2020 Theses Doctoral
Identifying and Modeling Code-Switched Language
Code-switching is the phenomenon by which bilingual speakers switch between multiple languages during written or spoken communication. The importance of developing language technologies that are able to process code-switched language is immense, given the large populations that routinely code-switch. Current NLP and Speech models break down when used on code-switched data, interrupting the language processing pipeline in back-end systems and forcing users to communicate in ways which for them are unnatural.
There are four main challenges that arise in building code-switched models: lack of code-switched data on which to train generative language models; lack of multilingual language annotations on code-switched examples which are needed to train supervised models; little understanding of how to leverage monolingual and parallel resources to build better code-switched models; and finally, how to use these models to learn why and when code-switching happens across language pairs. In this thesis, I look into different aspects of these four challenges.
The first part of this thesis focuses on how to obtain reliable corpora of code-switched language. We collected a large corpus of code-switched language from social media using a combination of sets of anchor words that exist in one language and sentence-level language taggers. The newly obtained corpus is superior to other corpora collected via different strategies when it comes to the amount and type of bilingualism in it. It also helps train better language tagging models. We also have proposed a new annotation scheme to obtain part-of-speech tags for code-switched English-Spanish language. The annotation scheme is composed of three different subtasks including automatic labeling, word-specific questions labeling and question-tree word labeling. The part-of-speech labels obtained for the Miami Bangor corpus of English-Spanish conversational speech show very high agreement and accuracy.
The second section of this thesis focuses on the tasks of part-of-speech tagging and language modeling. For the first task, we proposed a state-of-the-art approach to part-of-speech tagging of code-switched English-Spanish data based on recurrent neural networks.Our models were tested on the Miami Bangor corpus on the task of POS tagging alone, for which we achieved 96.34% accuracy, and joint part-of-speech and language ID tagging,which achieved similar POS tagging accuracy (96.39%) and very high language ID accuracy (98.78%).
For the task of language modeling, we first conducted an exhaustive analysis of the relationship between cognate words and code-switching. We then proposed a set of cognate-based features that helped improve language modeling performance by 12% relative points. Furthermore, we showed that these features can also be used across language pairs and still obtain performance improvements.
Finally, we tackled the question of how to use monolingual resources for code-switching models by pre-training state-of-the-art cross-lingual language models on large monolingual corpora and fine-tuning them on the tasks of language modeling and word-level language tagging on code-switched data. We obtained state-of-the-art results on both tasks.
- SotoMartinez_columbia_0054D_15749.pdf application/pdf 1.62 MB Download File
- mets.xml application/xml 12.9 KB Download File
More About This Work
- Academic Units
- Computer Science
- Thesis Advisors
- Hirschberg, Julia
- Ph.D., Columbia University
- Published Here
- February 12, 2020