In the framework of our Strand 5 (Computational Semantic Analysis), we are looking for a postdoctoral researcher in Diachronic Computational Semantics.
Research area : interpretable computational models for automatic detection and monitoring of semantic evolutions: combination of Contextual Embeddings and Pattern Mining approaches
Contract duration: 18 months
Research Laboratory: Sorbonne Paris Nord University, LIPN UMR7030 CNRS
Application deadline: January 15, 2022
Audition period: January 15-30, 2022
Job Starting date: from February 1, 2022
Context, Issues and research axes
Languages are constantly evolving, driven by the need to adapt to socio-cultural and technological developments and to make communication more efficient and expressive. In particular, new words are forged or borrowed from other languages, some words become obsolete, others acquire new meanings or lose existing meanings.
In NLP, the study of language dynamics, especially from the lexical point of view, has gained audience in recent years, complementing synchronic approaches. The field of research is structuring itself, with recent state of the art (Monteirol et al., 2021; Tahmasebi et al., 2021) and several scientific events (International Workshop on Computational Approaches to Historical Language Change 2019 and 2021, ACL 2019 and 2020). Two initial evaluation tasks have been proposed (Unsupervised Lexical Semantic Change Detection Task, SemEval2020) and reference sets have been set up for four languages (English, Latin, Swedish and German).
Lexical change detection systems have followed advances in NLP methods: after the first systems essentially based on frequency changes (for example Gulordova & Baroni, 2011), systems used word embeddings (Kim et al., 2014, Schletchweg et al., 2019) and more recently contextual embeddings (Hu et al., 2019; Martinc et al., 2019; Giulianelli et al., 2020). These latter systems generally proceed by grouping the contextual vector representations of the different uses into clusters of meaning, then detect changes according to different metrics (Monteirol et al. 2021). Current systems still face many limitations. Mainly, the opacity of neural models does not make it possible to characterize these evolutions, in particular it is difficult, if not impossible, to link the semantic changes to linguistic morphological, syntactic or lexico-syntactic features, or to categorize the types of changes (extension, restriction, metaphor, metonymy, etc.). To this end, one avenue would be to combine neural approaches with Pattern Mining (Béchet et al. 2015) or collocation extraction approaches from corpus linguistics (for example Gries, 2012) which make it possible to extract the most salient lexico-syntactic patterns of a given meaning from a corpus of occurrences and thus identify the evolution. It would also be interesting to use the contextual information of the occurrences (date, type of source, domain, diatopic and diastratic features, etc.) to characterize and follow the evolution of usages.
The job main objective is therefore to set up a system combining these approaches to allow an automatic characterization of semantic evolutions. The first step will consist in experimenting with state-of-the-art models for detecting changes. The second step will then try to combine contextual embeddings and pattern mining approaches / collocation extraction to highlight the linguistic characteristics of each of the meaning clusters and their evolution. The studied corpora will be mainly in English and French. The postdoctoral fellow will work in collaboration with computer scientists and linguists from the Labex who are currently building a reference corpus of semantic evolutions for French (following the Durel methodology: Schlechtweg et al., 2018).
Other issues may also be addressed by the recruited person, and in particular: current systems do not take into account the graduality of evolutions, generally being limited to comparing two synchronic language states; to get the vector representation of a lexis in a context, it is possible to use one of the hidden layers or a combination of them. There is currently no consensus on the most adequate layer to take into account to obtain the most adequate semantic representation.
The recruited person will join the strand 5 (“Computational Semantics”) of the Labex, specifically the research team working on the “Semantic Variation and Change” operation which aims to:
develop new models and methods for the automatic detection of lexical semantic changes, the typology of changes from intra- and extra-linguistic points of view;
develop a reference dataset of semantic evolutions in contemporary French, based on available diachronic corpora.
- PhD in computer science specialised in Computational Linguistics and Machine Learning
- deep learning methods and language models attested training and experience
- working language: French and / or English
Please send :
• a cover letter
• a description of the research project related to the research questions
• a CV with a list of publications and 3 representative publications (pdf or link),
• letters of recommendation or names of two referees.
to firstname.lastname@example.org and email@example.com before January 15, 2022. The auditions of the pre-selected candidates will take place at the end of January 2022.
Béchet N., Cellier P., Charnois T. and Crémilleux B. (2015). “Sequence mining under multiple constraints”. In Proceedings of the 30th Annual ACM Symposium on Applied Computing (SAC 2015), ACM Press, Salamanca, Spain, pages. 908--914.
Giulianelli, M., Tredici, M.D., & Fernández, R. (2020). “Analysing Lexical Semantic Change with Contextualised Word Representations”. Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 3960–3973 July 5 - 10, 2020. https://www.aclweb.org/anthology/2020.acl-main.365.pdf
Gries Stefan Th. (2012). "Behavioral Profiles: a fine-grained and quantitative approach in corpus-based lexical semantics". In Gonia Jarema, Gary Libben, Chris Westbury (eds.), Methodological and analytic frontiers in lexical research, 57-80. Amsterdam Philadelphia: John Benjamins.
Montariol, S. (2021). Models of diachronic semantic change using word embeddings. (Modèles diachroniques à base de plongements de mot pour l'analyse du changement sémantique). PhD Thesis, Paris-Saclay. 223 pages https://tel.archives-ouvertes.fr/tel-03199801/document
Montariol S., Doucet A. and Allauzen A. (2021). “Etat de l’art du changement sémantique à partir de plongements contextualisés”. In Coria 2021, http://coria.asso-aria.org/2021/articles/court_27/main.pdf
Montariol, S., Martinc, M., & Pivovarova, L.