Hiroyuki Sakai

In this paper, we propose a method that extracts positive comments (e.g. "もっと評価されるべき"(motto hyouka sarerubeki: should be valued) from Nico Nico Douga automatically. For example, positive comments extracted by our method are beneficial for analysis of campaign broadcasts.

In this paper, we propose a method that annotates tags corresponding to search queries in our company search system. These tags are estimated by our method automatically. For example, if a search query is "textmining", our method estimates "development", "offer" and "introduction" as tags. Then, our company search system is able to retrieve companies that develops "textminig" technology by using the estimated tags.

In recent years, animation sites, such as animation share services and animation distribution services, have spread by development of the Internet. Our focus is to associate an animation and the Nikkei Press-Releases via animation's tags and press-release's words.

In this paper, we introduce a method of extracting character's impressions from Twitter. We focus Yuru-chara that is one of the mascot characters and extract impression expressions (e. g., "Pretty". "Heal") corresponding to Yuru-chara's names. We make an impression expressions dictionary which consists of 363 impression expressions classified into "delighted", "angry", "lament", "frightened", "embarassed", "like", "annoyed", "excited", "amazed", "reliable""entertain""other". Our method acquires Tweets containing the impression expressions and the Yuru-chara's names and extracts impression expressions each Yuru-chara from the Tweets. Finally, we evaluated our method.

In this paper, we propose a method for extracting causal information from PDF files of the summary of financial statements of companies, e.g., the sales of smart phones was expanded continually. Our method extracts causal information form a PDF file of the summary of financial statements of a company by using clue expressions (e.g., "was good") and keywords relevant to the company. The clue expressions are extracted from articles concerning business performance of the financial press report automatically and keywords relevant to a company are extracted from the Web pages of the company.

Currently, the number of private investors is increasing. However, all private investors don't have the knowledge on investment. Therefore, we proposed a method for extracting important articles that influence the stock price of companies from financial articles. Actually, we calculate change of stock price from previous day. Then, we extract newspaper articles related to companies from set of articles issued on the day. We create a training data set from extracted newspaper articles. Our method extracts features from words contained in the training data set, and classifies the extracted articles into important articles that influence the stock price of companies and unimportant articles by SVM. Experimental results showed that our method achieved 44 F-measure. Moreover, we adds keywords extracted from web pages of companies to features of SVM for improving F-measure. In this result, our method achieved 57 F-measure and 73% accuracy.

In this paper, we propose a method for extracting keywords related to business enterprise from web pages of the companies. We define keywords to be extracted as business enterprise or goods of the companies. We collect web pages of companies and our method extracts keywords from the web pages by using document frequencies of words contained in the web pages.

We propose a method for determining causal relation in newspaper sentences with clue expressions. Clue expressions indicating causal relations sometimes have mean other than causal relation. For example, clue expression "から"(kara:from) means causal relation or other mean (e.g. start points). Therefore, we need to exclude clue expressins that mean other than causal relation for extracting causal knowledge correctly. Our method uses expanded linguistic ontology as semantic feature and extracts additional learning data automatically. As a result, performance of our method is improved.

We propose a method Cross-Bootstrapping for extracting solution-effect expressions from patent documents automatically. Solution-effect expressions are useful for generating patent maps. Our method extracts expressions using two clues and statistical information via a bootstrapping method. Furthermore, the advantage of our method is to extract expressions without dictionaries and patterns given manually. Finally, we experimented our method. As a result, our method achieved sufficient performance for generating patent maps.

In this research, we propose a method to automatically extract basis expressions that indicate economic trends from newspaper articles by using a statistical method. We also propose a method to classify them into positive expressions that indicate upbeat, and negative expressions that indicate downturn in economy. It is important to foresee the economic trends for companies, governments and investors to forecast stock places and sales of goods. Therefore, we considered that the companies, governments and investors are able to forecast economic trends by using basis expressions extracted from newspaper articles concerning economic trends. In this time, we extracted basis expressions, and classified them into positive expressions or negative expressions as information to forecast economic trends. We evaluated our methods using NIKKEI newspaper articles from 1990 to 2005. The results showed that the method to extract basis expressions, attained presicion 71.43% and recall 33.33%. The classification method attained F-measure with positive expressions 0.695, and F-measure with negative expressions 0.849.

In this research, we propose a method to automatically extract basis expressions that indicate economic trends from newspaper articles by using a statistical method. We also propose a method to classify them into positive expressions that indicate upbeat, and negative expressions that indicate downturn in economy. It is important to foresee the economic trends for companies, governments and investors to forecast stock places and sales of goods. Therefore, we considered that the companies, governments and investors are able to forecast economic trends by using basis expressions extracted from newspaper articles concerning economic trends. In this time, we extracted basis expressions, and classified them into positive expressions or negative expressions as information to forecast economic trends. We evaluated our methods using NIKKEI newspaper articles from 1990 to 2005. The results showed that the method to extract basis expressions, attained presicion 71.43% and recall 33.33%. The classification method attained F-measure with positive expressions 0.695, and F-measure with negative expressions 0.849.

Many articles about each company are distributed on the newspaper or Internet in a day. However, an important article for human is an article containing a story that influences the corporate performance. In this research, we propose a method for identifying an article containing a story that influences the corporate performance and extracting such articles from a newspaper corpus. Our method judges whether the story contained in the extracted article is positive or negative to the corporate performance. Moreover, we target the articles of the announcement on the corporate performance, we propose a method for extracting sentences containing its key factor(good business or bad business). Experimental results showed that our method for extracting articles containing a story that influences the corporate performance attained 85.8% precision and 66.8% recall and our method for extracting sentences containing the key factor attained 82.2% precision and 26.3% recall.

