Search

Class-based language modeling (LM) is a long-studied and effective approach to overcome data sparsity in the context of n-gram model training. In statistical machine translation (SMT), differ- ent forms of class-based LMs have been shown to improve baseline translation quality when used in combination with standard word-level LMs but no published work has systematically com- pared different kinds of classes, model forms and LM combination methods in a unified SMT setting. This paper aims to fill these gaps by focusing on the challenging problem of translating into Russian, a language with rich inflectional morphology and complex agreement phenomena. We conduct our evaluation in a large-data scenario and report statistically significant BLEU im- provements of up to 0.6 points when using a refined variant of the class-based model originally proposed by Brown et al. (1992).

In this work, we discuss the benefits of using automatically parsed corpora to study language variation. The study of language variation is an area of linguistics in which quantitative methods have been particularly successful. We argue that the large datasets that can be obtained using automatic annotation can help drive further research in this direction, providing sufficient data for the increasingly complex models used to describe variation. We demonstrate this by replicating and extending a previous quantitative variation study that used manually and semi-automatically annotated data. We show that while the study cannot be replicated completely due to limitations of the existing automatic annotation, we can draw at least the same conclusions as the original study. In addition, we demonstrate the flexibility of this method by extending the findings to related linguistic constructions and to another domain of text, using additional data.

Temporal evidence classification, i.e., finding associations between temporal expressions and relations expressed in text, is an important part of temporal relation extraction. To capture the variations found in this setting, we employ a distant supervision approach, modeling the task as multi-class text classification. There are two main challenges with distant supervision: (1) noise generated by incorrect heuristic labeling, and (2) distribution mismatch between the target and distant supervision examples. We are particularly interested in addressing the second problem and propose a sampling approach to handle the distribution mismatch. Our prior-informed distant supervision approach improves over basic distant supervision and outperforms a purely supervised approach when evaluated on TAC-KBP data, both on classification and end-to-end metrics.