Twitter Users #CodeSwitch Hashtags! #MoltoImportante #wow

David Jurgens, Stefan Dimitrov, Derek Ruths

Abstract

This study investigates the phenomenon of code-switching—the practice of mixing multiple languages—within Twitter hashtags. By analyzing a global dataset of nearly one billion tweets and focusing on bilingual hubs like Montreal and Paris, the researchers explore how hashtags function as a linguistic "side channel." The findings reveal that while hashtags often facilitate intentional multilingual expression, a significant portion of cross-lingual hashtag usage is driven by automated applications and spam. Furthermore, the study identifies a "bilingualism paradox" where users in less bilingual environments may use code-switched hashtags more frequently as a stylistic choice.

Introduction

Code-switching (CS) is a well-documented linguistic behavior in spoken and formal written contexts, but its dynamics in "microtext" platforms like Twitter remain under-explored. A critical gap exists in understanding the role of the hashtag. Unlike body text, hashtags often serve as a metadata layer or commentary. For instance, an English post tagged with #yoquiero (Spanish for "I want it") adds cultural or emotional nuance. This research aims to distinguish between "true" intentional code-switching by humans and "spurious" adoption caused by global brands, spam bots, or automated third-party applications.

Methods and Data

The researchers employed a two-pronged experimental approach:

  1. Global Large-Scale Analysis: Nearly one billion tweets (March–July 2014) were processed to identify hashtags used across different languages. This was filtered to 19.4 million posts covering 4,624 unique hashtags appearing at least 1,000 times.
  2. Taxonomy Development: To categorize hashtag usage, the authors developed a taxonomy of eight distinct types, ranging from "Annotation" to "Spam."
Table 1: Taxonomy of hashtag types used to categorize linguistic behavior.
  1. Bilingual City Comparison: The study focused on Montreal, Quebec City, and Paris—cities with high French-English bilingualism—to compare how users incorporate hashtags from a secondary language. Language identification was performed using langid.py and validated against aspell dictionaries.

Key Insights and Results

Hashtags as "Side Channels"

The analysis confirms that hashtags often act as a linguistic bridge, allowing users to comment on their own text in a different language. However, the most "multilingual" hashtags (appearing in the most languages) are rarely examples of genuine code-switching. Tags like #lastfm or #job are typically generated by software, highlighting the risk of relying on raw frequency counts in digital linguistics.

The Bilingualism Paradox

A counter-intuitive finding emerged from the city-specific data: French speakers in Paris used English hashtags at a higher rate than French speakers in Montreal.

Figure 1: Comparison of cross-lingual hashtag usage in bilingual cities.

In Paris, approximately 13% of French tweets included an English hashtag, whereas in Montreal and Quebec City, this rate was roughly 7%. The authors hypothesize that in highly bilingual environments like Montreal, a user wishing to use English is more likely to write the entire tweet in English. In contrast, a Parisian might use an English hashtag as a stylistic "switch" within a predominantly French tweet.

Emotional Resonance

Manual validation revealed that "Annotation" hashtags (e.g., #fail, #magic) and "Community" hashtags are the primary drivers of genuine code-switching. These tags allow for emotional expression that transcends the primary language of the post.

Conclusion and Future Work

While this study provides a robust framework for analyzing microtext code-switching, limitations remain. Language identification in short, slang-heavy microtext is inherently noisy. Future research could leverage automated classifiers for the hashtag taxonomy and incorporate longitudinal user data to better understand the social context of bilinguality. By examining a user's historical data and social network, researchers can more accurately distinguish between stylistic switching and functional language use.

Summary updated

BibTeX
@inproceedings{jurgens2014twitter,
  author = {David Jurgens and Stefan Dimitrov and Derek Ruths},
  title = {{Twitter Users \#CodeSwitch Hashtags! \#MoltoImportante \#wow}},
  booktitle = {CodeSwitch @ EMNLP},
  year = {2014},
  doi = {10.3115/v1/w14-3906},
}
Figure