Exploring the use of dependency parsing in automatic erroneous collocation extraction in learner English - Archive ouverte HAL Accéder directement au contenu
Communication Dans Un Congrès Année : 2022

Exploring the use of dependency parsing in automatic erroneous collocation extraction in learner English

Résumé

Second language learners usually encounter difficulties in collocations both for writing and for oral expression (Garner et al., 2020; Granger & Larsson, 2021; Nesselhauf, 2003; Uchihara et al., 2021). As a subset of semantic phraseme, collocation is not free: without freedom of selection of its signified and without freedom of combination of its components (Mel’čuk, 1998; Tutin, 2013). In this sense, we considered a collocation is erroneous when it is not a standard prefabricated pattern, for example, “to *create [construct] a taller building”. The correction of collocations in written essays could help learners increase their competence and thus their proficiency in English writing (Meunier & Granger, 2008). A collocation extraction module could facilitate corpus-based phraseological analysis and thus help understand the interlanguage development stages (Liu & Lu, 2020; Schneider & Smith, 2015). It could also be incorporated into a computer-assisted language learning (CALL) system to help learners write and use collocations appropriately. A prototype of automatic Verb-Noun (VN) collocation detection was developed and reported in our previous study (Li & Gaillat, 2020). We found that there were mainly six causes that degraded the performance (Li & Gaillat, 2021). First was that some extracted pairs were frequent composites but not collocations, for example, “see section”, “thank president”. The second was that the noun was not the object of the verb, for example, “give (something) (to) dog”, “remove (something) (from) heat”. Other remaining causes were in principle due to Part-Of-Speech (POS) errors. Recently, dependency parsing was reported to improve the quality of collocation extraction (Uhrig & Proisl, 2012). Dependency parsing reveals the pairwise syntactic relations between words in the sentences, i.e. the dependence of a word on a head-word. We considered that syntactic parsing is a line of research to explore. On the previous basis, we shall use a dependency parser to further improve the performance of our module. With dependency annotation, we may restrict the word association analysis to those pairs that have a specific grammatical relation: for example, the VN pairs in accusative cases in which the noun is the direct object (dobj) of the verb. This may also help find long-distance pairs that are outside a window of adjacent tokens. Uhrig et al. (2018) systematically studied various dependency parsers and schemes for the extraction of standard collocations. However, since automatic parsers are generally developed for native language data, Huang et al. (2018) demonstrated that, despite the high accuracy (>80%), parsers built from standard English are not robust to learner errors: 63% of the learner errors caused at least one parsing error. Berzak et al. (2016) proposed a Treebank of Learner English and measured the effect of grammatical errors on parsing accuracy. Yet, to the best of our knowledge, the impact of dependency parsing on erroneous collocation extraction has not been studied. This research presents a work-in-progress report about extracting erroneous VN collocations. The research question is: can a dependency parser built from standard English data be used to extract erroneous collocations in a learner corpus? For evaluation, we used the National University of Singapore Corpus of Learner English (NUCLE) (Dahlmeier et al., 2013) where grammatical errors, including wrong collocations, are annotated by professional instructors. The main principle is, firstly, to extract all possible collocations in the learner corpus, and then identify standard collocations by comparing extractions with examples from a reference corpus (in our case collocations extracted from British National Corpus) (Li & Gaillat, 2021); the remainder of the items are considered as erroneous collocations. We will compare the performance (precision and recall rate) of the extraction of erroneous collocations on the basis of dependency relations to the previously developed window- based approach. The results will be evaluated by manual inspection to investigate the effect of learner errors on dependency parsing and erroneous collocation extraction. Future improvement is envisaged based on a large web corpus as a reference (Paquot et al., 2021) and other statistical measures (Gries & Durrant, 2020). Note: NUCLE is a collection of 1,414 essays (in a total of 1.2 million words) written by students who are non-native English speakers. It is available by submitting a license agreement via https://www.comp.nus.edu.sg/~nlp/corpora.html
Fichier non déposé

Dates et versions

hal-03962631 , version 1 (30-01-2023)

Identifiants

  • HAL Id : hal-03962631 , version 1

Citer

Jen-Yu Li, Thomas Gaillat, Élisabeth Richard. Exploring the use of dependency parsing in automatic erroneous collocation extraction in learner English. LCR 2022, Sep 2022, Padua, Italy. ⟨hal-03962631⟩
22 Consultations
0 Téléchargements

Partager

Gmail Facebook X LinkedIn More