Comparatives in Context

Steffen Staab, Udo Hahn

Abstract

Natural language processing of comparative sentences (e.g., "X is faster than Y") often falters when faced with the irregularities of real-world text, such as metonymy or omitted complements. This paper proposes a novel model based on "semantic copying" that integrates world knowledge and discourse context to resolve these complexities. By prioritizing conceptual similarity over rigid syntax, the system achieves high accuracy in interpreting both explicit and implicit comparisons.

The Challenge of Comparatives

In natural language processing, comparatives are surprisingly tricky. While a human easily understands that "The Oki-X11 is faster than the HP" implies a comparison between two printers, a computer often struggles. Traditional approaches relied heavily on syntax—the formal structure of the sentence. However, real-world text is messy. We often omit parts of the comparison ("This printer is faster [than the other one]") or use metonymy, where one concept stands in for another (using "HP" to mean "the printer manufactured by HP").

Table 1: Examples of various comparative structures handled by the model.

Syntax-oriented models fail because they lack the semantic and conceptual "glue" needed to connect the dots when the text doesn't explicitly spell everything out. The variety of comparative forms—ranging from predicative to metonymic—requires a more flexible approach.

Methodology: Semantic Copying

The authors introduce a multi-layered interpretation process based on Semantic Copying. Instead of just looking at sentence structure, the model follows a four-step process:

  1. Quasi-Coordination: The system identifies potential pairs of objects to be contrasted. If syntax fails to find a match, it looks for "conceptually plausible" pairs using a description logic framework.
  2. Conceptual Filtering: A predicate called PlausiblePair ensures the comparison makes sense. For example, comparing a "Person" to a "Time Interval" is flagged as implausible, whereas comparing two "Printers" is highly similar and preferred.
  3. Semantic Reconstruction: Once a pair is found (e.g., Printer A and Printer B), the model "copies" the semantic relations from the main part of the sentence (the matrix clause) to the comparison part (the complement). If the sentence says Printer A has a thickness of X, the model reconstructs that Printer B has a thickness of Y.
  4. Contextual Integration: To handle omitted complements, the authors integrate Centering Theory. This allows the model to look back at previous sentences to find the most likely "center" of discourse to fill in the missing piece of the comparison.

Empirical Evaluation

The model was evaluated using a corpus of 32 real-world texts from the information technology (IT) domain, totaling approximately 30,000 words. This dataset contained 160 instances of comparatives, providing a realistic benchmark for technical documentation and product descriptions.

Table 2: Quantitative performance of the semantic copying model across different comparative types.

The results demonstrate the effectiveness of the semantic approach:

  • Explicit Comparatives: For sentences where the complement was fully present, the model achieved a 91% correct analysis rate.
  • Omitted Complements: The model successfully resolved 48% of cases where the comparison was implicit. While lower than explicit cases, this represents a significant advancement, as previous models typically could not handle these cases at all.
  • Metonymy Resolution: The inclusion of world knowledge (e.g., knowing HP makes printers) added a roughly 10% "bonus" to the overall coverage.

Discussion and Future Work

The "aha moment" of this research is that conceptual similarity is often a better guide for resolution than grammatical rules. By defining a "Similarity Ordering," the model can prioritize the most logical comparison even when the sentence structure is irregular. Addressing omitted complements is essential, as they accounted for 61% of the comparatives in the study.

However, limitations remain. About 25% of omitted complements are based on pragmatic inferences rather than anaphoric references (e.g., "With this device, you store data more efficiently" implies a comparison to not using the device). Future work will explore deeper pragmatic reasoning and the interaction between comparative interpretation and broader quantifier scope.

Summary updated

BibTeX
@inproceedings{staab1997comparatives,
  author = {Steffen Staab and Udo Hahn},
  title = {{Comparatives in Context}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1997},
}
Figure