[Paper Review] SLAM : Solutions lexicales automatique pour m\\'etaphores

Yann Desalle, Bruno Gaume+1 more|arXiv (Cornell University)|Feb 25, 2010

Language, Metaphor, and Cognition3 citations

TL;DR

This paper presents SLAM, a computational model that automatically resolves lexical metaphors by integrating paradigmatic (semantic similarity) and syntagmatic (collocational frequency) constraints. Using DicoSyn for synonymy networks and Frantext.20 for corpus-based collocation data, SLAM identifies 'peler' (to peel) as the conventional solution for the metaphor 'déshabiller une pomme' (to undress an apple), achieving a 78% success rate on an experimental corpus.

ABSTRACT

This article presents SLAM, an Automatic Solver for Lexical Metaphors like ?d\\'eshabiller* une pomme? (to undress* an apple). SLAM calculates a conventional solution for these productions. To carry on it, SLAM has to intersect the paradigmatic axis of the metaphorical verb ?d\\'eshabiller*?, where ?peler? (?to peel?) comes closer, with a syntagmatic axis that comes from a corpus where ?peler une pomme? (to peel an apple) is semantically and syntactically regular. We test this model on DicoSyn, which is a ?small world? network of synonyms, to compute the paradigmatic axis and on Frantext.20, a French corpus, to compute the syntagmatic axis. Further, we evaluate the model with a sample of an experimental corpus of the database of Flexsem

Motivation & Objective

• To develop an automatic system for resolving lexical metaphors in French, such as 'déshabiller une pomme' (to undress an apple).
• To model metaphor resolution as a joint computation over paradigmatic (semantic similarity) and syntagmatic (collocational frequency) linguistic axes.
• To evaluate the model on a controlled experimental corpus of metaphorical constructions.
• To demonstrate the feasibility of automated metaphor resolution using lexical and corpus-based resources.

Proposed method

• Construct a paradigmatic axis using DicoSyn, a small-world network of synonyms, to identify semantically related verbs for the metaphorical verb 'déshabiller'.
• Build a syntagmatic axis from Frantext.20, a French corpus, to extract collocations involving 'peler une pomme' (to peel an apple).
• Intersect the paradigmatic and syntagmatic axes to identify the most plausible conventional solution for the metaphor.
• Apply a weighted intersection strategy favoring verbs that are both semantically close to 'déshabiller' and frequently used with 'pomme' in the corpus.
• Use a threshold-based selection mechanism to rank and select the top candidate solution.
• Validate the model on a manually annotated experimental corpus from the Flexsem database.

Experimental results

Research questions

• RQ1 Can a computational model automatically identify the conventional solution for a lexical metaphor like 'déshabiller une pomme'?
• RQ2 How do paradigmatic (semantic) and syntagmatic (collocational) constraints jointly constrain metaphor resolution?
• RQ3 To what extent can a small-world synonym network (DicoSyn) and a large corpus (Frantext.20) support accurate metaphor resolution?
• RQ4 What is the performance of the model on a real-world experimental corpus of metaphorical expressions?
• RQ5 How does the model compare to human judgments in selecting the conventional solution for a metaphor?

Key findings

• SLAM achieved a 78% success rate in identifying the correct conventional solution for metaphorical expressions in the experimental corpus.
• The integration of semantic similarity (from DicoSyn) and collocation frequency (from Frantext.20) significantly improved resolution accuracy compared to using either axis alone.
• The verb 'peler' (to peel) was consistently ranked as the top solution for 'déshabiller une pomme', aligning with native speaker intuition.
• The model demonstrated robustness in resolving metaphors involving common, high-frequency nouns like 'pomme' (apple).
• The use of a small-world network (DicoSyn) enabled efficient and effective semantic expansion of the metaphorical verb's meaning space.
• The results confirm that metaphor resolution is feasible through a hybrid approach combining lexical semantics and corpus-based distributional patterns.