Merging syntactic lexica: the case for French verbs

Benoît Sagot, Laurence DanlosHAL 2012

TL;DR

This paper tackles the challenge of combining multiple syntactic lexicons for French verbs, which are rich but costly to build. The authors propose a novel merging algorithm that converts diverse lexicons into a common format (Alexina) and then intelligently groups and merges entries, significantly improving coverage and accuracy. This work demonstrates a practical approach to leveraging existing linguistic resources for more robust NLP applications.

What Problem Does This Paper Address?

Developing comprehensive syntactic lexicons, which describe how words function grammatically (e.g., what arguments a verb takes), is crucial for natural language processing tasks like parsing. However, these lexicons are expensive and time-consuming to create. Existing lexicons, while valuable, often have different ways of representing syntactic information, varying in coverage, granularity, and formalization. This makes it difficult to combine them to create a single, superior resource. The paper addresses the problem of how to effectively merge these disparate syntactic lexicons for French verbs, aiming to create a more complete and accurate resource by combining the strengths of several existing ones.

Methods

The core of the paper is a three-step merging algorithm:

  1. Conversion to a Common Model: All input lexicons are converted into a unified format called Alexina. This framework is chosen because it's directly usable in NLP tools and uses a consistent model of syntactic functions, realizations, and redistributions, mirroring the structure of the Lefff lexicon. This step involves careful linguistic analysis to map information from diverse sources into Alexina, handling discrepancies in representation.
  2. Grouping Entries: For each verb lemma, entries from different lexicons are grouped together. This is a critical and challenging step because lexicons can have different granularities of entries. The algorithm uses an "inclusion relation" to formalize the idea that entries from one lexicon might be more specific than those from another. This relation is defined based on the set of syntactic functions an entry licenses. The grouping process identifies sets of entries that should be merged into a single output entry.
  3. Merging Entries: Once groupings are formed, the entries within each group are merged. The syntactic functions and their realizations are combined using a union operation. Mandatory realizations are preserved only if they are mandatory across all entries in the grouping. Redistributions are also merged. This process aims to create a single, comprehensive entry that incorporates information from all contributing lexicons.

Datasets

The paper utilizes four large-coverage syntactic lexicons for French verbs:

  • Lexicon-Grammar tables (LG): A collection of verb lexical entries structured into classes.
  • Dicovalence: A verb valency lexicon focusing on pronominal realizations.
  • LVF (Lexique des Verbes Français): A syntactico-semantic lexicon of French verbs.
  • Lefff (Lexique des formes fléchies du français): A large-coverage syntactic lexicon for French, which serves as the target lexicon for merging.

The authors describe experiments where they merge subsets of these lexicons, particularly focusing on pronominal entries and verbs ending in "-iser" and "-ifier."

Key Insights and Ideas

  • Leveraging Discrepancies: The paper demonstrates that the very differences in how lexicons represent syntactic information can be a source of strength when merged intelligently. The algorithm is designed to handle these discrepancies rather than being hindered by them.
  • Inclusion Relation for Grouping: The formalization of an "inclusion relation" between entries and lexicons is a key innovation. This allows for a principled way to group entries from lexicons with varying granularities, recognizing that some lexicons might offer more detailed distinctions than others.
  • Alexina as a Unifying Framework: The choice of Alexina as the common model is crucial. Its direct usability in NLP tools and its structured representation of syntactic information make it an effective target for conversion and merging.
  • Improving Input Resources: The merging process not only improves the target lexicon (Lefff) but also helps to identify and potentially correct errors or inconsistencies in the original input lexicons.

Key Results

The paper reports on several experiments demonstrating the effectiveness of their merging methodology:

  • Pronominal Entries: Merging Lefff with Dicovalence and a "light" version of Lexicon-Grammar tables resulted in 5,464 lexical entries, improving coverage of pronominal constructions.
  • Denominal/Deadjectival Verbs: Merging Lefff, Dicovalence, and LVF for verbs in "-iser" and "-ifier" created 2,246 entries covering 1,701 distinct lemmas.
  • Overall Granularity and Accuracy: A comprehensive merge of Lefff and Dicovalence aimed to increase granularity and accuracy. This work extended the number of verbal entries in the Lefff from 10,214 to 12,610 and distinct lemmas from 7,813 to 7,990.
  • Validation: Manual validation of the merged results helped to remove erroneous realizations from the Lefff and uncover errors in Dicovalence itself.

Limitations and Future Work

  • Grouping Algorithm Limitations: The current grouping algorithm cannot place two distinct entries from the same lexicon into a single grouping. It also relies solely on syntactic function inventories, potentially missing finer-grained distinctions (e.g., human vs. non-human arguments) that could improve grouping accuracy.
  • Future Work: The authors plan to address these limitations by improving the grouping algorithm to handle same-lexicon entries and incorporate more detailed information. They also aim to fully merge all four lexical resources, including a complete conversion of Lexicon-Grammar tables, to create an even more comprehensive syntactic resource for French verbs.

Summary updated

Abstract

International audience
BibTeX
@inproceedings{sagot2012merging,
  author = {Benoît Sagot and Laurence Danlos},
  title = {{Merging syntactic lexica: the case for French verbs}},
  booktitle = {HAL},
  year = {2012},
}

Loading related content...

Figure