Abstract
Natural language processing often struggles to balance the rigidity of symbolic rules with the fluidity of statistical probabilities. This paper presents a hybrid architecture that bridges this gap, focusing on the interpretation of compound nouns. By integrating symbolic grammar, statistical word-sense probabilities, and pragmatic reasoning, the proposed framework efficiently resolves linguistic ambiguity.
Introduction
Interpreting language is a complex balancing act between a word's typical meaning and its specific meaning within a given conversation. Traditional NLP systems often fail at the extremes: purely rule-based systems suffer from "symbolic overgeneration," producing too many valid but unlikely interpretations, while deep pragmatic reasoning is often too computationally expensive for real-time use. This paper by Ann Copestake and Alex Lascarides introduces a "Lexicon Pragmatics Interface" designed to handle these challenges, particularly in the context of productive compound nouns like "cotton bag" or "apple juice seat."
The Lexicon Pragmatics Interface
The core of the proposed system is a three-layered architecture that combines symbolic, statistical, and pragmatic components.
Symbolic Schemata
The system utilizes a hierarchy of rules, or schemata, that define how nouns can combine. These range from specific relations like made-of (e.g., "cardboard box") to a "catch-all" general-nn relation for novel or unusual cases.
Statistical Preferences and Pragmatic Reasoning
To manage the overgeneration of symbolic rules, the authors incorporate statistical data from the British National Corpus (BNC). This data allows the system to calculate the probability of specific word senses and the "productivity" of different compound rules.
The system uses Segmented Discourse Representation Theory (SDRT) and Discourse in Commonsense Entailment (DICE) to model conversation structure and apply logic. The "Prefer Frequent Senses" rule acts as a search heuristic: the system first attempts the most statistically likely interpretation. If this interpretation is contextually coherent, the search concludes; otherwise, it moves to the next most likely sense.
Key Insights
- Probabilities as Search Heuristics: Statistics are used to guide logic rather than replace it. This significantly reduces the computational burden on the pragmatic module by prioritizing likely interpretations.
- Coherence Trumps Frequency: The system remains flexible. If a common meaning (e.g., "diet" as food) is nonsensical in context (e.g., referring to a parliament), the system successfully falls back to rarer senses.
- Productivity Matters: The likelihood of novel compounds can be estimated by the productivity of the underlying rule. For instance, the
made-ofrule is highly productive, making such interpretations statistically "cheaper" to verify.
Results and Discussion
The framework provides a robust explanation for both standard ambiguity resolution and the interpretation of novel compounds. For example, it correctly identifies "made of cotton" for "cotton bag" in neutral contexts but can switch to a "purpose" interpretation if the discourse context demands it. Furthermore, it handles highly unusual compounds like "apple juice seat" by allowing pragmatics to "fill in the blanks" when the discourse provides a specific bridge.
Conclusion
This research demonstrates that integrating symbolic and statistical representations provides a powerful tool for NLP. By using probabilities to hint at meanings and pragmatic coherence to validate them, the Lexicon Pragmatics Interface offers a scalable and theoretically sound approach to language interpretation. Future work will focus on the large-scale acquisition of frequency data and further empirical validation of the mathematical models for novel compounds.
Summary updated