Many articles about each company are distributed on the newspaper or Internet in a day. However, an important article for human is an article containing a story that influences the corporate performance. In this research, we propose a method for identifying an article containing a story that influences the corporate performance and extracting such articles from a newspaper corpus. Our method judges whether the story contained in the extracted article is positive or negative to the corporate performance. Moreover, we target the articles of the announcement on the corporate performance, we propose a method for extracting sentences containing its key factor(good business or bad business). Experimental results showed that our method for extracting articles containing a story that influences the corporate performance attained 85.8% precision and 66.8% recall and our method for extracting sentences containing the key factor attained 82.2% precision and 26.3% recall.

It is impossible to simply interpret a link on a Webpage as "confidence to the linked page" due to the increase in the amount of information on the Web, and the availability of simple information dispatch tools, such as Weblog. By getting to know the feeling implicitly expressed by the link on the Web, we think that it becomes possible to find more densely cooperated Web Communities. Thus, in this research, we tried to estimate a sentiment polarity score by supervised learning to the link in Weblog. By taking into consideration the deviation of expression between in a positive document and in a negative document, and document structure peculiar to Weblog, we found that about 0.45 can be conjectured with remarkable correlation by Pearson's correlation coefficient.

We propose a method of acquiring knowledge about the possibility of deletion of adnominal verb phrases from a corpus. Our method acquires such items as the frequency of modification of the noun by adnominal verb phrases and the variety of adnominal verb phrases modifying the noun, from a newspaper corpus, and yields knowledge about the possibility of deletion of adnominal verb phrases by using their statistical information. Our method also identifies adnominal verb phrases modifying nouns with important content as deletable ones. Experimental results show that the precision and the recall attained 77.3% and 69.4%, respectively. © 2006 Wiley Periodicals, Inc.

We propose a method for extracting articles concerning traffic accident and expressions concerning accident cause (e.g., "mishandling of the steering wheel control") from a newspaper corpus. Our method extracts articles concerning traffic accident from a newspaper corpus by using SVMs, and extracts expressions concerning accident cause from the extracted articles. Here, we define an expression modified by expressions concerning accident cause as "a seed expression". Our method acquires expressions concerning accident cause from an initial seed expression provided manually. Moreover, our method acquires seed expressions from the expressions concerning accident cause and acquires new expressions concerning accident cause from the acquired seed expressions. By iterating these processes, expressions concerning accident cause and seed expressions are acquired. Experimental results showed that our method of extraction of articles concerning traffic accident from a newspaper corpus attained 82.0% precison and 84.3% recall. Here, we define a sentence containing both an expression concerning accident cause and a seed expression as a cause sentence and the precision and the recall of extraction of cause sentences attained 77.7% and 39.8%, respectively.

We propose a method for extracting articles concerning traffic accident and expressions concerning accident cause (e.g., "mishandling of the steering wheel control") from a newspaper corpus. Our method extracts articles concerning traffic accident from a newspaper corpus by using SVMs, and extracts expressions concerning accident cause from the extracted articles. Here, we define an expression modified by expressions concerning accident cause as "a seed expression". Our method acquires expressions concerning accident cause from an initial seed expression provided manually. Moreover, our method acquires seed expressions from the expressions concerning accident cause and acquires new expressions concerning accident cause from the acquired seed expressions. By iterating these processes, expressions concerning accident cause and seed expressions are acquired. Experimental results showed that our method of extraction of articles concerning traffic accident from a newspaper corpus attained 82.0% precison and 84.3% recall. Here, we define a sentence containing both an expression concerning accident cause and a seed expression as a cause sentence and the precision and the recall of extraction of cause sentences attained 77.7% and 39.8%, respectively.

In this paper, we report about knowledge acquisition of paraphrasing katakana words adopted from English for improving readability. Recently, the problem of using difficult katakana words in public documents has been drawing much attention. Mostly these words are from English. Moreover, they often have synonyms in Japanese. However, it is impossible to obtain rules manually because of time and effort. Therefore, we propose a method of paraphrasing katakana to Japanese words automatically. Experimental results show that recall 82.8% and precision 81.1% are attained for English words restoration, recall 14.6% and precision 70.8% are attained for acquisition of paraphrasing.

In this paper, we report about knowledge acquisition of paraphrasing katakana words adopted from English for improving readability. Recently, the problem of using difficult katakana words in public documents has been drawing much attention. Mostly these words are from English. Moreover, they often have synonyms in Japanese. However, it is impossible to obtain rules manually because of time and effort. Therefore, we propose a method of paraphrasing katakana to Japanese words automatically. Experimental results show that recall 82.8% and precision 81.1% are attained for English words restoration, recall 14.6% and precision 70.8% are attained for acquisition of paraphrasing.

This paper proposes an unsupervised method of acquiring knowledge about the abbreviation possibility of multiplex verb/noun phrases. Our method calculates weight values of verb/noun phrases and the weight values are calculated by mutual information based on degree of combination of verb/noun phrases and modified verbs/nouns. And, our method recognizes verb/noun phrases possible to be abbreviated by comparing weight values of verb/noun phrases. The evaluation of our method by experiments shows that the precision attains 73.4% and the recall attains 45.9% in case of noun phrases and the precision attains 74.7% and the recall attains 42.9% in case of verb phrases.

This paper proposes an unsupervised method of acquiring knowledge about the abbreviation possibility of multiplex verb/noun phrases. Our method calculates weight values of verb/noun phrases and the weight values are calculated by mutual information based on degree of combination of verb/noun phrases and modified verbs/nouns. And, our method recognizes verb/noun phrases possible to be abbreviated by comparing weight values of verb/noun phrases. The evaluation of our method by experiments shows that the precision attains 73.4% and the recall attains 45.9% in case of noun phrases and the precision attains 74.7% and the recall attains 42.9% in case of verb phrases.