Minoru Yoshida

Reidentification of Persons Using Clothing Features in Real-Life Video

Person reidentification, which aims to track people across nonoverlapping cameras, is a fundamental task in automated video processing. Moving people often appear differently when viewed from different nonoverlapping cameras because of differences in illumination, pose, and camera properties. The color histogram is a global feature of an object that can be used for identification. This histogram describes the distribution of all colors on the object. However, the use of color histograms has two disadvantages. First, colors change differently under different lighting and at different angles. Second, traditional color histograms lack spatial information. We used a perception-based color space to solve the illumination problem of traditional histograms. We also used the spatial pyramid matching (SPM) model to improve the image spatial information in color histograms. Finally, we used the Gaussian mixture model (GMM) to show features for person reidentification, because the main color feature of GMM is more adaptable for scene changes, and improve the stability of the retrieved results for different color spaces in various scenes. Through a series of experiments, we found the relationships of different features that impact person reidentification.

Emotion predicting method based on emotion state change of personae according to the other's utterances

It is a very difficult task for a dialogue system or robot to predict others emotion considering course of dialogue, although it is natural task for human. We thought that human emotion changes at all times during having a conversation and these changes are caused to our internal emotional state by responding to stimulus from external world. Most recent researches on dialogue system were focusing on superficial feature of utterance aiming to the application for question-answering system, thus, they rarely considered internal emotional state and stimulus from external world. In this paper, we focused on the dialogue sentences in a drama script. By making various annotations that were considered useful to analyze the course of dialogue and the relationships among the characters, we constructed a scenario emotion corpus. We studied how the emotion of a character changes responding to the other characters utterances or emotions by analyzing this corpus. We proposed a method based on the emotion predicting weight to predict the other characters emotional transition from the target characters emotion. As the result of an evaluation experiment, the effectiveness of the proposed method was confirmed.

Extraction Japanese Slang from Weblog Data based on Script Type and Stroke Count

Abstract Young people commonly use slang in the texts for weblogs or Social Networking Sites. How to treat such slang words properly is one of the problems in the field of text mining. In this paper, we examined several methods to extract Japanese slang called “ Wakamono Kotoba ,” which is particularly used by young people, by focusing on its script type and stroke count. In the evaluation experiment, a high precision was obtained when we adopted script type for extraction.

Distributed Representations for Words on Tables

We consider a problem of word embedding for tables, and we obtain distributed representations for words found in tables. We propose a table word-embedding method, which considers both horizontal and vertical relations between cells to estimate appropriate word embedding for words in tables. We propose objective functions that make use of horizontal and vertical relations, both individually and jointly.

Table Topic Models for Hidden Unit Estimation

We propose a method to estimate hidden units of numbers written in tables. We focus on Wikipedia tables and propose an algorithm to estimate which units are appropriate for a given cell that has a number but no unit words. We try to estimate such hidden units using surrounding contexts such as a cell in the first row. To improve the performance, we propose the table topic model that can model tables and surrounding sentences simultaneously.

Emotion recognition for sentences with unknown expressions based on semantic similarity by using Bag of Concepts

In studies of emotion estimation from text, varieties of methods have been attempted such as emotional expression dictionary or sentence structure dictionary and emotion corpus. However, most of these methods targeted the expressions included in the existing morphological analysis dictionaries, as a result, they did not pay enough attention to unknown words, especially newly coined words. In Japan, the growth of Internet communication sites such as weblogs and social networking sites brought younger people especially in teens and in their 20s to create new words and use them very often. We prepared an emotion corpus by collecting weblog article texts including new words, analyzed the corpus statistically, and proposed a method to estimate emotions of the texts. Most slang such as Youth Slang is too ambiguous in sense classification to be registered into the existing dictionaries such as thesaurus. To cope with these words, we created a large scale of Twitter corpus to calculate sense similarity between words. We proposed to convert unknown to sense class id to process the words that were not included in learning data. We defined this as a method using Bag of Concepts as feature. As a result of the evaluation experiment using several classifies, the proposed method was proved robustness for unknown expression.

Unsupervised Analysis of Web Page Semantic Structures by Hierarchical Bayesian Modeling

We propose a Bayesian probabilistic modeling of the semantic structures of HTML documents. We assume that HTML documents have logically hierarchical structures and model them as links between blocks. These links or dependency structures are estimated by sampling methods. We use hierarchical Bayesian modeling where each block is given labels such as “heading” or “contents”, and words and layout features (i.e., symbols and HTML tags) are generated simultaneously, based on these labels.

