Hiroyuki Sakai

<p>In this research, we propose a method for extracting sentences containing causal information from articles describing the market conditions of the Nikkei Stock Average. The sentences containing causal information are needed to generate market analysis comments. Our method extracts articles describing the market conditions of the Nikkei Stock Average from economic newspaper articles and extracting sentences containing causal information from the extracted articles by deep learning. Here, our method automatically generates the training data necessary to extract the articles describing the market conditions and sentences containing causal information by deep learning and achieved high accuracy. Moreover, our method extracts complementary information of the content described in the causal sentences by using economic causal-chain search.</p>

The proportion of individual investors in security markets has been rising at a rapid rate in Japan in recent times. As a result, the need for technology to support individual investors in making smarter investment decisions is ever increasing. When individual investors make investment decisions, what is the most important information they should hunt for? It is to be able to sense the company's future performance forecast. This is because even if the current performance is in the red (unfavorable) zone, stock prices will tend to rise if it shows that the future performance will recover from the enterprise side. Therefore, in this research, we focus on the financial results as an information source including performance forecasts. The summaries of financial statements include sentences concerning business performance forecast of companies and other associated sentences predicting the upcoming economic conditions. In this study, we extract and classify sentences indicating business performance forecasts and economic forecasts from summaries of financial statements. This allows individual investors to easily grasp the company's performance forecast. Our approach consists of two steps. First, our method extracts sentences relating to 'prediction' by using clue expressions. Second, we automatically classify these prediction oriented sentences into two classes by using deep learning. Our method is capable of automatically generating training data for deep learning and greatly omits efforts to create the training data set by hand.

What would happen if temperatures were subdued and result in a cool summer? One can easily imagine that air conditioner, ice cream or beer sales would be suppressed as a result of this. Less obvious is that agricultural shipments might be delayed, or that sound proofing material sales might decrease. The ability to extract such causal knowledge is important, but it is also important to distinguish between cause-effect pairs that are known and those that are likely to be unknown, or rare. Therefore, in this paper, we propose a method for extracting rare causal knowledge from Japanese financial statement summaries produced by companies. Our method consists of three steps. First, it extracts sentences that include causal knowledge from the summaries using a machine learning method based on an extended language ontology. Second, it obtains causal knowledge from the extracted sentences using syntactic patterns. Finally, it extracts the rarest causal knowledge from the knowledge it has obtained.

In this paper, we propose a method for extracting laboratory front pages from university websites. There are more than 779 universities and colleges in Japan. For selecting a university or a college, some high school students want to know what laboratories these universities or colleges have. To learn about these laboratories, high school students have to search the laboratory front pages from the university websites. However, sometimes it is difficult to find a laboratory front page because they are sometimes buried deep in the hierarchy of university websites. Our method extracts laboratory front pages by using a support vector machine model and applying certain rules. We also developed a laboratory search system that can be used to retrieve laboratory front pages extracted with our method. We evaluated our method and confirmed that is attained 85.0% precision and 65.5% recall.

<p>In this paper, we describe research on applied systems for realizing efficiency of work to store information of notice of annual meeting of shareholders in the database by using text mining technology. We aim to estimate start pages of proposals stated in notice of the meeting of shareholders and classify which proposal the page is. And we developed a system that automatically performs these tasks using text information of the notice of convocation of shareholders, and actually operates it. As a result of comparative experiment between our implemented system and conventional manual work, the working time was shortened to about 1/10. We propose three methods for classifying proposals. The first method classifies proposals by specialized terms extracted from training data. The second method classifies proposals by using deep learning. The final method classifies proposals by extracted proposal title. We evaluated our methods, and the effectiveness of each method was verified. </p>

In this research, we aim to predict start pages of proposals stated in notice of the meeting of shareholders and classify which proposal the page is. We propose two methods that classification method of proposals. The first method heuristically predicts the page on which the proposal is described. Moreover our method extracts specialized terms of each proposal and assigns weights to them. After that, our method classifies proposals by specialized terms. The second method classifies proposals using deep learning. Each methods were evaluated, and the effectiveness of each methods was verified.

In this paper, we propose a method that estimates new tags from user comments on videos on the Nico Nico Douga website. On Nico Nico Douga, users can post tags and comments on videos. However, users cannot post more than 12 tags on a video; therefore, there are some important tags that could be posted but are sometimes missed. We present a technique to acquire some of these missing tags by choosing new tags that score well in a scoring method developed by us.

