Learning to Predict Code-Switching Points

Proceedings of the 2008 Conference on Empirical Methods in Natural Language Processing, pages 973–981, Honolulu, October 2008. c⃝2008 Association for Computational Linguistics Learning to Predict Code-Switching Points Thamar Solorio and Yang Liu Human Language Technology Research Institute The University of Texas at Dallas Richardson, TX 75080, USA tsolorio,yangl@hlt.utdallas.edu Abstract Predicting possible code-switching points can help develop more accurate methods for au- tomatically processing mixed-language text, such as multilingual language models for speech recognition systems and syntactic an- alyzers. We present in this paper exploratory results on learning to predict potential code- switching points in Spanish-English. We trained different learning algorithms using a transcription of code-switched discourse. To evaluate the performance of the classifiers, we used two different criteria: 1) measuring pre- cision, recall, and F-measure of the predic- tions against the reference in the transcrip- tion, and 2) rating the naturalness of artifi- cially generated code-switched sentences. Av- erage scores for the code-switched sentences generated by our machine learning approach were close to the scores of those generated by humans. 1 Introduction Multilingual speakers often switch back and forth between languages when speaking or writing, mostly in informal settings. The mixing of lan- guages involves very elaborated patterns and forms and we usually use the term Code-Switching (CS) to encompass all of them (Lipski, 1978). Before the Internet era, CS was mainly used in its spoken form. But with so many different informal interaction set- tings, such as chats, forums, blogs, and web sites like Myspace and Facebook, CS is being used more and more in written form. For English and Spanish, CS has taken a step further. It has become a hall- mark of the chicano culture as it is evident by the growing number of chicano writers publishing work in Spanish-English CS. We have not completely discovered the process of human language acquisition, especially dual lan- guage acquisition. Findings in linguistics, soci- olinguistics, and psycholinguistics show that the production of code-switched discourse requires a very sophisticated knowledge of the languages be- ing mixed. Some theories suggest bilingual speak- ers might have a third grammar for processing this type of discourse. The general agreement regarding CS is that switches do not take place at random and instead it is possible to identify rules that bilingual speakers adhere to. Understanding the CS process can lead to accu- rate methods for the automatic processing of bilin- gual discourse, and corpus-driven studies about CS can also inform linguistic theories. In this paper we present exploratory work on learning to predict CS points using a machine learning approach. Such an approach can be used to reduce perplexity of lan- guage models for bilingual discourse. We believe that CS behavior can be learned by a classifier and the results presented in this paper support our belief. One of the difficult aspects of trying to predict CS points is how to evaluate the performance of the learner since switching is intrinsically motivated and there are no forced switches (Sankoff, 1998b). Therefore, standard classification measures for …

Content truncated to 500 words