“Chère maison” or “maison chère”? Transformer-based prediction of adjective placement in French

Findings of the Association for Computational Linguistics: EACL 2023, pages 1668–1683 May 2-6, 2023 ©2023 Association for Computational Linguistics Chère maison or maison chère? Transformer-based Prediction of Adjective Placement in French Eleni Metheniti ① ② Tim Van de Cruys ③ Wissam Kerkri ④ ① CLLE-CNRS ④ Université Toulouse - Jean Jaurès (UT2J) Juliette Thuilier ① ② IRIT Université Toulouse - Paul Sabatier (UT3) firstname.lastname@{univ-tlse2.fr,irit.fr,kuleuven.be} Nabil Hathout ① ③ UK Leuven Department of Linguistics Leuven.AI institute Abstract In French, the placement of the adjective within a noun phrase is subject to variation: it can ap- pear either before or after the noun. We con- duct experiments to assess whether transformer- based language models are able to learn the ad- jective position in noun phrases in French—a position which depends on several linguistic factors. Prior findings have shown that trans- former models are insensitive to permutated word order, but in this work, we show that fine- tuned models are successful at learning and selecting the correct position of the adjective. However, this success can be attributed to the process of finetuning rather than the linguistic knowledge acquired during pretraining, as evi- denced by the low accuracy of experiments of classification that make use of pretrained em- beddings. Comparing the finetuned models to the choices of native speakers (with a question- naire), we notice that the models favor context and global syntactic roles, and are weaker with complex structures and fixed expressions. 1 Introduction In French, the placement of the adjective is sub- ject to a considerable amount of variation—a phe- nomenon that has been under close scrutiny among linguists. Generally speaking, adjective placement in anteposition or postposition is attributed to many intertwining, linguistic processes, rather than a few rigid grammatical rules. However, the order of the adjective can be crucial to the meaning of the noun phrase; in the titular example, chère maison means “dear house" but maison chère means “expensive house". Meanwhile, natural language processing re- searchers investigate whether language models built by transformer architectures are able to cap- ture some of the inner workings of human language during their learning process. So far, research has shown that the high performance of such models does not imply the understanding of basic concepts such as grammatical order because the transformer architecture is non-sequential by design. We are exploring whether transformer-based lan- guage models are capable of perceiving the adjec- tive’s position in a sentence with regard to its head noun, with a variety of experiments. Our goal is not to set a new state-of-the-art, but to explore if and how this information on word order is learned and used in tandem with the contextual word em- bedding information. While previous work has shown that transformer models are insensitive to word order (Pham et al., 2021; Gupta et al., 2021), finetuned models were successful in classifying adjective position (Sinha et al., 2021b). We also tested variations of finetuning training sizes and the use of attention masks to hide either the context of the …

Content truncated to 500 words