The NL2KR Platform for building Natural Language Translation Systems

Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing, pages 899–908, Beijing, China, July 26-31, 2015. c⃝2015 Association for Computational Linguistics The NL2KR Platform for building Natural Language Translation Systems Nguyen H. Vo, Arindam Mitra and Chitta Baral School of Computing, Informatics and Decision Systems Engineering Arizona State University {nguyen.h.vo, amitra7, chitta }@asu.edu Abstract This paper presents the NL2KR platform to build systems that can translate text to different formal languages. It is freely- available1, customizable, and comes with an Interactive GUI support that is use- ful in the development of a translation system. Our key contribution is a user- friendly system based on an interactive multistage learning algorithm. This effec- tive algorithm employs Inverse-λ, Gener- alization and user provided dictionary to learn new meanings of words from sen- tences and their representations. Using the learned meanings, and the Generaliza- tion approach, it is able to translate new sentences. NL2KR is evaluated on two standard corpora, Jobs and GeoQuery and it exhibits state-of-the-art performance on both of them. 1 Introduction and Related Work For natural language interaction with systems one needs to translate natural language text to the input language of that system. Since different systems (such as a robot or database system) may have dif- ferent input language, we need a way to translate natural language to different formal languages as needed by the application. We have developed a user friendly platform, NL2KR, that takes exam- ples of sentences and their translations (in a de- sired target language that varies with the applica- tion), and some bootstrap information (an initial lexicon), and constructs a translation system from text to that desired target language. 1http://nl2kr.engineering.asu.edu/ Our approach to translate natural language text to formal representation is inspired by Montague’s work (Montague, 1974) where the meanings of words and phrases are expressed as λ-calculus ex- pressions and the meaning of a sentence is built from semantics of constituent words through ap- propriate λ-calculus (Church, 1936) applications. A major challenge in using this approach has been the difficulty of coming up with the λ-calculus representation of words. Montague’s approach has been widely used in (Zettlemoyer and Collins, 2005; Kwiatkowski et al., 2010) to translate natural language to formal languages. In ZC05 (Zettlemoyer and Collins, 2005) the learning algorithm requires the user to provide the semantic templates for all words. A semantic template is a λ-expression (e.g. λx.p(x) for an arity one predicate), which describes a par- ticular pattern of representation in that formal lan- guage. With all these possible templates, the learning algorithm extracts the semantic represen- tation of the words from the formal representa- tion of a sentence. It then associates the extracted meanings to the words of the sentence in all possi- ble ways and ranks the associations according to some goodness measure. While manually com- ing up with semantic templates for one target lan- guage is perhaps reasonable, manually doing it for different target languages corresponding …

Content truncated to 500 words