080 | Evaluating the semantic representation contextualization of polysemic words

Cognition, Behavior, and Memory

Author: Bruno Bianchi | Email: brunobian@gmail.com


Bruno Bianchi , Julieta Laurino , Diego Fernández Slezak , Laura Kackzer , Juan Esteban Kamienkowski

1° Laboratorio de Inteligencia Artificial Aplicada, Instituto de Ciencias de Computación, Facultad de Ciencias Exactas y Naturales (UBA-CONICET)
2° Departamento de Computación, Facultad de Ciencias Exactas y Naturales, Universidad de Buenos Aires
3° Instituto de Fisiología, Biología Molecular y Neurociencias, Facultad de Ciencias Exactas y Naturales (UBA-CONICET)
4° Maestría de Explotación de Datos y Descubrimiento del Conocimiento, Facultad de Ciencias Exactas y Naturales, Universidad de Buenos Aires

Polysemy (words with the same spelling and different, but related, meaning) is a feature present in many languages. For example, in English it is estimated that 80% of words are polysemic. This generates that, when we come across a word, either visually or aurally, our brain must quickly interpret what meaning is being referred to, based on the context in which it appeared. The mechanisms that operate to carry out this processing are not precisely known.

Meanwhile, state-of-the-art computational Language Models have achieved a level of language comprehension that allows them to recognize the correct meaning that is being assigned to polysemic words when used in context.

In the present work we present a preliminary comparative analysis between behavioral responses of humans and the mechanisms that operate in Language Models against the interpretation of neutral sentences (which do not allow defining the meaning of the polysemic words present) contextualized with biasing paragraphs. This work is a first step towards a better understanding of the brain mechanisms underlying this process.