We address the problem of mining code-switched data from the web, where code-switching is defined as the tendency of bilinguals to switch between their multiple languages both across and within utterances. We propose a method that identifies data as code-switched in languages L1 and L2 when a language classifier labels the document as language L1 but the document also contains words that can only belong to L2. We apply our method to Twitter data and collect a set of more than 43,000 tweets. We obtain language identifiers for a subset of 8,000 tweets using crowd-sourcing with high inter-annotator agreement and accuracy. We validate our Twitter corpus by comparing it to the Spanish-English corpus of code-switched tweets collected for the EMNLP 2016 Shared Task for Language Identification, in terms of code-switching rates, language composition and amount of code-switch types found in both datasets. We then trained language taggers on both corpora and show that a tagger trained on the EMNLP corpus exhibits a considerable drop in accuracy when tested on the new corpus and a tagger trained on our new corpus achieves very high accuracy when tested on both corpora.
@InProceedings{MENDELS18.92, author = {Gideon Mendels and Victor Soto and Aaron Jaech and Julia Hirschberg}, title = "{Collecting Code-Switched Data from Social Media}", booktitle = {Proceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC 2018)}, year = {2018}, month = {May 7-12, 2018}, address = {Miyazaki, Japan}, editor = {Nicoletta Calzolari (Conference chair) and Khalid Choukri and Christopher Cieri and Thierry Declerck and Sara Goggi and Koiti Hasida and Hitoshi Isahara and Bente Maegaard and Joseph Mariani and Hélène Mazo and Asuncion Moreno and Jan Odijk and Stelios Piperidis and Takenobu Tokunaga}, publisher = {European Language Resources Association (ELRA)}, isbn = {979-10-95546-00-9}, language = {english} }