Abstract
In this paper we introduce an empirical approach to the semantic interpretation of superlative adjectives. We present a corpus annotated for superlatives and propose an interpretation algorithm that uses a wide-coverage parser and produces semantic representations. We achieve F-scores between 0.84 and 0.91 for detecting attributive superlatives and an accuracy in the range of 0.69--0.84 for determining the correct comparison set. As far as we are aware, this is the first automated approach to superlatives for open-domain texts and questions.
Introduction
Superlatives like "tallest" or "most complex" are ubiquitous in natural language but present significant challenges for automated semantic interpretation. Understanding a phrase such as "the oldest building in Rome" requires a system to identify the target entity (a building), the property being compared (age), and the comparison set (all buildings in Rome).
Historically, formal linguistics provided theoretical models for these structures, often focusing on complex, artificial examples. Conversely, mainstream NLP systems frequently treated superlatives as simple modifiers, ignoring their internal logic. This research bridges the gap by introducing a system capable of handling the diverse superlatives found in real-world news articles and search engine queries.
Methodology
The researchers developed a "Deep Linguistic Analysis" (DLA) pipeline that transitions from raw text to formal semantic representations.
Syntactic Parsing
The first stage utilizes a wide-coverage Combinatory Categorial Grammar (CCG) parser. CCG is particularly effective for this task because it offers a transparent interface between syntactic structure and semantic meaning. The parser handles various superlative constructions, including those with modifiers and ordinal numbers.
Semantic Representation
The syntactic output is mapped into Discourse Representation Theory (DRT). This framework allows the system to explicitly represent the comparison set as a logical condition. By using lambda calculus, the system builds complex meanings from individual word components.
Refinement Rules
Standard parsers often struggle with the "scope" of superlatives. To address this, the authors implemented post-processing rules for specific cases:
- Possessives: Ensuring "Jaguar’s largest shareholder" correctly identifies the comparison set as shareholders of Jaguar.
- NP Modifiers: Ensuring "the largest lab in New England" compares labs within New England, rather than comparing all labs globally.
Data and Annotation
The study introduced a newly annotated corpus of 3,045 superlatives from diverse sources, including the Wall Street Journal (WSJ), Glasgow Herald, TREC QA questions, and Excite search logs. Instances were categorized into four types: attributive (89.1%), predicative, adverbial, and idiomatic.
The high prevalence of attributive superlatives in real-world data justified the focus on this category. The annotation process revealed that while detecting the superlative word itself is straightforward, determining the "comparison set"—the syntactic neighborhood that defines the group being compared—is the primary semantic challenge. Interestingly, "comparative readings" (e.g., "John climbed the highest mountain" meaning higher than others, not the tallest overall) were found to be extremely rare in news text despite their prominence in theoretical linguistics.
Results
The system was evaluated on two primary tasks:
- Superlative Detection: The DLA system achieved F-scores between 0.84 and 0.91, significantly outperforming baseline methods using simple Part-of-Speech (POS) tags.
- Comparison Set Determination: This core semantic challenge saw the most advanced version of the system (DLA 4) achieve an accuracy of 0.78–0.83 on WSJ data and 0.82 on TREC questions, demonstrating its utility for Question Answering.
Conclusion
This work represents the first automated, corpus-grounded approach to interpreting superlatives in open-domain text. While the system performs well on grammatical text, performance drops on ungrammatical web queries where the underlying CCG parser struggles. Future work will aim to address the "dimension" of comparison and extend the analysis to predicative and adverbial forms.
Summary updated