@inproceedings{dutta-2022-word,
    title = "Word-level Language Identification Using Subword Embeddings for Code-mixed {B}angla-{E}nglish Social Media Data",
    author = "Dutta, Aparna",
    editor = {S{\"a}lev{\"a}, Jonne  and
      Lignos, Constantine},
    booktitle = "Proceedings of the Workshop on Dataset Creation for Lower-Resourced Languages within the 13th Language Resources and Evaluation Conference",
    month = jun,
    year = "2022",
    address = "Marseille, France",
    publisher = "European Language Resources Association",
    url = "https://preview.aclanthology.org/fix-sig-urls/2022.dclrl-1.10/",
    pages = "76--82",
    abstract = "This paper reports work on building a word-level language identification (LID) model for code-mixed Bangla-English social media data using subword embeddings, with an ultimate goal of using this LID module as the first step in a modular part-of-speech (POS) tagger in future research. This work reports preliminary results of a word-level LID model that uses a single bidirectional LSTM with subword embeddings trained on very limited code-mixed resources. At the time of writing, there are no previous reported results available in which subword embeddings are used for language identification with the Bangla-English code-mixed language pair. As part of the current work, a labeled resource for word-level language identification is also presented, by correcting 85.7{\%} of labels from the 2016 ICON Whatsapp Bangla-English dataset. The trained model was evaluated on a test set of 4,015 tokens compiled from the 2015 and 2016 ICON datasets, and achieved a test accuracy of 93.61{\%}."
}