Encoding information on adjectives in a lexical-semantic net for computational applications

Antonietta Alonge, Francesca Bertagna, Nicoletta Calzolari, Adriana Roventini, Antonio ZampolliNAACL 2000

TL;DR

This paper details how the EuroWordNet framework is being extended to incorporate rich semantic information for adjectives, a previously underserved part of the lexical network. By introducing new semantic relations and refining the Top Ontology, the authors aim to create a more comprehensive lexical resource crucial for advanced natural language processing tasks.

What Problem Does This Paper Address?

Lexical-semantic networks, like Princeton WordNet (WN), represent word meanings and their relationships. While WN has extensively modeled nouns and verbs, adjectives have received less attention, often being treated as secondary or only linked as targets of relations from nouns and verbs. The EuroWordNet (EWN) project built upon WN with multilingual capabilities and enriched relations for nouns and verbs, but still lacked a deep representation for adjectives. This paper addresses the gap in representing adjective semantics by detailing modifications to the EWN framework within the Italian National Project (ItalWordNet - IWN) to encode a wider array of adjective-specific semantic relations and integrate them into a unified conceptual ontology. This is crucial for computational applications that increasingly rely on nuanced understanding of word meanings, including those expressed by adjectives.

Methods

The authors extend the EuroWordNet (EWN) framework by:

  • Introducing New Semantic Relations: Beyond standard synonymy and antonymy, they encode relations like NEAR_SYNONYMY, PERTAINS_TO (linking adjectives to nouns they describe, e.g., chemical/chemistry), IS_A_VALUE_OF (linking adjectives to abstract properties, e.g., tall/stature), INVOLVED (linking adjectives to concepts incorporated in their meaning, e.g., filoso/fili), and CAUSE (linking adjectives to the actions or states that cause them, e.g., depurative/depurare). A novel LIABLE_TO relation is introduced for adjectives indicating the possibility of events.
  • Revising the Top Ontology (TO): The existing TO, which categorizes concepts into "orders" (1st for concrete entities, 2nd for events/states/properties, 3rd for propositions), is adapted. Specifically, the "PROPERTY" and "RELATION" nodes are moved under "SITUATION COMPONENT" to better accommodate adjectives that express properties or relations, and to allow for dynamic/static distinctions. New top concepts like "ATTRIBUTE" and "FUNCTIONAL" are created under "PROPERTY" to distinguish between descriptive adjectives (referent-modifiers) and reference-modifying adjectives.
  • Leveraging Machine-Readable Dictionaries and Theoretical Work: The encoding of relations and ontology modifications are informed by linguistic theories, EAGLES recommendations, and data extracted from machine-readable dictionaries.
  • Diagnostic Frames and Substitution Tests: These established methods are used to verify the existence and consistency of encoded relations between synsets.

Datasets

The paper discusses the use of:

  • Princeton WordNet (WN): Serves as the foundational lexical-semantic network.
  • EuroWordNet (EWN): The multilingual framework upon which ItalWordNet is built.
  • Machine-Readable Dictionaries: Used as a primary source for extracting adjective semantic information and identifying new relations.
  • ItalWordNet (IWN): The Italian National Project's large wordnet, which is being extended with this new adjective information.

No new benchmark datasets are explicitly introduced, but the paper contributes to the development of a richer lexical resource (IWN) that can serve as a dataset for future NLP tasks.

Key Insights and Ideas

  • Adjectives are More Than Just Modifiers: The paper highlights that adjectives carry substantial semantic weight beyond simply describing nouns. New relations like CAUSE and INVOLVED reveal deeper conceptual links.
  • The Need for a Granular Ontology: The revision of the Top Ontology, particularly the introduction of "ATTRIBUTE" and "FUNCTIONAL" under "PROPERTY," is crucial for distinguishing between different types of adjectives and their semantic roles, especially for reference-modifying adjectives.
  • Cross-POS Relations are Key: The ability to link adjectives to nouns and verbs through relations like PERTAINS_TO and CAUSE is vital for capturing how adjectives interact with other word classes and for disambiguation.
  • Ambiguity Resolution through Rich Semantics: The example of "depresso" demonstrates how encoding multiple senses with distinct relations and TO links can effectively disambiguate adjective usage in context.
  • Bridging the Gap for Computational Applications: The ultimate goal is to provide a resource that computational linguists can leverage for tasks like information retrieval, question answering, and machine translation, where nuanced adjective understanding is often critical.

Key Results

The paper details the design and implementation of a richer semantic representation for adjectives within the ItalWordNet framework. While quantitative results (e.g., accuracy scores) are not presented, the key outcome is the establishment of:

  • A comprehensive set of new lexical-semantic relations specifically for adjectives.
  • A revised Top Ontology that better accommodates adjective semantics and cross-POS relationships.
  • A methodology for encoding adjective information that can be applied to other languages and lexical resources.

The paper provides illustrative examples of how these new relations and ontological structures can capture complex adjective meanings and aid in disambiguation.

Limitations and Future Work

  • Focus on Italian: While the framework is designed for multilingualism, the detailed examples and development are presented within the context of the Italian National Project (IWN).
  • Adverb Encoding: The paper notes that only adverbs derived from adjectives via the "-mente" suffix are being encoded, implying a limited scope for other adverb types.
  • Completeness of Relations: The authors acknowledge that many adjectives may not have easily definable antonyms or relations like PERTAINS_TO or IS_A_VALUE_OF, necessitating the new relations. However, the exhaustive coverage of all possible adjective relations is an ongoing challenge.
  • Future Work: The authors suggest that the encoded information will be highly useful for computational applications, particularly in resolving ambiguities and drawing inferences. The inclusion of this detailed semantic information in a large database like IWN is intended to facilitate these applications. Further work will likely involve applying this framework to other languages and evaluating its impact on NLP tasks.

Summary updated

BibTeX
@inproceedings{alonge2000encoding,
  author = {Antonietta Alonge and Francesca Bertagna and Nicoletta Calzolari and Adriana Roventini and Antonio Zampolli},
  title = {{Encoding information on adjectives in a lexical-semantic net for computational applications}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2000},
}
Figure