Understanding Linguistic Accommodation in Code-Switched Human-Machine Dialogues

Proceedings of the 24th Conference on Computational Natural Language Learning, pages 565–577 Online, November 19-20, 2020. c⃝2020 Association for Computational Linguistics https://doi.org/10.18653/v1/P17 565 Understanding Linguistic Accommodation in Code-Switched Human-Machine Dialogues Tanmay Parekh Language Technologies Institute Carnegie Mellon University tparekh@cs.cmu.edu Emily P. Ahn Department of Linguistics University of Washington eahn@uw.edu Yulia Tsvetkov Language Technologies Institute Carnegie Mellon University ytsvetko@cs.cmu.edu Alan W Black Language Technologies Institute Carnegie Mellon University awb@cs.cmu.edu Abstract Code-switching is a ubiquitous phenomenon in multilingual communities. Natural lan- guage technologies that wish to communicate like humans must therefore adaptively incorpo- rate code-switching techniques when they are deployed in multilingual settings. To this end, we propose a Hindi-English human–machine dialogue system that elicits code-switching conversations in a controlled setting. It uses different code-switching agent strategies to un- derstand how users respond and accommodate to the agent’s language choice. Through this system, we collect and release a new dataset COMMONDOST, comprising of 439 human– machine multilingual conversations. We adapt pre-defined metrics to discover linguistic ac- commodation from users to agents. Finally, we compare these dialogues with Spanish- English dialogues collected in a similar set- ting, and analyze the impact of linguistic and socio-cultural factors on code-switching pat- terns across the two language pairs.1 1 Introduction When interlocutors share more than one lan- guage, they nearly inevitably engage in code- switching (CS): shifting from one language to an- other (Sankoff and Poplack, 1981; Muysken, 2000; Auer, 2013). Since most people in the world to- day are multilingual (Grosjean and Li, 2013), CS is a ubiquitous phenomenon in multilingual com- munities. It goes beyond simple lexical borrow- ing to blending of languages at syntactic, gram- matical and morphological levels (Sitaram et al., 2019). Code-switching has been studied in lin- guistics and sociolinguistics for decades (Poplack, 1980; Gumperz, 1982; Milroy et al., 1995; Auer, 1The code and data is available at https://github. com/TanmayParekh/commonDost Kya tumhare paas koi dost hai who like to eat mangoes? Nahi. Mere kisi friend ko aam pasand nahi Acha…mere 2 dost bank mein kaam karte hai and they like eating apple Mere paas bhi 3 dost hai who like to eat apple INS ALT ALT ALT Figure 1: We present a bilingual dialogue system for human-machine conversations in Hindi-English (Red: Hindi, Blue: English). We discover that humans pos- itively adopt the agent’s code-switching style (ALT and INS) and the language choice for keywords (high- lighted in bold). 2013; Gardner-Chloros and Weston, 2015) since it reveals various linguistic and socio-cultural be- haviours (Heller, 1982). However, NLP studies of written CS are limited to social media texts, rather than natural conversation, and tend to focus on sin- gle sentences, rather than be contextualized in a dialogue (Rabinovich et al., 2019). Advances in dialogue research (Vinyals and Le, 2015; Zhang et al., 2020; Serban et al., 2016) have enabled conversational AI technologies for human– machine interactions, like Alexa and Siri. Although these technologies are pervasive, they still have limited abilities to accommodate to the user, and they do not account for the ubiquity …

Content truncated to 500 words