Syllables for Sentence Classification in Morphologically Rich Languages

Syllables for Sentence Classification in Morphologically Rich Languages Madhuri Tummalapalli Language Technologies Research Center KCIS IIIT Hyderabad madhuri.tummalapalli@research.iiit.ac.in Radhika Mamidi Language Technologies Research Center KCIS IIIT Hyderabad radhika.mamidi@iiit.ac.in Abstract Sentence Classification is one of the most fun- damental tasks in NLP, where the aim is to classify a given sentence into a pre-defined set of classes. A lot of work has been done in English in the last few years, which vary in their methodologies. A huge proportion of these works represent the input sentences as a sequence of words in their models. Only a few of them rely on character level repre- sentation. Through this work, we introduce a new method for representing a sentence - as a sequence of syllables. As we show in this work, syllables are a better choice to represent the sub-word level information in a sentence, which is essential for morphologically rich languages. We consider the tasks of Sentiment Analysis and Question Classification in three languages showing varied morphological rich- ness - English, Hindi and Telugu. Through ex- tensive evaluation, we show that syllables are the best performing input type when compared to words or characters for the morphologically rich languages - Hindi and Telugu. 1 Introduction Sentence Classification is one of the most funda- mental tasks in Natural Language Processing, where a given sentence is required to be classified into a pre-defined set of classes. This pre-defined set de- pends upon the specific task under consideration, which could be any of Sentiment Analysis, Ques- tion Classification, Subjectivity Analysis etc. There has been a lot of work done on each of these tasks separately, where the features and techniques cho- sen for classification are specific to the particular task (Silva et al., 2011; Huang et al., 2008; Pang and Lee, 2005a; Kouloumpis et al., 2011). How- ever, the recent popularity and developments in Neu- ral Networks have enabled researchers to build clas- sifier models that are more generic in nature (Kim, 2014; Kalchbrenner et al., 2014). Given sufficient training data, these networks have the ability to learn task specific features for classification. Since these models do not require hand-crafted features, they make an attractive choice for the construction of task-independent or language-independent systems. There are a huge number of languages used in different parts of the world and it is difficult and expensive to build resources for each language. Also, building an end-to-end task-specific language- specific system for every major language is a very tedious and expensive task. We, thus need to build systems which are generic in nature. Most of the systems and techniques that have been built previously rely on word-level input. A few works use character level input (Zhang et al., 2015; Zhang and LeCun, 2015; Tummalapalli et al., 2018), as they show the benefit of being able to capture sub-word level or morphological information and are also helpful in dealing with Out Of Vocabulary (OOV) words. This is especially important in the case of morphological rich languages where a num- …

Content truncated to 500 words