Say Anything: Automatic Semantic Infelicity Detection in L2 English Indefinite Pronouns

Proceedings of the 23rd Conference on Computational Natural Language Learning, pages 77–86 Hong Kong, China, November 3-4, 2019. c⃝2019 Association for Computational Linguistics 77 Say Anything: Automatic Semantic Infelicity Detection in L2 English Indefinite Pronouns Ella Rabinovich1 Julia Watson1 Barend Beekhuizen2 Suzanne Stevenson1 1Dept. of Computer Science, University of Toronto, Canada 2Dept. of Language Studies, University of Toronto, Canada {ella,jwatson,suzanne}@cs.toronto.edu barendbeekhuizen@utoronto.ca Abstract Computational research on error detection in second language speakers has mainly ad- dressed clear grammatical anomalies typical to learners at the beginner-to-intermediate level. We focus instead on acquisition of subtle se- mantic nuances of English indefinite pronouns by non-native speakers at varying levels of proficiency. We first lay out theoretical, lin- guistically motivated hypotheses, and support- ing empirical evidence on the nature of the challenges posed by indefinite pronouns to En- glish learners. We then suggest and evaluate an automatic approach for detection of atypical usage patterns, demonstrating that deep learn- ing architectures are promising for this task in- volving nuanced semantic anomalies. 1 Introduction The ubiquity of English as an online lingua franca offers a rich opportunity for computational re- search on second language acquisition and on tools for aiding non-native speakers. Most com- putational research in second language (L2) has focused on spelling and grammar errors, and has been conducted on learners with beginner-to- intermediate proficiency level (henceforth, “learn- ers”) (e.g. Ji et al., 2017; Sakaguchi et al., 2017; Rozovskaya et al., 2017; Lo et al., 2018). Lit- tle empirical work has looked at semantic errors, with existing research mostly focusing on colloca- tions (e.g., Dahlmeier and Ng, 2011; Vecchi et al., 2011; Kochmar and Briscoe, 2013). Also, highly proficient, advanced L2 speakers (henceforth, “ad- vanced L2s”) have received little attention (though see Daudaravicius et al., 2016). In contrast to learners, these speakers rarely violate grammati- cal norms of the L2, but rather deviate from native usage in much more nuanced ways, often exhibit- ing mild infelicities rather than outright errors. We aim to explore an elusive aspect of master- ing the subtle contours of a word’s meaning that are shaped by its context. Specifically, we in- vestigate patterns of acquisition of English indefi- nite pronouns by L2 speakers. Indefinite pronouns (IPs) are linguistic devices that refer to an entity (such as a person or thing) that has not yet been introduced in discourse. In English, examples are words like someone, anything, and nobody. Con- sider the following sentences, taken verbatim from corpora of L2 speakers (original pronoun is bold- faced; less felicitous usages marked with ‘?’).1 1. Do you know someone/anyone who was dis- criminated based on gender? 2. It was a little amazing, because they didn’t stole ?something/anything. 3. ??Anyone/Someone told me the company has millions in debts and isn’t able to pay it. Clearly, mastery of IPs in English relies on recog- nizing subtle factors that determine their appropri- ate usage in various contexts. Here, in Section 2, we develop a linguistic anal- ysis with detailed hypotheses on precisely how the tangled relations …

Content truncated to 500 words