Construction d'un lexique des adjectifs dénominaux (Construction of a lexicon of denominal adjectives)

Jana Strnadová, Benoît Sagot

TL;DR

Researchers have developed a new, freely available lexicon for French denominal adjectives (adjectives derived from nouns, like artisanal from artisan). By combining manually defined morphological rules with large-scale corpus validation, they successfully extracted over 12,029 noun-adjective pairs, significantly boosting the resources available for French Natural Language Processing (NLP).

What Problem Does This Paper Address?

In NLP, understanding how words are constructed—morphology—is vital for tasks like machine translation, information retrieval, and handling "out-of-vocabulary" words (neologisms or technical terms). While French already had robust resources for nouns derived from verbs (deverbal nouns), it lacked a comprehensive, open-source lexicon for adjectives derived from nouns (denominal adjectives).

The challenge lies in the complexity of French derivation. It isn't just about adding a suffix; it often involves "sandhi" phenomena (spelling changes at the word boundary, like muscle becoming muscul-aire) or "suppletion," where the adjective is based on a Latin or Greek root entirely different from the modern French noun (e.g., oiseau vs. aviaire). This paper focuses on automating the capture of the regular and semi-regular patterns.

Methods

The authors employed a hybrid approach combining expert linguistic knowledge with "big data" validation:

  1. Rule-Based Generation: Using the Alexina framework, the authors manually wrote 596 morphological rules defining how suffixes (like -al, -esque, -eux) attach to nouns.
  2. Candidate Production: They applied these rules to over 65,000 nouns, generating nearly 887,000 potential noun-adjective pairs.
  3. Corpus-Based Validation: To separate legitimate words from "hallucinated" ones, they checked candidates against massive text corpora using an iterative scoring function.
  4. Noise Filtering: They used the sxspell tool to filter out OCR errors and applied length constraints to remove short, ambiguous strings.
Filtering process and candidate retention statistics.

Datasets

The study leveraged several major French linguistic resources:

  • Lexicons: Lefff (Lexique des Formes Fléchies du Français) and Morphalou provided the base nouns and known adjectives.
  • Corpora: Est Républicain (newspaper), frwiki (Wikipedia), and Google 1-grams (scanned books).
  • Evaluation Benchmark: A subset of Wiktionary was used to create a "gold standard" list of 1,433 noun-adjective pairs to test the system's coverage.

Key Insights and Ideas

  • The "Suppletion" Gap: Nearly half of denominal adjectives in some categories use suppletive (Latin/Greek) bases, highlighting a ceiling for purely rule-based systems.
  • Iterative Discovery: Their scoring method specifically targets "unknown" words by penalizing candidates already in the lexicon, forcing the system to find new, valid adjectives.
  • Morphology as a Filter: Even when the system suggests a "wrong" base noun, the adjective itself is often a valid word, making the tool useful for vocabulary discovery.

Key Results

  • Scale: The final lexicon contains 12,029 noun-adjective pairs, covering 9,692 distinct adjectives.
  • Precision: Manual evaluation of a sample (words starting with 'g') showed a precision of 78% for known adjectives and 51% for previously uncatalogued adjectives found in the corpora.
  • Coverage: The system successfully captured about 75% of regular adjectives in Wiktionary.
Evaluation of precision for known and unknown adjectives starting with 'g'.

Limitations and Future Work

The primary limitation is the system's inability to handle suppletive bases (like jeu → ludique) and semantic drift. Future work aims to integrate distributional semantics (word embeddings) to ensure noun-adjective pairs are related in meaning and to use French-Latin bilingual lexicons to bridge the gap for suppletive adjectives. The resulting lexicon is released under the LGPL-LR license.

Summary updated

Abstract

Après une brève analyse linguistique des adjectifs dénominaux en français, nous décrivons le processus automatique que nous avons mis en place à partir de lexiques et de corpus volumineux pour construire un lexique d’adjectifs dénominaux dérivés de manière régulière. Nous estimons à la fois la précision et la couverture du lexique dérivationnel obtenu. À terme, ce lexique librement disponible aura été validé manuellement et contiendra également les adjectifs dénominaux à base supplétive.
BibTeX
@inproceedings{strnadova2011construction,
  author = {Jana Strnadová and Benoît Sagot},
  title = {{Construction d'un lexique des adjectifs dénominaux (Construction of a lexicon of denominal adjectives)}},
  booktitle = {TALN (Articles Courts)},
  year = {2011},
}
Figure