Cross-Cultural Similarity Features for Cross-Lingual Transfer Learning of Pragmatically Motivated Tasks

Abstract

Much work in cross-lingual transfer learning explored how to select better transfer languages for multilingual tasks, primarily focusing on typological and genealogical similarities between languages. We hypothesize that these measures of linguistic proximity are not enough when working with pragmaticallymotivated tasks, such as sentiment analysis. As an alternative, we introduce three linguistic features that capture cross-cultural similarities that manifest in linguistic patterns and quantify distinct aspects of language pragmatics: language context-level, figurative language, and the lexification of emotion concepts. Our analyses show that the proposed pragmatic features do capture cross-cultural similarities and align well with existing work in sociolinguistics and linguistic anthropology. We further corroborate the effectiveness of pragmatically-driven transfer in the downstream task of choosing transfer languages for cross-lingual sentiment analysis.

EnCBP: A New Benchmark Dataset for Finer-Grained Cultural Background Prediction in English

“…As Table 5 shows, the two-stage training strategy brings noticeable performance improvements to the SA models. This agrees with prior psychological research (Sun et al, 2021), since the expressions of sentiments and attitudes differ across culture groups. Similarly, since NEs are usually mentioned differently across cultures, training the model to distinguish culture-specific writing styles helps resolve the conflict between the training domain of BERT and that of the CoNLL-2003 dataset and improves the performance of the NER model.…”

“…Psychological research has revealed that people from different cultural background behave differently in the ways they think (Nisbett et al, 2001), talk (Kim, 2002), write (Krampetz, 2005;Almuhailib, 2019;Kitano, 1990), and express emotions (Hareli et al, 2015;Sun et al, 2021;Acheampong et al, 2020). NLP researchers have applied cultural background information to model differences in linguistic expressions across culture groups especially for psycholinguistic tasks 1 , e.g., distributional perspective identification (Tian et al, 2021) and sentiment analysis (Sun et al, 2021). In prior research, culture groups are usually defined by official language (Tian et al, 2021) (e.g., US, UK, and India are considered part of the same culture group) or, even more coarse-grained, by ideology (Imran et al, 2020) (e.g., "Western" countries and "Eastern" countries).…”

Fairness in Language Models Beyond English: Gaps and Challenges

“…Whilst utilizing linguistic features in methods to evaluate and mitigate gender bias is a relatively new field of study, previous work has demonstrated that additional linguistic context can result in performance gains (Volkova et al, 2013;Wallace et al, 2014), thus in alignment with the claim from Hovy and Yang (2021) that LMs must utilize social context to be able to reach human-level performance on tasks. Sun et al (2021) utilizes linguistic features to capture cross-cultural similarities, and thus, to select languages that are optimal for cross-lingual transfer. However, it is essential to acknowledge that languages are susceptible to cultural and linguistic shifts that occur at both global and local levels over time, as noted in Hamilton et al (2016).…”

Challenges and Strategies in Cross-Cultural NLP

“…Among linguistic reasons, cultural factors (e.g., conceptualisation) can play a role in the mismatch between the spaces. Though hardly ever considered when selecting the source language for model transfer, considering cultural factors improves performance on target languages in pragmatically motivated tasks (Sun et al, 2021).…”