Mixed Language and Code-Switching in the Canadian Hansard

While there has been lots of interest in code-switching in informal text such as tweets and online content, we ask whether code-switching occurs in the proceedings of multilingual institutions. We focus on the Canadian Hansard, and automatically detect mixed language segments based on simple corpus-based rules and an existing word-level language tagger. Manual evaluation shows that the performance of automatic detection varies significantly depending on the primary language. While 95% precision can be achieved when the original language is French, common words generate many false positives which hurt precision in English. Furthermore, we found that codeswitching does occur within the mixed languages examples detected in the Canadian Hansard, and it might be used differently by French and English speakers. This analysis suggests that parallel corpora such as the Hansard can provide interesting test beds for studying multilingual practices, including code-switching and its translation, and encourages us to collect more gold annotations to improve the characterization and detection of mixed language and code-switching in parallel corpora.