IUCL: Combining Information Sources for SemEval Task 5

Proceedings of the 8th International Workshop on Semantic Evaluation (SemEval 2014), pages 356–360, Dublin, Ireland, August 23-24, 2014. IUCL: Combining Information Sources for SemEval Task 5 Alex Rudnick, Levi King, Can Liu, Markus Dickinson, Sandra K¨ubler Indiana University Bloomington, IN, USA {alexr,leviking,liucan,md7,skuebler}@indiana.edu Abstract We describe the Indiana University sys- tem for SemEval Task 5, the L2 writ- ing assistant task, as well as some exten- sions to the system that were completed after the main evaluation. Our team sub- mitted translations for all four language pairs in the evaluation, yielding the top scores for English-German. The system is based on combining several information sources to arrive at a final L2 translation for a given L1 text fragment, incorporating phrase tables extracted from bitexts, an L2 language model, a multilingual dictionary, and dependency-based collocational mod- els derived from large samples of target- language text. 1 Introduction In the L2 writing assistant task, we must translate an L1 fragment in the midst of an existing, nearly complete, L2 sentence. With the presence of this rich target-language context, the task is rather dif- ferent from a standard machine translation setting, and our goal with our design was to make effec- tive use of the L2 context, exploiting collocational relationships between tokens anywhere in the L2 context and the proposed fragment translations. Our system proceeds in several stages: (1) look- ing up or constructing candidate translations for the L1 fragment, (2) scoring candidate transla- tions via a language model of the L2, (3) scoring candidate translations with a dependency-driven word similarity measure (Lin, 1998) (which we call SIM), and (4) combining the previous scores in a log-linear model to arrive at a final n-best list. Step 1 models transfer knowledge between This work is licenced under a Creative Commons Attribution 4.0 International License. Page numbers and proceedings footer are added by the organizers. License details: http: //creativecommons.org/licenses/by/4.0/ the L1 and L2; step 2 models facts about the L2 syntax, i.e., which translations fit well into the lo- cal context; step 3 models collocational and se- mantic tendencies of the L2; and step 4 gives dif- ferent weights to each of the three sources of in- formation. Although we did not finish step 3 in time for the official results, we discuss it here, as it represents the most novel aspect of the system – namely, steps towards the exploitation of the rich L2 context. In general, our approach is language- independent, with accuracy varying due to the size of data sources and quality of input technology (e.g., syntactic parse accuracy). More features could easily be added to the log-linear model, and further explorations of ways to make use of target- language knowledge could be promising. 2 Data Sources The data sources serve two major purposes for our system: For L2 candidate generation, we use Eu- roparl and BabelNet; and for candidate ranking based on L2 context, we use Wikipedia and the Google Books Syntactic N-grams. Europarl The Europarl Parallel Corpus (Eu- roparl, v7) (Koehn, …

Content truncated to 500 words