Analyzing Transformers Representations. A Linguistic Perspective

old_uid19851
titleAnalyzing Transformers Representations. A Linguistic Perspective
start_date2021/12/17
schedule11h UTC+1
onlineno
detailsLocation/connection: Due to current restrictions, you are invited to attend online (Google Meet). The link will be shared on the day in this google doc https://docs.google.com/document/d/1bnuIxo9F2WLPsdI9VIn233fMVwybFSUequrUh4vlYkk/edit?usp=sharing.
summaryTransformers have become a key component in many NLP models, arguably because of their capacity to uncover contextualized distributed representation of tokens from raw texts. Many works have striven to analyze these representations to find out whether they are consistent with models derived from linguistic theories and how they could explain their ability to solve an impressive number of NLP tasks. In this talk, I will present two series of experiments falling within this line of research and aiming at highlighting the information flows within a Transformer network. The first series of experiments focuses on the long distance agreement task (e.g. between a verb and its subject), one of the most popular methods to assess neural networks’ ability to encode syntactic information. I will present several experimental results showing that transformers are able to build an abstract, high-level sentence representation rather than solely capturing surface statistical regularities. In a second series of experiments, I will use a controlled set of examples to investigate how gender information circulates in an encoder-decoder architecture considering both probing techniques as well as interventions on the internal representations used in the MT system. Joint work with Bingzhi Li, Benoit Crabbé, Lichao Zhu, Nicolas Bailler and François Yvon
responsiblesSeddah