Generating Code-Switched Text from Monolingual Text with Dependency Tree

Generating Code-Switched Text from Monolingual Text with Dependency Tree Bryan Gregorius School of Science and Technology Kwansei Gakuin University Hyogo, Japan contact@selubi.tech Takeshi Okadome School of Science and Technology Kwansei Gakuin University Hyogo, Japan tokadome@acm.org Abstract Various methods have been proposed to generate code-switched texts. Many of these involve training neural networks and, in turn, require some (albeit small) amounts of code-switched texts or parallel corpora to train the model itself. In this pa- per, we propose a method to convert mono- lingual text into a bilingual code-switched sentence using a dependency parser and machine translator. We leverage the char- acteristics of the dependency tree to iden- tify the switching point and then pass it to machine translation to generate the code-switched sentence. We then surveyed multilingual people of respective language pairs to review the generated sentences and categorize the result. We found that our method is capable of generating nat- ural code-switched text for various lan- guage pairs with the same algorithm. Our method does not require training and thus does not require training data. Our imple- mentation of the model uses off-the-shelf components. The implementation is also built with the possibility of using purpose- built components and rapid deployability in mind. 1 Introduction Code-Switching (CS) is a phenomenon where a speaker alternates between two or more lan- guages in a single conversation. This phe- nomenon is frequently observed in multilin- gual communities, with sentences alternating between a base language and one or more in- serted languages. One of the reasons for do- ing CS is to clarify important information that cannot be explained in one language or code. 1.An example of this can be seen in Table Here, while the English translation ap- pears natural, the concept “内々定” (informal promise of employment) does not exist in En- glish. Therefore, the word “offer” is a good substitute. However, for people who under- stand Japanese, “内々定” provides more con- text. In this example, the base language that provides the grammatical structure is English, and the inserted language is Japanese. CS-related research is integral to Natural Language Processing (NLP) research, as it helps us understand how multilingual people use and understand languages. Most NLP- related corpora are an aggregation of scripts taken from books, movies, and other media. However, those media are primarily targeted at a particular demographic and, thus, mostly monolingual. This makes CS-related corpus scarce and CS corpus generation a research topic of interest. Furthermore, CS corpus gen- eration is but a step in building CS language models. As such, there is a demand for rapidly deployable CS sentence generators. In this paper, we propose a method to gener- ate CS sentences from monolingual sentences. Our model is designed to work on various lan- guage pairs. Our model implementation is eas- ily expandable to other language pairs and is made with the possibility of being used in tan- dem with custom components in mind. To our knowledge, we are the first to build a highly extensible code-switched generator that …

Content truncated to 500 words