In this paper, we propose a method of extracting causal information from PDF files of the summary of financial statements of companies, e.g., ”The sales of smart phones was expanded continually”. Cause information is useful for investors in selecting companies to invest. We downloaded 106,885 PDF files of the summary of financial statements of companies fromWeb pages of the companies automatically. Our method extracts causal information from the PDF files by using clue expressions (e.g., ”was expanded”) and keywords relevant to a company. The clue expressions are extracted from the PDF files of the summary of financial statements of companies and articles concerning business performance of companies automatically. We developed the search system which is able to retrieve causal informations extracted by our method. The search system shows causal information containing a keyword inputted by users, and the summary of financial statements containing the retrieved causal information. We evaluated our method and it attained 83.91% precision and 55.04% recall, respectively. Moreover, we compared our method with Sakai et al’s method originally proposed for extracting causal information from financial articles concerning business performance of companies and experimental results showed that our method outperforms Sakai et al’s method.

In this paper, we propose a method that extracts positive comments (e.g. (motto hyouka sarerubeki: should be valued)") and negative comments (e.g. (hetakuso: dub)") from Nico Nico Douga automatically. For example, positive comments extracted by our method are beneficial for analysis of campaign broadcast.

A patent map, a visual representation of related patent information, is an effective strategic tool for analysis of patent application trends. In particular, an effect-technology type patent map is commonly used because patent examinations are based on technology and effect terms of the inventions. However, the patent map creation is a costly task, because most of patent maps are manually created. Therefore, we propose a method for automatic patent map construction. As a first step to develop such a method, we focused on technological terms that represent technologies for realizing the effect. In this step, we develop our method for clustering technological terms, because we need to use clusters of technological terms for our patent maps. The automatic technological terms clustering requires determining the relationship between technological terms. We developed a method to find relationship between technological terms by using a similarity between morphemes of these words as well as existing thesaurus knowledge.

We propose a method to extract a lot of correspondences between questions and answers from a Web message board automatically. We use Web message boards as information sources because Web messasge boards have a lot of articles posted by general users. We extract correspondences between questions and answers that can be used in question answering systems to support natural language sentence input. At first, our proposed method classifies messages of a Web message board into either questions or others. Next, our method extracts a set of root-node pairs from the thread tree of a Web message board, where we define the thread tree when the root is an article classified as a question, and nodes are articles classified as answer candidates. Our method finds correspondences between questions and answers using two clues, (1)similarity between their articles, (2)link count between their articles. We experimented the proposed method, discussed results, and analyzed errors.

We propose a method of assigning polarity to causal information extracted from Japanese financial articles concerning business performance of companies. Our method assigns polarity (positive or negative) to causal information in accordance with business performance, e.g. "zidousya no uriage ga koutyou: (Sales of cars are good)" (The polarity positive is assigned in this example). We may use causal expressions assigned polarity by our method, e.g., to analyze content of articles concerning business performance circumstantially. First, our method classifies articles concerning business performance into positive articles and negative articles. Using them, our method assigns polarity (positive or negative) to causal information extracted from the set of articles concerning business performance. Although Our method needs training dataset for classifying articles concerning business performance into positive and negative ones. our method does not need a training dataset for assigning polarity to causal information. Hence. even if causal information not appearing in the training dataset for classifying articles concerning business performance into positive and negative ones exist, our method is able to assign it polarity by using statistical information of this classified sets of articles. We evaluated our method and confirmed that it attained 74.4% precision and 50.4% recall of assigning polarity positive, and 76.8% precision and 61.5% recall of assigning polarity negative, respectively.

We propose a method for extracting information on the technical effect from a patent document. The information on the technical effect extracted by our method is useful for generating patent maps (see e.g., Figure 1.) automatically or analyzing the technical trend from patent documents. Our method extracts expressions containing the information on the technical effect by using frequent expressions and clue expressions effective for extracting them. The frequent expressions and clue expressions are extracted by using statistical information and initial clue expressions automatically. Our method extracts expressions containing the information on the technical effect without predetermined patterns given by hand, and is expected to be applied to other tasks for acquiring expressions that have a particular meaning (e.g., information on the means for solving the problems) not limited to the information on the technical effect. Our method achieves not only high precision (78.0%) but also high recall (77.6%) by acquiring such clue expressions automatically from patent documents.

We propose a method of assigning polarity to causal information extracted from Japanese financial articles concerning business performance of companies. Our method assigns polarity (positive or negative) according to business performance to causal information, e.g. "zidousya no uriage ga koutyou: (Sales of cars are good)" (The polarity positive is assigned in this example.). First, our method classifies articles concerning business performance into positive articles and negative articles. Using this classified sets of articles, our method assigns polarity (positive or negative) to causal information extracted from the set of articles concerning business performance. We evaluated our method and it attained 75.3% precision and 47.9% recall of assigning polarity positive, and 77.0% precision and 58.5% recall of assigning polarity negative, respectively.

We propose a method of informative DOM* subtree identification from a Web page in an unfamiliar Web site. Our method uses layout data of DOM nodes generated by a generic Web browser. The results show that our method outperforms a baseline method, and was able to identify informative DOM subtrees from Web pages robustly.

