Predicting Foreign Language Usage from English-Only Social Media Posts

Abstract

Social media is known for its multi-cultural and multilingual interactions, a natural product of which is code-mixing. Multilingual speakers mix languages they tweet to address a different audience, express certain feelings, or attract attention. This paper presents a largescale analysis of 6 million tweets produced by 27 thousand multilingual users speaking 12 other languages besides English. We rely on this corpus to build predictive models to infer non-English languages that users speak exclusively from their English tweets. Unlike native language identification task, we rely on large amounts of informal social media communications rather than ESL essays. We contrast the predictive power of the state-of-theart machine learning models trained on lexical, syntactic, and stylistic signals with neural network models learned from word, character and byte representations extracted from English only tweets. We report that content, style and syntax are the most predictive of nonEnglish languages that users speak on Twitter. Neural network models learned from byte representations of user content combined with transfer learning yield the best performance. Finally, by analyzing cross-lingual transferthe influence of non-English languages on various levels of linguistic performance in English, we present novel findings on stylistic and syntactic variations across speakers of 12 languages in social media.

Native Language Identification with User Generated Content

“…These are content based features that are highly domain-dependent and are not likely to generalize across domains. Volkova et al (2018) explored the contribution of various (lexical, syntactic, and stylistic) signals for predicting the foreign language of non-English speakers based on their English posts on Twitter. This effectively results in a 12-way classification task, with 12 different L1s (data sizes are distributed very unevenly), and the best results are unsurprisingly obtained with word unigrams and bigrams.…”

Native Language Identification of Fluent and Advanced Non-Native Writers

“…• Feature Space. Previous NLI studies with UGC [9,18,45] rely on (i) the content-specifc features (linguistic features) such as word and character based n-grams, which are not likely to generalize to other domains [9] (see Section 2.2 for more details); and (ii) the social network features which are specifc to the Reddit dataset 3 only [30] and are hard to generalize to other datasets. These features include karma, average score, average number of submissions, average number of comments and most popular subreddits [9].…”

“…These features include karma, average score, average number of submissions, average number of comments and most popular subreddits [9]. Unlike previous NLI studies with UGC [9,18,45], we defne a topic-independent feature space without relying on the social network features and content-specifc features, which makes our solution generalizable to other domains and datasets. Specifcally, our feature space is based on (i) part-of-speech (POS) n-grams, (ii) function words, (iii) context-free grammar production rules (CFG rules), and (iv) the structure of the text sample such as average sentence length (see Section 4.1 for more details).…”

On the Detection of False Information: From Rumors to Fake News

“…We model the presence of the words in one of the following cues categories: assertives verbs [109] , bias [187] , factive verbs [125] , implicative verbs [123] , hedges citehyland2018metadiscourse , report verbs . A previous work has used these bias cues to identify bias in suspicious news posts in Twitter [226] (6 Features).…”

“…We aim to detect IRA trolls by identifying their way of writing English tweets. As shown in [226], English tweets generated by non-English speakers have a different syntactic pattern. Thus, we use SOTA NLI features to detect this unique pattern [45,142,96].…”