducing the total number of proof steps required for complex discourses.

Conclusion

This research provides a robust theoretical framework for more efficient contextual inference. By making logic "smarter" about nested structures, the proposed method eliminates the redundant theorem-proving that plagues traditional DRT implementations. Future work aims to extend this system to handle local consistency checks and integrate model generation for improved satisfiability testing.

Summary updated

Abstract

In this paper, an application of automated theorem proving techniques to computational semantics is considered. In order to compute the presuppositions of a natural language discourse, several inference tasks arise. Instead of treating these inferences independently of each other, we show how integrating techniques from formal approaches to context into deduction can help to compute presuppositions more efficiently. Contexts are represented as Discourse Representation Structures and the way they are nested is made explicit. In addition, a tableau calculus is present which keeps track of contextual information, and thereby allows to avoid carrying out redundant inference steps as it happens in approaches that neglect explicit nesting of contexts.
BibTeX
@inproceedings{monz1999contextual,
  author = {Christof Monz},
  title = {{Contextual Inference in Computational Semantics}},
  booktitle = {CONTEXT},
  year = {1999},
  url = {https://openreview.net/forum?id=hdaXG_AZNqfS},
}

Loading related content...

Figure