Toshinori Miyoshi

Learning Sentence Ordering for Opinion Generation of Debate

We propose a sentence ordering method to help compose persuasive opinions for debating. In debate texts, support of an opinion such as evidence and reason typically follows the main claim. We focused on this claimsupport structure to order sentences, and developed a two-step method. First, we select from among candidate sentences a first sentence that is likely to be a claim. Second, we order the remaining sentences by using a ranking-based method. We tested the effectiveness of the proposed method by comparing it with a general-purpose method of sentence ordering and found through experiment that it improves the accuracy of first sentence selection by about 19 percentage points and had a superior performance over all metrics. We also applied the proposed method to a constructive speech generation task.

Character Normalization Methods Using Moments of Gradient Features and Normalization Cooperated Feature Extraction

Normalization is a particular important preprocessing operation, and has a large effect on the performance of character recognition. One of the purposes of normalization is to regulate the size, position, and shape of character images so as to reduce within-class shape variations. Among various methods of normalization, moment-based normalizations are known to greatly improve the performance of character recognition. However, conventional moment-based normalization methods are susceptible to the variations of stroke length and/or thickness. In order to alleviate this problem, we propose moment normalization methods that use the moments of character contours instead of character images themselves to estimate the transformation parameters. Our experiments show that the proposed methods are effective particularly for printed character recognition.

End-to-end Argument Generation System in Debating

We introduce an argument generation system in debating, one that is based on sentence retrieval. Users can specify a motion such as This house should ban gambling, and a stance on whether the system agrees or disagrees with the motion. Then the system outputs three argument paragraphs based on “values” automatically decided by the system. The “value” indicates a topic that is considered as a positive or negative for people or communities, such as health and education. Each paragraph is related to one value and composed of about seven sentences. An evaluation over 50 motions from a popular debate website showed that the generated arguments are understandable in 64 paragraphs out of 150.

Simplified polynomial network classifier for handwritten character recognition

Class-specific feature polynomial classifier (CFPC), a variant of a polynomial classifier (PC), yields high classification accuracy especially in high dimensional feature spaces. However, the computational cost for classification in such a high dimensional space is rather expensive. To overcome this difficulty, we propose a simplified polynomial network (SPN) classifier that reduces the complexity of polynomial networks with little deterioration of classification accuracy. In experiments of handwritten digit recognition on USPS, SPN using features of 30.0 dimensions on average achieved higher classification accuracy and a classification speed about 12.8 times faster than CFPC using features of 250 dimensions. In experiments on MNIST, SPN using features of 40.0 dimensions on average achieved a classification speed about 2.0 times faster than CFPC using features of 100 dimensions with nearly the same classification accuracy.

bunji at SemEval-2016 Task 5: Neural and Syntactic Models of Entity-Attribute Relationship for Aspect-based Sentiment Analysis.

This paper describes a sentiment analysis system developed by the bunji team in SemEval2016 Task 5. In this task, we estimate the sentimental polarity of a given entity-attribute (E#A) pair in a sentence. Our approach is to estimate the relationship between target entities and sentimental expressions. We use two different methods to estimate the relationship. The first one is based on a neural attention model that learns relations between tokens and E#A pairs through backpropagation. The second one is based on a rule-based system that examines several verb-centric relations related to E#A pairs. We confirmed the effectiveness of the proposed methods in a target estimation task and a polarity estimation task in the restaurant domain, while our overall ranks were modest.

Moment-Based Character-Normalization Methods Using a Contour Image Combined with an Original Image

Moment-based character-normalization methods are known to improve character-recognition accuracy. These methods use the moments of an input image, which has two dimensions because of the thickness of its stroke lines, to estimate transformation parameters, whereas character is essentially composed of one dimensional stroke lines. This implies that these methods overestimate the moments of the thick parts of character strokes. To solve this problem, moment-based normalization methods, which use the moments of a contour image of a character combined with the input image of that character, are proposed. To extract the contours of character strokes, two methods, chain code contour (CC) and gradient contour (GC), are used. Character-recognition experiments on two printed-character databases and on two handwritten-character databases show that the character-recognition accuracies of the proposed methods are comparable to or significantly higher than those of conventional methods. In particular, the proposed methods are more effective for printed-character recognition.

Severity estimation of finger-tapping caused by Parkinson's disease by using linear discriminant regression analysis

We propose a linear discriminant regression analysis (LDRA) that provides an estimated severity marker for discriminating between healthy and patient groups and estimating severities of the patient group simultaneously. This method combines an evaluation function for discriminating between two groups and one for estimating severities of one group. The combined function is optimized to obtain an equation for calculating estimated severities. The method was evaluated with finger-tapping data of healthy and Parkinsons disease (PD) groups and PD severities assessed by a doctor. As a result, the discrimination ability of LDRA (AUC: 0.8835) was higher than that of discriminant analysis (DA. AUC: 0.8442), which is a conventional method for classification, and the regression ability of LDRA (mean square error (MSE): 1.27) was superior to that of multiple regression analysis (MRA. MSE: 1.68), which is a conventional method for regression. The results show that LDRA is an effective method for estimating the presence and severity of Parkinsons disease.