All Words Unsupervised Semantic Category Labeling for Hindi

International Conference RANLP 2009 - Borovets, Bulgaria, pages 365–369 All Words Unsupervised Semantic Category Labeling for Hindi Siva Reddy, Abhilash Inumella, Rajeev Sangal, Soma Paul Language Technologies Research Center International Institute of Information Technology, Gachibowli, Hyderabad, India - 500032 {gvsreddy, abhilashi}@students.iiit.ac.in, {sangal, soma}@iiit.ac.in Abstract In the task of semantic category labeling, given a text, every word in it has to be assigned a semantic category. Our language of interest is Hindi. We use the ontolog- ical categories defined in Hindi Wordnet as semantic category inventories. In this paper we present two un- supervised approaches namely Flat Semantic Category Labeler (FSCL) and Hierarchical Semantic Category Labeler (HSCL ). The former method treats semantic categories as a flat list, whereas the latter one exploits the hierarchy among the semantic categories in a top down manner. Further our methods use simple proba- bilistic models, using which the category labeling be- comes a simple table look up with little extra compu- tation and thus opening the possibility of it’s use in real-time interactive systems. Keywords word sense disambiguation, Semantic category labeling, unsupervised 1 Introduction Given a word, its admissible semantic categories and its context, the task of semantic category labeling is to assign the most appropriate semantic category to the word. Our language of interest is Hindi.1An example is shown in Ta- ble 1. We use Hindi Wordnet Ontological categories [5] as semantic category inventories rather than Wordnet synsets. We justify the reason behind this later. 1.1 Ontological Categories Hindi Wordnet’s Ontology is a hierarchical organization of concepts. A separate ontological hierarchy exists for each syntactic category (noun, verb, adjective adverb). Num- ber of nodes (categories) in noun, verb, adjective and ad- verb hierarchy are 101, 31, 25 and 11 respectively. The maximum depth of the hierarchy is 5. Each synset of the wordnet is mapped to some place in the hierarchy. Figure 1 depicts an example showing the mapping from the synsets of the word billA to ontological categories. 1.2 Ontological Categories versus Synsets ? During manual annotation of few hindi sentences, we found that the inter annotator agreements were more when 1 Hindi is the official language of India. Urdu is a close cousin to Hindi. Hindi and Urdu are spoken by approximately 500 million people in the world. Fig. 1: Hindi wordnet entry of the word billA. The word has two senses meaning male cat and badge. Ontological cate- gory mappings of the two senses are shown on the left side of the figure. On the right, the semantic category tree(SCT) of the word is shown. we used ontological categories compared to synsets. This is because various synsets (fine grained) are mappped to the same ontological category (coarse grained). A simi- lar behavior was observed for English in earlier works. In the English Lexical Sample Task [7, 11] of Senseval-2, the inter annotator agreement of verbs rose to 82% using the grouped senses (coarse) from 73% using Wordnet 1.7 un- grouped senses. Ramakrishnan et al. [4] states that, sense disambiguation systems …

Content truncated to 500 words