“nee intention enti?” towards dialog act recognition in code-mixed conversations

Abstract

Code-Mixing (CM) is a very commonly observed mode of communication in a multilingual configuration. The trends of using this newly emerging language has its effect as a culling option especially in platforms like social media. This becomes particularly important in the context of technology and health, where expressing the upcoming advancements is difficult in native language. Despite the change of such language dynamics, current dialog systems cannot handle a switch between languages across sentences and mixing within a sentence. Everyday conversations are fabricated in this mixed language and analyzing dialog acts in this language is very essential in further advancements of making interaction with personal assistants more natural. The problem is further compounded with crossing the script barriers in code-mixing. In this paper we take the first step towards understanding code-mixing in dialog processing, by recognizing dialog act (intention) of the code-mixed utterance. Considering the dearth of resources in code-mixed languages, we design our current system using only wordlevel resources such as language identification, transliteration and lexical translation. Our best performing system is HMM based with an F-score of 76.67.