Automatisation multilingue du prétraitement de transcriptions dans la détection de la maladie d’Alzheimer

Abstract

Alzheimer’s disease (AD) is a degenerative disease that is characterised mainly by progressive cognitive alterations. Linguistic functions alterations are a key factor in AD detection, as multiple studies have proven that they appear at an early stage. In order to measure those functions, we process and analyse transcripts from various corpora based on the Cookie Theft picture description task. More and more semi-automatic systems are trying to take advantage of image descriptions form the Cookie Theft picture description task. However, extracting quantitative measures of a transcript requires considerable effort. In addition, when working with a multilingual corpus, it is important to adapt the system in order to clearly differentiate the linguistic variants. Currently, no study has focused on the creation of a universal pipeline for this type of task, which could improve the reproducibility of scientific experiments. This thesis therefore presents a simple and effective method for dealing with transcripts, by sequencing a series of sub-tasks that cleans, normalize and extract measures from transcripts making it possible to identify cognitive deteriorations. Since some tasks are language dependant, they have been adapted to be easily configurable. Results have demonstrated that our work improves reproducibility of experiments. In fact, we were able to automatically normalise and extract linguistic measures from a French and English corpus of the Cookie Theft picture description task. In fact, the retracing’s frequency, extracted from transcripts, revealed a significant correlation with the severity of cognitive impairment (> 0.5). Then, we trained predictive models that produced similar results with previous studies, using the Pitt Corpus (76%). This demonstrates that our automated preprocessing task is reliable. However, given the large diversity of languages across the world and their own language structure, this method has its own limitations. Thus, this work contributes in Alzheimer’s disease literature by presenting a pipeline that improves reproducibility of experiments when analysing transcripts from Cookie Theft picture description task. Also, we think that our work could eventually be used for different type of cognitive tasks since it is not dependent of the context.

Date	15 Nov 2020
Original language	French
Awarding Institution	École de technologie supérieure
Supervisor	Sylvie Ratté (Supervisor)