Automatic Discovery of Adposition Typology

5830

Authors

nan

Abstract

Natural languages (NL) can be classified as prepositional or postpositional based on the order of the noun phrase and the adposition. Categorizing a language by its adposition typology helps in addressing several challenges in linguistics and natural language processing (NLP). Understand- ing the adposition typologies for less-studied languages by manual analysis of large text corpora can be quite expensive, yet automatic discovery of the same has received very little attention till date. This research presents a simple unsupervised technique to automatically predict the adpo- sition typology for a language. Most of the function words of a language are adpositions, and we show that function words can be effectively separated from content words by leveraging differ- ences in their distributional properties in a corpus. Using this principle, we show that languages can be classified as prepositional or postpositional based on the rank correlations derived from entropies of word co-occurrence distributions. Our claims are substantiated through experiments on 23 languages from ten diverse families, 19 of which are correctly classified by our technique. 1

Keywords

nan

Full Content

Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers, pages 1037–1046, Dublin, Ireland, August 23-29 2014. Automatic Discovery of Adposition Typology Rishiraj Saha Roy IIT Kharagpur Kharagpur, India – 721302. rishiraj@cse.iitkgp.ernet.in Rahul Katare∗ IIT Kharagpur Kharagpur, India – 721302. rah.ykg@gmail.com Niloy Ganguly IIT Kharagpur Kharagpur, India – 721302. niloy@cse.iitkgp.ernet.in Monojit Choudhury Microsoft Research India Bangalore, India – 560001. monojitc@microsoft.com Abstract Natural languages (NL) can be classified as prepositional or postpositional based on the order of the noun phrase and the adposition. Categorizing a language by its adposition typology helps in addressing several challenges in linguistics and natural language processing (NLP). Understand- ing the adposition typologies for less-studied languages by manual analysis of large text corpora can be quite expensive, yet automatic discovery of the same has received very little attention till date. This research presents a simple unsupervised technique to automatically predict the adpo- sition typology for a language. Most of the function words of a language are adpositions, and we show that function words can be effectively separated from content words by leveraging differ- ences in their distributional properties in a corpus. Using this principle, we show that languages can be classified as prepositional or postpositional based on the rank correlations derived from entropies of word co-occurrence distributions. Our claims are substantiated through experiments on 23 languages from ten diverse families, 19 of which are correctly classified by our technique. 1 Introduction Adpositions form a subcategory of function words that combine with noun phrases to denote their se- mantic or grammatical relationships with verbs, and sometimes other noun phrases. NLs can be neatly divided into a few basic typologies based on the order of the noun phrase and its adposition. If the ad- position is placed before the noun phrase, it is called a preposition. Postpositions and inpositions, on the other hand, are adpositions that are placed after and inside noun phrases respectively. If prepositions are predominantly used in the language, for example in English, Bulgarian and Russian, then the language is said to be prepositional. Similarly, Japanese, Hindi and Turkish are some examples of postpositional languages, which predominantly use postpositions. These two are the most commonly found adposition typologies across the globe. Out of 1185 languages analyzed on the World Atlas of Language Structures (WALS)1 (Dryer and Haspelmath, 2011), there are 577 postpositional, 512 prepositional and only 8 in- positional languages. There are a few (30 and 58 respectively) languages which use no or both kinds of adpositions. The order of adpositions is strongly correlated with many other word order typologies. For instance, postpositional languages usually have Object-Verb ordering, whereas prepositional languages have Verb-Object ordering (Greenberg, 1963). Daum´e and Campbell (2007) present a statistical model for automatically discovering such implications from a large typological database and discuss many other typological implications involving adpositions. Motivation. Knowledge of the typological characteristics of languages is not only of interest to lin- guists, but also very useful in NLP for two main reasons. First, typological information, if appropriately …

Content truncated to 500 words

Labeled Text

Natural languages (NL) can be classified as prepositional or postpositional based on the order of the noun phrase and the adposition. Categorizing a language by its adposition typology helps in addressing several challenges in linguistics and natural language processing (NLP). Understanding the adposition typologies for less-studied languages by manual analysis of large text corpora can be quite expensive, yet automatic discovery of the same has received very little attention till date. This research presents a simple unsupervised technique to automatically predict the adposition typology for a language. Most of the function words of a language are adpositions, and we show that function words can be effectively separated from content words by leveraging differences in their distributional properties in a corpus. Using this principle, we show that languages can be classified as prepositional or postpositional based on the rank correlations derived from entropies of word co-occurrence distributions. Our claims are substantiated through experiments on 23 languages from ten diverse families, 19 of which are correctly classified by our technique. 1

Publication Details

Paper ID:
5830
Year:
2014
Publication:
TIPSTER
Publication Type:
conf
DOI:
nan