In general, arguments are located near the predicate. A previous study has exploited this characteristic to group candidates by positional relations between a predicate and its candidate arguments and then searched for the final candidate using a predetermined priority list of the groups. However, in such an analysis, candidates in different groups cannot be compared. Therefore, we propose a Japanese predicate argument structure analysis model that gathers the most likely candidates from all the groups and then selects the final candidate amongst them. We can account for candidates with less priority before making a final decision to perform global optimization. Experimental results show that our model outperforms deterministic models.

View full abstract

This paper presents a simple yet effective approach to sentence-level uncertainty detection which does not require cue word annotation. Unlike previous works, the proposed method focuses on cue selection, decoupling it from disambiguation and by optimizing it over sentence hedging error rate. High performance for the task is achieved in experiments, even for settings with poor disambiguation, without cue annotation and with otherwise unreliable corpora from a machine learning point-of-view.

View full abstract

The main challenge in hierarchical multi-label document classification is the means by which hierarchically organized labels are leveraged. In this paper, we propose to exploit dependencies among multiple labels to be output, which has not been considered in previous studies. To accomplish this, we first formalize the task as a structured prediction problem and propose (1) a global model that jointly outputs multiple labels and (2) a decoding algorithm that finds an exact solution with dynamic programming. We then introduce features that capture inter-label dependencies. Experiments show that these features improve performance while reducing the model size.

View full abstract

In this report, we show that the problem of domain adaptation for word sense disambiguation (WSD) can be treated as a covariate shift problem, and we try to solve it by maximizing the log-likelihood by weighting the probability density ratio, which is the standard solution of covariate shift. The key to solving this problem lies in the estimation of the probability density ratio. We estimate the probability density ratio using simple method employing the Naive Bayes model. In our proposed method, we apply the covariate shift method to the training data expanded by the Daumé’s feature augmentation method. In the experiment, we solve six types of domain adaptations for WSD using three domains, viz., OC (Yahoo! Chiebukuro), PB (Book), and PN (Newspaper) in the BCCWJ corpus. The results show that our proposed method outperforms the Daumé’s method. This report shows that even our simple method of estimating the probability density ratio is effective for use in the covariate shift method. In future, we intend to investigate and find a method of estimating the probability density ratio more accurately. Further, we intend to use the SVM instead of the maximum entropy method. Moreover, the method of covariate shift is also effective for unsupervised domain adaptations and is a promising approach for WSD domain adaptations.

View full abstract