Synonym-Grouping Method: Mapping Synonyms to Senses

• The Synonym-Grouping Method is a knowledge-based approach that refines broad synonym sets into sense-specific subgroups using gloss-based similarity metrics.
• The methodology indexes dictionary definitions and merges synonym glosses, computing both simple and vector-based similarities to map synonyms to appropriate senses.
• Performance metrics show a precision of 67%, recall of 71%, and an F1 score of approximately 0.70, outperforming random baselines and approaching human agreement.

A synonym-grouping method is an approach for mapping the set of synonyms associated with a lexical item ("lemma") onto the distinct senses (i.e., dictionary definitions) delineated for that item, refining broad synonym sets into semantically focused subgroups. The methodology presented in "Grouping Synonyms by Definitions" (0909.3445) centers on French verbs, leveraging high-quality manually constructed resources—the TLFi (Trésor de la Langue Française informatisé) for definitions and five established synonym dictionaries for candidate synonym sets. The process assigns each synonym to the specific sense(s) of the lemma it best matches by quantifying similarities between definition indices and merged synonym glosses using advanced overlap and vector-based metrics.

1. Lexical Resource Integration and Problem Definition

The input resources consist of:

• TLFi, which provides semantically rich, manually edited verb definitions, including open-class content words, domain tags, and inherited taxonomic cues.
• Five French synonym dictionaries offering lists of candidate synonyms per lemma.

For each verb $V$:

• The set of possible senses is $D_V = \{d_1, d_2, ..., d_n\}$, each $d_i$ a definition from TLFi.
• The candidate synonym set is the union of all synonyms for $V$ across the five dictionaries.

The core problem is: For each synonym $S$ of $V$, assign $S$ to one or more $d_i \in D_V$ such that the assignment respects the semantic correspondence as estimated from available textual cues.

2. Methodological Workflow

Each step is designed to enable high-fidelity mapping of synonyms to senses:

1. Definition Indexing: For each $d_i$, create an index of open-class words and, optionally, domain and synonymic cues:

$$\text{index}(d_i) = \langle w_1, w_2, ..., w_k \rangle$$

For example, the definition "Jeter loin en avant avec force" is indexed as $\langle$ jeter, loin, avant, force $\rangle$.

1. Synonym Definition Merging: For each synonym $S$ found in the dictionaries, gather all TLFi definitions for $S$ (if present), merging their indices into a "sense profile" for $S$.
2. Gloss-based Similarity Computation:

Compute a similarity score between each $d_i$'s index and $S$'s merged profile using: - Simple Overlap (Lesk-inspired): Count of common tokens. - Extended Overlap: Rewards phrase overlap, weighting $n$-gram overlaps by $n^2$. - Vector approaches:

For one method, assign weighted word vectors based on definition frequency, e.g.,

$$w_i = \frac{\text{tf}(w_i, d_i)}{\text{tf}(w_i, D_V)}$$

Compute vector dot products as similarity.

1. Assignment: For each synonym $S$, assign it to the $d_i$ with the maximal (nonzero) similarity.
2. Reflexive/Non-reflexive Disambiguation: To address French verb alternation (e.g., "abandonner" vs. "s’abandonner"), the entries are automatically partitioned so synonym comparisons are restricted to compatible usages.

3. Evaluation Protocol and Gold Standard

A rigorous evaluation combines manual and automatic assessment:

• Sample: 27 French verbs spanning diverse degrees of polysemy, frequency, and genericity.
• Manual Annotation:

For each (verb, synonym) pair, four expert lexicographers specify which TLFi definitions are adequate matches for the synonym, establishing a (verb, definition, synonym) gold standard.

• Inter-annotator Agreement:
  • Pairwise: 74.06%–87.07%
  • Four-way: 63.37%
  • These rates quantify both the intrinsic difficulty and the plausible "human upper bound" for the task.

4. Performance Metrics

Algorithmic performance (precision, recall, F1) is reported relative to the gold standard.

• Precision: 67% (fraction of assigned synonym-to-definition pairs found correct according to the lexicographers)
• Recall: 71% (fraction of gold standard synonym-to-definition pairs captured by the method)
• F-score: $\approx 0.70$

These figures substantially exceed a random-baseline, indicating that definition text (despite its brevity) provides sufficient discrimination for effective synonym sense grouping. Human agreement not reaching 100% further contextualizes the practical ceiling.

5. Technical Formulations

Key technical aspects include:

• Definition Index Construction: Uses open-class tokens and, as needed, augmenting cues (domain, synonym).
• Similarity Scoring Formulations:
  • Overlap-based: $|\text{index}(d_i) \cap \text{profile}(S)|$
  • Extended: Summing $n^2$ for all $n$-gram matches.
  • First-order vector: Weighted as detailed above.
  • Second-order vector: Embeds distributional similarities across all definitions.
• Assignment Rule: Select $d_i$ for $S$ such that $\text{sim}(d_i, S)$ is maximal and $> 0$.
• Reflexive/Non-reflexive Filtering: Partition verbs and synonyms contextually before index and similarity calculation.

The method includes a comparative assessment of six similarity metrics, with metrics documented in source tables, analyzed with and without the reflexive/non-reflexive alternation distinction.

6. Relation to Prior Research

Comparison axes with established approaches include:

• Word Sense Disambiguation (WSD):

Unlike corpus-based (supervised or context-sensitive) WSD, this approach is strictly knowledge-based, aligning dictionary sense indices without recourse to explicit usage corpora.

• Synonym Lexicon Acquisition:

Recent work emphasizing automatic synonym extraction often conflates true synonyms with associatively related terms (hypernyms, antonyms). By starting with expert-validated synonymy and refining it with sense alignment, this method achieves higher semantic precision.

• WordNet Construction:

Projects like WOLF for French attempt automatic synset induction via English-French mapping, but the present method aligns directly at the French lexicographic sense level—yielding finer-grained, TLFi-compatible synonym groupings and complementing the WordNet tradition.

Unique features include the explicit accounting for reflexive alternations, direct mapping to authoritative definitions, and the systematic merging of multiple synonym dictionaries into a single, sense-annotated synonym resource.

7. Applications and Implications

Applications of this synonym-grouping method span:

• Resource Construction:

Unified, high-recall synonym lexica for French verbs, indexed by sense, can be constructed by merging dictionary data via this technique.

• Query Expansion and Semantic Search:

Synonyms tagged with sense enable precise query rewriting strategies in search systems and more accurate text summarization.

• WSD and Machine Translation:

Sense distinctions facilitate finer disambiguation in translation pipelines, improving alignment and context-sensitive translation.

• WordNet and Lexicographic Research:

The sense-grouped synonym resource is a precursor/foundation for French WordNet construction, ontology alignment, and multilingual resource development.

• Meta-lexicography:

Empirical findings about the correlation (and divergence) between synonym sets and sense differentiation inform future dictionary design and annotation schemes.

---

The synonym-grouping method described thus constitutes a robust, linguistically grounded framework for aligning synonyms with fine-grained senses, yielding linguistically rich, high-precision lexical resources applicable to multiple computational linguistic tasks and lexicographic studies. Its integration of gloss-based similarity, sense-level frequency weighting, and alternation-aware filtering enables a performance level that approaches human annotator agreement, while the underlying methodology is extensible to other languages and lexical domains.