Principled Induction of Phrasal Bilexica

Principled Induction of Phrasal Bilexica Markus Saers and Dekai Wu Human Language Technology Center Dept. of Computer Science and Engineering Hong Kong University of Science and Technology Hong Kong {masaers|dekai}@cs.ust.hk Abstract We aim to replace the long and com- plicated, pipeline employed to produce probabilistic phrasal bilexica with a the- oretically principled, grammar based, ap- proach. To this end, we introduce a learning regime to learn a phrasal gram- mar equivalent to linear transduction gram- mars. The stochastic version of this new grammar type also has the property that the set of biterminals constitute a natural probability distribution, making it similar to a probabilistic translation lexicon. Since we learn a phrasal grammar, we are, in ef- fect, learning a probabilistic phrasal bilex- icon. As a proof of concept, we show that phrasal bilexica, induced in this manner, can be used to improve the performance of a traditional phrase-based SMT system. 1 Introduction The aim of the paper is to show that it is possi- ble to replace the long and complicated pipeline that turns a parallel corpus into a phrasal bilexicon, with a theoretically principled, grammar based ap- proach. The ultimate aim, towards which this pa- per is a step, is to eliminate the mismatch between learning and translation that plagues the statistical machine translation SMT systems of today. When phrase-based statistical machine transla- tion (SMT) systems replaced the token-based, a significant boost to translation quality was experi- enced. Rather than having to rely on broad gener- alizations, the system simply memorizes chunks of text and their translation. The degrees of freedom c⃝ 2011 European Association for Machine Translation. to make errors are thus severely restricted, mak- ing the system more accurate at the cost of storing huge amounts of fixed chunks. Since a surface- based system has no mechanism for generalizing in a systematic way, this is a good work-around. Although structured SMT systems are capable of making generalizations beyond the scope of surface-based systems, it is also imperative to be able to handle chunks of text. This is the cor- rect way to capture phenomena such as figures of speech, whose translations go beyond the valid generalizations of the language. However, distin- guishing between the phrasal biterminals that can be handled by generalization and the ones that can- not, is a hard problem. Indeed, finding a can- didate set of phrasal biterminals is a hard prob- lem. Na¨ıvely enumerating all possible phrasal en- tries found in a parallel corpus and determine their probability by relative frequency, is doomed to fail because of the sheer amount of possible phrasal entries. The same is true for constructing transduc- tion grammars with phrasal biterminals: the size of the grammar makes it impossible to handle, which is why transduction grammars are usually thought of as being restricted to handle single-token ter- minals. Although formal proofs are often easier to construct for the normal form of a grammar, the grammar itself is not restricted to this normal form; and although the runtime complexities …

Content truncated to 500 words