We propose a method of extracting cause information from Japanese financial articles concerning business performance. Our method acquires cause information, e.g. (zidousya no uriage ga koutyou: Sales of cars were good)". Cause information is useful for investors in selecting companies to invest. Our method extracts cause information as a form of causal expression by using statistical information and initial clue expressions automatically. Our method can extract causal expressions without predetermined patterns or complex rules given by hand, and is expected to be applied to other tasks for acquiring phrases that have a particular meaning not limited to cause information. We compared our method with our previous one originally proposed for extracting phrases concerning traffic accident causes and experimental results showed that our new method outperforms our previous one.

This paper proposes 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, respectively. It is important for companies, governments and investors to predict economic trends in order to forecast revenue, sales of products, prices of commodities and stock prices. We considered that basis expressions are useful for the companies, governments and investors to forecast economic trends. We extracted basis expressions, and classified them into positive expressions or negative expressions as information to forecast economic trends. Our method used a bootstrap method that was minimally a supervised algorithm for extracting basis expressions. Moreover, our method classified basis expressions into positive expressions or negative ones without dictionaries.

We propose a method of extracting cause information from Japanese newspaper articles concerning business performance. Cause information is useful for investors in selecting companies to invest. Our method extracts cause information as a form of causal expression by using statistical information and initial clue phrases automatically. Our method can extract causal expressions without predetermined patterns or complex rules given by hand, and is expected to be applied to other tasks or language for acquiring phrases that have a particular meaning not limited to cause information. We compared our method with our previous method originally proposed for extracting phrases concerning traffic accident causes and experimental results showed that our new method outperforms our previous one.

At the presentation of his/her research results, it is necessary and indispensable for a researcher to prepare presentation slides for making audience understood his/her research results within limited time. However, making slides requires a lot of time and much effort. Therefore, many researchers hope to prepare slides more efficiently. In this research, we propose a method for automatically generating slides from a LATEX manuscript of a paper, which aims at reduction of researchers&#039; workload. Our method analyzes a LATEX file of a paper and allocates content contained in the paper to slides and generates itemization. In the analysis of a LATEX file, our method uses only necessary information for slide generation and deletes unnecessary information. Here, our method determines necessary information for slide generation by using a typical structure of the LATEX file. In the allocation of content to slides, our method calculates the score of nouns based on frequency, entropy and idf score and extracts importance sentences contained in the paper by using the score and allocates the extracted important sentences to the slides. In generation of itemization, our method generates itemization by using the conjunction that signifies a parallel relation. The reason why our method generates itemization by using the conjunction is as follows: A sentence corresponding to a sentence including the conjunction that signifies a parallel relation may be contained in the paper. Hence, our method generates itemization that consists of the sentence including the conjunction and the sentence corresponding to it. We evaluated our method, and the experimental results showed that our method turned out to be e?ective for generating slides from a LATEX file of a paper.

We propose a multiple-document summarization system with user interaction for coping appropriately with the user&#039;s varying summarization needs. Generally, automatic document summarization is a technology for producing a summary corresponding to a single document. However, in order that a person can better perform intellectual activities, a technology for producing a summary of more than one document (i.e. multiple-document summarization) becomes more important than producing a summary of a single document. Our multiple-document summarization system extracts keywords from the document set to be summarized and displays the k best keywords scored by our system to a user on the screen. From the displayed keywords, the user selects those reflecting the user&#039;s summarization needs. Here, in this paper, we define a user&#039;s &quot;summarization needs&quot; as content in which that particular user is interested. Our multiple-document summarization system produces a summary suitable for a user&#039;s summarization needs by using user-selected keywords. For evaluation of our method, we participated in the TSC3 of NTCIR4, an evaluation workshop for information retrieval and summarization held by National Institute of Informatics. We participated in this workshop by having our system select the 12 best keywords as scored by our system, and our entered system exhibited good performance in content evaluation of multiple-document summarization task. Moreover, we evaluate the effectiveness of user interaction, and the experimental results show that our user interaction system is effective.

This paper proposes a method of acquiring knowledge about the abbreviation possibility of verb phrases. In a certain clause containing a verb and including verb phrases, the proposed method extracts some clauses which contain the same verb and have different case postpositional particles from a large corpus. Then, our method recognizes verb phrases possible to be abbreviated by comparing the verb phrases with the verb phrases contained in the extracted clauses. In our method, the verb phrases containing important piece of information is hard to recognize as being possible to be abbreviated, and the verb phrases containing information which appear in previous sentences is easy to recognize as being possible to be abbreviated. The evaluation of our method by experiments shows that the precision is 78.0% and the recall is 67.9%.We compare our method with the method which recognizes verb phrases possible to be abbreviated by recognizing optional case elements described in a case frame dictionary as being possible to be abbreviated. By the evaluation results, we conclude that our method outperforms the method which recognizes verb phrases possible to be abbreviated by using a case frame dictionary.