Linguistic interoperability within a unified architecture - Productions scientifiques du CLILLAC-ARP
Communication Dans Un Congrès Année : 2024

Linguistic interoperability within a unified architecture

Résumé

Modern approaches to quantitative linguistics rely on large datasets. These datasets are representations of linguistic observations made up of features of various dimensions. Analyses rely on these data representations. As a result, the question of their construction is essential. In the case of quantitative research methods, datasets are built atomatically. This includes the sequential use of various types of NLP tools. These tools are the components of software architectures that ensure data interoperability (Rehm et al., 2020; Sérasset et al., 2009). In this paper, we present the data architecture used in the Analytics for Language Learning (A4LL) project (Gaillat, 2022a) whose aim is to produce linguistic analytics for foreign language teachers. The system transforms learner writings into linguistic feature sets which are used to produce linguistic indicators of L2 writing performance. The system relies on an architecture composed of NLP tools. Multilingual Linguistic annotation is provided by UDPipe (Straka et al., 2016) and Stanza (Qi et al., 2020) and stored in a database. Several tools exploit this annotation to compute textual measures of different linguistic dimensions. Syntactic complexity ratios are extracted with the use of TAASSC (Kyle, 2016). Text cohesion ratios are computed with TAACO (Crossley et al., 2016). Noun-Verb collocations are extracted (Li et al., 2022) and their association strength is computed by NLTK (Bird & Loper, 2004), yielding a number of scores. Specific linguistic structural complexity indicators are computed on by language models predicting small groups of words related to each other by the same paradgmitatic relationships (Gaillat, 2022b). Some lexical sophistication indicators are computed via BERT-trained language models (Devlin et al., 2019). All these indicators are filtered out as part of a graphical vizualisation module of the architecture. Overall this process relies on interoperable data representations. The architecture is hosted on a virtual machine of the Huma-Num consortium. The Docker technology supports a modular implementation of the system in which all modules act as standalone programmes with the aformentionned roles. These programmes are related to each other thanks to APIs and a data transfer queuing system. Automation and interoperability are the cornerstones of quantitative data processing.
Fichier principal
Vignette du fichier
A4LL_LLcD_2024.pdf (1.87 Mo) Télécharger le fichier
Origine Fichiers produits par l'(les) auteur(s)

Dates et versions

hal-04712737 , version 1 (27-09-2024)

Identifiants

  • HAL Id : hal-04712737 , version 1

Citer

Thomas Gaillat, Cyrielle Mallart, Andrew J. Simpkin, Rémi Venant, Nicolas Ballier, et al.. Linguistic interoperability within a unified architecture. Langues & Langage à la croisée des Disciplines - 1ère Rencontre annuelle LLcD, Sorbonne Université; cnrs, Sep 2024, Paris, France. ⟨hal-04712737⟩
16 Consultations
2 Téléchargements

Partager

More