Extension et couplage de ressources syntaxiques et sémantiques sur les adverbes du français

Benoît Sagot, Karën Fort, Fabienne Venant

TL;DR

This paper tackles the challenge of limited semantic resources for French by extending the WOLF wordnet with adverbial information and coupling it with the syntactic Lefff lexicon. The authors leverage derivation and synonymy relations to significantly increase the coverage of French adverbs in WOLF, making a valuable semantic resource more comprehensive and useful for NLP applications.

What Problem Does This Paper Address?

The field of Natural Language Processing (NLP) heavily relies on comprehensive lexical resources, including syntactic and semantic information about words. While resources for English are relatively abundant, French has historically lagged behind, particularly in the domain of semantic resources. This paper focuses on a specific, yet crucial, part of speech: adverbs. The authors highlight that while morphological and syntactic resources for French are improving, high-quality semantic resources remain scarce. Specifically, they aim to address the limited coverage of adverbs in the French wordnet, WOLF, and to bridge the gap between syntactic and semantic information for these words. This is essential for building more sophisticated NLP tools that can understand and generate language with greater nuance.

Methods

The authors employ two primary techniques to extend the WOLF wordnet's adverbial entries:

  • Extension by Derivation: This method exploits the morphological and semantic relationship between adjectives and adverbs formed by adding the suffix "-ment" in French (analogous to "-ly" in English). They identify French adjectives linked to adverbial synsets in WOLF and apply the derivation rule to generate potential adverbs. These generated adverbs are then checked against the Lefff lexicon, and if found, are added to the corresponding WOLF synset. This technique leverages existing syntactic information to enrich semantic data.
  • Extension by Synonymy: This approach utilizes the DicoSyn database, a synonym dictionary for French, to identify groups of synonymous adverbs (cliques). They first assign weights to existing lexeme-synset links in WOLF based on their origin (e.g., bilingual resources, multilingual corpora). Then, for each WOLF adverbial synset, they find the DicoSyn clique that maximizes the sum of weights of common lexemes. Finally, adverbs from these selected cliques are added to the WOLF synset, further expanding its coverage.

Datasets

  • WOLF (WOrdnet Libre du Français): A freely available semantic lexical resource for French, structured as a wordnet. The paper focuses on extending its adverbial entries.
  • Lefff (Lexique des Formes Fléchies du Français): A broad-coverage morphological and syntactic lexicon for French. It serves as a source of French adverbs and is used to validate derived adverbs.
  • DicoSyn: An electronic synonym dictionary for French, used to extract synonymy relations and form cliques of synonymous adverbs.
  • Lexicon-Grammar Tables of "-ment" Adverbs (Molinier & Levrier, 2000): These tables, primarily used to construct Lefff's adverbial entries, provide syntactic-semantic classes for adverbs ending in "-ment."

Key Insights and Ideas

  • Leveraging Existing Resources is Key: The paper demonstrates the power of combining different types of lexical resources (syntactic, semantic, synonymy databases) to overcome the limitations of individual resources.
  • Derivation as a Bridge: The morphological relationship between adjectives and adverbs is a surprisingly effective way to automatically expand semantic coverage, especially for a productive class like "-ment" adverbs.
  • Synonymy as a Semantic Anchor: Using synonymy cliques from DicoSyn provides a robust way to group semantically related adverbs and link them to existing synsets, even if they weren't directly derivable.
  • The Importance of Manual Validation: Despite sophisticated automatic methods, the paper underscores the necessity of manual evaluation to ensure the quality and correctness of the extended semantic resource.

Key Results

The paper reports significant improvements in the WOLF adverbial resource:

  • Increased Lexeme-Synset Relations: The number of adverbial lexeme-synset relations in WOLF increased from 983 to 1,536 (a 56% augmentation) after extension by derivation, and further to 2,149 (a 28.5% increase) after extension by synonymy.
  • Increased Non-Empty Synsets: The number of adverbial synsets with at least one known French lexeme rose from 676 to 969 (a 43% augmentation).
  • Increased Lexemes: The total number of adverbial lexemes in WOLF grew from 665 to 889 (a 23% augmentation).
  • Manual Validation Success: Over 68% of the automatically generated lexeme-synset associations were deemed correct (OK) during manual evaluation. This resulted in 1,461 validated relations, a substantial increase from the initial 983.

Limitations and Future Work

  • Inter-Validator Agreement: The manual evaluation revealed a notable level of disagreement among the three annotators, particularly in distinguishing between closely related semantic codes (SC, SR, NR). This highlights the inherent subjectivity in semantic annotation and the difficulty of defining precise boundaries between categories.
  • Coupling with Lefff: While the paper describes the process of extending WOLF and the potential for coupling with Lefff, the actual manual validation of these couplings is planned for the near future.
  • Future Directions: The authors suggest exploring other lexical relations, such as hyperonymy and antonymy, and extending their approach to other parts of speech (nouns, verbs) and other types of relations (syntagmatic relations via corpus analysis). They also envision further collaboration between different lexical resources to enrich the French NLP landscape.

Summary updated

Abstract

This paper presents a work on extending the adverbial entries of the WOLF, a semantic lexical resource for French, and connecting them with those of the syntactic lexicon Lefff, which were mostly extracted from the lexicon-grammar tables from (Molinier & Levrier, 2000). This work relies on the exploitation of the derivation and the synonyms relations ; the latter are extracted from the DicoSyn synonyms database. The resulting semantic resource, which is freely available, is manually evaluated and validated in an exhaustive manner.
BibTeX
@inproceedings{sagot2009extension,
  author = {Benoît Sagot and Karën Fort and Fabienne Venant},
  title = {{Extension et couplage de ressources syntaxiques et sémantiques sur les adverbes du français}},
  booktitle = {Lingvisticae},
  year = {2009},
  doi = {10.1075/li.32.2.12sag},
}
Figure