Improving NLP Systems Using Unconventional, Freely-Available Data

Abstract

Sentence labeling is a type of pattern recognition task that involves the assignment of a categorical label to each member of a sentence of observed words. Standard supervised sentence-labeling systems often have poor generalization: it is difficult to estimate parameters for words which appear in the test set, but seldom (or never) appear in the training set, because they only use words as features in their prediction tasks. Representation learning is a promising technique for discovering features that allow a supervised classifier to generalize from a source domain dataset to arbitrary new domains. We demonstrate that features which are learned from distributional representations of unlabeled data can be used to improve performance on out-of-vocabulary words and help the model to generalize. We also argue that it is important for a representation learner to be able to incorporate expert knowledge during its search for helpful features. We investigate techniques for building open-domain sentence labeling systems that approach the ideal of a system whose accuracy is high and consistent across domains. In particular, we investigate unsupervised techniques for language model representation learning that provide new features which are stable across domains, in that they are predictive in both the training and out-of-domain test data. In experiments, our best system with the proposed techniques reduce error by as much as 11.4% relative to the previous system using traditional representations on the Part-of-Speech tagging task. Moreover, we leverage the Posterior Regularization framework, and develop an architecture for incorporating biases from prior knowledge into representation learning. We investigate three types of biases: entropy bias, distance bias and predictive bias. Experiments on two domain adaptation tasks show that our biased learners identify significantly better sets of features than unbiased learners. This results in a relative reduction in error of more than 16% for both tasks with respect to existing state-of-the-art representation learning techniques. We also extend the idea of using additional unlabeled data to improve the system's performance on a different NLP task, word alignment. Traditional word alignment only takes a sentence-level aligned parallel corpus as input and generates the word-level alignments. However, as the integration of different cultures, more and more people are competent in multiple languages, and they often use elements of multiple languages in conversations. Linguist Code Switching (LCS) is such a situation where two or more languages show up in the context of a single conversation. Traditional machine translation (MT) systems treat LCS data as noise, or just as regular sentences. However, if LCS data is processed intelligently, it can provide a useful signal for training word alignment and MT models. In this work, we first extract constraints from this code switching data and then incorporate them into a word alignment model training procedure. We also show that by using the code switching data, we can jointly train a word alignment model and a language model using co-training. Our techniques for incorporating LCS data improve by 2.64 in BLEU score over a baseline MT system trained using only standard sentence-aligned corpora.