Word Profile Automatic Tracking Through the Prism of Construction Grammars and Distributionalism

International audienceThis talk will detail the first steps to automatically track semantic change of alreadyentrenched form-meaning pairs in contemporary French. These experimentations arepart of Neoveille, an on-going collaborative project funded by the Sorbonne ParisCité University whose main goal is to track contemporary formal and semanticneologisms in seven languages (French, Greek, Polish, Czech, Brasilian Portuguese,Russian and Chinese).Tracking semantic change implies knowing the existing meaning of words and thentrack the changes occurring to this configuration. We here focus on the first step.Sophisticated linguistically-motivated models have been proposed to describemeaning. Among others, the Meaning-Text Theory is certainely the most detailed,whereas the Frame Semantics Theory has given rise to some resources. But theyresult in description of small parts of the lexicon, as they both rely on assumptionsthat requires linguistic and human expertise, and thus cannot be applied on a largescale.Two computational-aware models have been proposed to approach meaning, bothbeing based on the distributional hypothesis : the first assumption relies on countingsequences repetition and is summarized by the well-known formula that « we canknow the meaning of a word by the company it keeps » (Firth, 1957) ; the secondone is based on (Harris, 1970 ; 1988) assumption : « ...if we consider words ormorphemes A and B to be more different in meaning than A and C, then we will oftenfind that the distributions of A and B are more different than the distributions of A andC. In other words, difference of meaning correlates with difference of distribution.(Harris, 1970, p.786) ». A third derived assumption linked to the distributionnal theoryis that linguistic phenomena are a matter of scale, there is no discrete distinctions,even if we can identify distributional classes.The first approach aims at describing a word meaning through its combinatorialprofile, ie the most prototypical and frequent collocates or collostructions that a baseform has in corpus. It has been studied for at least twenty years, leading to a wealthof frequency calculus to overcome the biases of the absolute frequency. Importantnotions have emerged : collocation, collostruction and even word profile or sketch(Kilgariff, 2004, 2014).The second approach aims at clustering words on the basis of a common distributionfrom a representative corpus. It results in highlighting similarity, a notion coveringseveral classical semantic relations (synonymy, hyponymy-hyperonymy, antonymy,analogy).The Construction Grammar paradigm is also mostly usage-based and propose aframework complementing the distributional model. The most important assumptionsreside in the definition of the form-meaning pair as occurring at various levels, andthe refusal of classical intermediary levels between form and meaning. Constructionsare characterized by three parameters : size (from atomic to complex), phonologicalspecificity (from substantive to schematic) and type of concepts (from contentful toprocedural) (Traugott and Trousdale, 2013). The form-meaning pairs change throughtime is called entrenchment (Langacker, 1987), and is caused by several factors(Schmid, 2016) that affect their schematicity, productivity and compositionality.In the experiment described in this talk, we will apply the distributional assumptions(relative frequency, similarity distributions) and try to derive automatically formmeaningprofiles in the CG terms. This experiment is based on a large-scalemorpho-syntaxically annotated French corpus (Wikipedia). On this corpus, we applya Frequent Patterns calculation (Van Gompel et al., 2016) complemented by a stateof-the-art vector space model (Mikolov et al., 2013), giving rise to promising resultson several French verbs : we do not only identify form-meaning typical constructions,but also cluster words in the argumental positions. We analyze the results, focusingon several discursive constraints to be dealt with to effectively derive word profiles.We draw some perspectives for the work. Results will be made available on theNeoveille website