This paper presents a method to predict retrieval terms from relevant/surrounding words or descriptive texts in Japanese by using deep belief networks (DBN), one of two typical types of deep learning. To determine the effectiveness of using DBN for this task, we tested it along with baseline methods using example-based approaches and conventional machine learning methods, i.e., multi-layer perceptron (MLP) and support vector machines (SVM), for comparison. The data for training and testing were obtained from the Web in manual and automatic manners. Automatically created pseudo data was also used. A grid search was adopted for obtaining the optimal hyperparameters of these machine learning methods by performing cross-validation on training data. Experimental results showed that (1) using DBN has far higher prediction precisions than using baseline methods and higher prediction precisions than using either MLP or SVM; (2) adding automatically gathered data and pseudo data to the manually gathered data as training data is an effective measure for further improving the prediction precisions; and (3) DBN is able to deal with noisier training data than MLP, i.e., the prediction precision of DBN can be improved by adding noisy training data, but that of MLP cannot be.

View full abstract

In this article, we present an incremental dependency parsing algorithm with an arc-eager variant of the left-corner parsing strategy. Our algorithm’s stack depth captures the center-embeddedness of the recognized dependency structure. A higher stack depth occurs only when processing deeper center-embedded sentences in which people find difficulty in comprehension. We examine whether our algorithm can capture the syntactic regularity that universally exists in languages through two kinds of experiments across treebanks of 19 languages. We first show through oracle parsing experiments that our parsing algorithm consistently requires less stack depth to recognize annotated trees relative to other algorithms across languages. This result also suggests the existence of a syntactic universal by which deeper center-embedding is a rare construction across languages, a result that has yet to be quantitatively cross-linguistically examined. We further investigate the above claim through supervised parsing experiments and show that our proposed parser is consistently less sensitive to constraints on stack depth bounds when decoding across languages, while the performance of other parsers such as the arc-eager parser is largely affected by such constraints. We thus conclude that the stack depth of our parser represents a more meaningful measure for capturing syntactic regularity in languages than those of existing parsers.

View full abstract

This paper proposes a novel word alignment model based on a recurrent neural network (RNN), in which an unlimited alignment history is represented by recurrently connected hidden layers. In addition, we perform unsupervised learning inspired by (Dyer et al. 2011), which utilizes artificially generated negative samples. Our alignment model is directional, like the generative IBM models (Brown et al. 1993). To overcome this limitation, we encourage an agreement between the two directional models by introducing a penalty function, which ensures word embedding consistency across two directional models during training. The RNN-based model outperforms both the feed-forward NN-based model (Yang et al. 2013) and the IBM Model 4 under Japanese-English and French-English word alignment tasks, and achieves comparable translation performance to those baselines under Japanese-English and Chinese-English translation tasks.

View full abstract