Judgment of Slang Based on Character Feature and Feature Expression Based on Slang’s Context Feature

Our research aim was to develop the means to automatically identify a particular character string as slang and then connect the detected slang word to words with similar meaning in order to successfully process the sentence in which the word appears. By recognizing a slang word in this way, one can apply different processing to the word and avoid the distinctive problems associated with processing slang words. This paper proposes a method to distinguish standard words from slang words using information from the characters comprising the character string. An experiment testing the effectiveness of our method showed a 30 % or more improvement in classification accuracy compared to the baseline method. We also use a contextual feature related to emotion to expand the unregistered slang word in the training data into other expressions and propose an emotion estimation method based on the expanded expressions. In our experiment, successful emotion estimation was obtained in nearly 54 % of the cases, a notably higher rate than with the baseline method. Our proposed method was shown to have validity.

Refinement by Filtering Translation Candidates and Similarity Based Approach to Expand Emotion Tagged Corpus

Researches on emotion estimation from text mostly use machine learning method. Because machine learning requires a large amount of example corpora, how to acquire high quality training data has been discussed as one of its major problems. The existing language resources include emotion corpora; however, they are not available if the language is different. Constructing bilingual corpus manually is also financially difficult. We propose a method to convert a training data into different language using an existing Japanese-English parallel emotion corpus. With a bilingual dictionary, the translation candidates are extracted against every word of each sentence included in the corpus. Then the extracted translation candidates are narrowed down into a set of words that highly contribute to emotion estimation and we used the set of words as training data. Moreover, when one language’s unannotated linguistic resources can be obtained, the words can be expanded based on the word distributed expression. By using this expressions, we can improve accuracy without decreasing information volume of one sentence. Then, we tried the corpus expansion without translating target linguistic resource. As the result of the evaluation experiment using the machine learning algorithm, we could clear the effectiveness of the emotion corpus which expanded based on the original language’s unannotated sentences and based on similar sentence. Moreover, when large amount of linguistic resources without annotation can be obtained in one language, their words can be expanded based on distributed expressions of the words. By using distributed expressions, we can improve accuracy without decreasing information volume of one sentence. Then, we attempted to expand corpus without translating target linguistic resource. The result of the evaluation experiment using the machine learning algorithm showed the effectiveness of the expanded emotion corpus based on the original language’s unannotated sentences and their similar sentences.

An Approach to Refine Translation Candidates for Emotion Estimation in Japanese-English Language

Researches on emotion estimation from text mostly use machine learning method. Because machine learning n nrequires a large amount of example corpora, how to acquire high quality training data has been discussed as n none of its major problems. The existing language resources include emotion corpora; however, they are not n navailable if the language is different. Constructing bilingual corpus manually is also financially difficult. We n npropose a method to convert a training data into different language using an existing Japanese-English parallel n nemotion corpus. With a bilingual dictionary, the translation candidates are extracted against every word of n neach sentence included in the corpus. Then the extracted translation candidates are narrowed down into a set n nof words that highly contribute to emotion estimation and we used the set of words as training data. As the n nresult of the evaluation experiment using the training data created by our proposed method, the accuracy of n nemotion estimation increased up to 66.7% in Naive Bayes. n n1 INTRODUCTION n nRecently, there have been many researches on emotion n nestimation from text in the field of sentiment n nanalysis or opinion mining (Ren, 2009), (Ren and n nQuan, 2015), (Ren and Wu, 2013), (Quan and Ren, n n2010), (Quan and Ren, 2014), (Ren and Matsumoto, n n2015) and many of them adopted machine learning n nmethods that used words as a feature. When the type n nof the target sentence for emotion estimation and the n ntype of the sentence prepared as training data are different, n nas in the case of terminology in the problem n nof domain adaptation for document classification, the n nappearance tendency of the emotion words differs. n nThis causes a problem in fluctuation of accuracy. On n nthe other hand, when a word is used as a feature for n nemotion estimation, the sentence structure does not n nhave to be considered. As a result, it is easy to apply n nthe method to other languages. Only if we prepare a n nlarge number of corpora with annotation of emotion n ntags on each sentence, emotion would be easily estimated n nby using the machine learning method. In the n nmachine learning method, because manual definition n nof a rule is not necessary, we can reduce costs to apply n nthe method to other languages. n nHowever, just like the problem in the domain, depending n non the