TL;DR
This paper proposes a novel framework that bridges the gap between symbolic linguistic knowledge (lexicon and grammar) and statistical information (word frequencies and probabilities) to resolve ambiguity in compound noun interpretation. By integrating these two approaches, the authors aim to create a more robust and flexible system for understanding language, particularly for novel or less common word uses.
What Problem Does This Paper Address?
Interpreting the meaning of compound nouns (like "apple juice seat" or "cotton bag") is a significant challenge in natural language processing (NLP). Traditional approaches often fall into two camps: purely symbolic (lexicon/grammar-based) or purely statistical (frequency-based). Symbolic methods struggle with the vast productivity of language and the fact that many compound meanings are context-dependent, not strictly defined by rules. Statistical methods, while good at capturing common patterns, can fail when encountering novel or less frequent word combinations. This paper addresses the need for a unified approach that leverages both the structured knowledge of language and the probabilistic nature of its use to accurately disambiguate compound meanings, especially in discourse contexts.
Methods
The authors propose an architecture that integrates three key components:
- Symbolic Lexicon/Grammar: This provides a structured representation of word meanings and grammatical rules, including a hierarchy of "schemata" for compound nouns. These schemata define conventional interpretations (e.g., "made-of," "purpose-patient").
- Word-Sense Probabilities: Statistical information, derived from corpus frequencies, is used to assign probabilities to different senses of words and compound interpretations. This helps prioritize more common meanings.
- Pragmatic Component: A discourse theory (SDRT) and its accompanying pragmatic module (DICE) are used to reason about context and coherence. This component helps select the most plausible interpretation that fits the discourse, even if it's a less frequent one.
The core idea is that lexical probabilities offer "hints" to the pragmatic module, guiding its reasoning. When a statistically preferred interpretation leads to discourse incoherence, the pragmatic component can override it in favor of a less probable but contextually appropriate meaning. They also introduce a formula for estimating the "productivity" of compound schemata, which helps in assigning probabilities to unseen compounds.
Datasets
The paper mentions using the British National Corpus (BNC) for semi-automatic analysis of word and compound frequencies. WordNet is also cited as a source for semantic categories used in probability calculations. No new datasets were introduced.
Key Insights and Ideas
- The inadequacy of purely symbolic or statistical approaches: The paper strongly argues that neither approach alone is sufficient for handling the complexities of compound interpretation.
- Probabilities as "hints" for pragmatics: Lexical probabilities are not absolute deciders but rather guide the pragmatic reasoning process, helping to prune the search space of possible interpretations.
- Discourse coherence as the ultimate arbiter: Even statistically improbable interpretations can be favored if they are necessary for maintaining discourse coherence. This is a crucial aspect of their model.
- Productivity of linguistic rules: The concept of "schema productivity" is introduced to quantify how likely a particular type of compound interpretation is, which is essential for handling novel word formations.
- A unified architecture: The proposed framework offers a principled way to integrate symbolic linguistic knowledge with statistical data and discourse-level reasoning.
Key Results
The paper doesn't present quantitative results in terms of specific accuracy scores on a benchmark dataset. Instead, it focuses on demonstrating the theoretical framework and its qualitative capabilities through examples.
- Example: "cotton bag": The paper illustrates how the framework can distinguish between "bag made of cotton" (high probability, favored in general contexts) and "bag for cotton" (lower probability, potentially favored in specific contexts).
- Example: "diet": It shows how the food sense of "diet" is preferred in isolation, but the "legislature" sense might be favored in a discourse context that makes the food sense incoherent.
- Handling novel compounds: The framework is designed to handle compounds like "apple juice seat" by relying on the general-nn schema and then using pragmatics to infer the specific relationship based on context.
Limitations and Future Work
- Computational cost of pragmatics: The authors acknowledge that pragmatic reasoning is computationally expensive, even in restricted domains, and their approach aims to circumscribe it as much as possible.
- Formula validation: Further work is needed to validate or revise the probability estimation formulae presented in Section 3.
- Compound schema development: The set of compound schemata needs further development and refinement.
- Integration of other linguistic phenomena: While the paper focuses on compound nouns, the authors suggest that a similar approach could be applied to morphologically-derived forms and words in extended usages.
- Language specificity: The pragmatic component is described as not being language-specific, but the lexical probabilities are derived from English corpora, implying a degree of language dependence in the statistical component.
Summary updated