Yoshiyuki Kotani

Question answering system using Q & A site corpus Query expansion and answer candidate evaluation

Question Answering (QA) is a task of answering natural language questions with adequate sentences. This paper proposes two methods to improve the performance of the QA system using a Q&A site corpus. The first method is for the relevant document retrieval module. We proposed modification of measure of mutual information for the query expansion; we calculate it between two words in each question and a word in its answer in the Q&A site corpus not to choose the words that are not suitable.The second method is for the candidate answer evaluation module. We proposed to evaluate candidate answers using the two measures together, i.e., the Web relevance score and the translation probability. The experiments were carried out using a Japanese Q&A site corpus. They revealed that the first proposed method was significantly better than the original method when their accuracies and MRR (Mean Reciprocal Rank) were compared and the second method was significantly better than the original methods when their MRR were compared.

An Othello evaluation function based on Temporal Difference Learning using probability of winning

This paper presents a new reinforcement learning method, called temporal difference learning with Monte Carlo simulation (TDMC), which uses a combination of Temporal Difference Learning (TD) and winning probability in each non-terminal position. Studies on self-teaching evaluation functions as applied to logic games have been conducted for many years, however few successful results of employing TD have been reported. This is perhaps due to the fact that the only reward observable in logic games is their final outcome, with no obvious rewards present in non-terminal positions. TDMC(lambda) attempts to compensate this problem by introducing winning probabilities, obtained through Monte Carlo simulation, as substitute rewards. Using Othello as a testing environment, TDMC(lambda), in comparison to TD(lambda), has been seen to yield better learning results.

Nested Monte-Carlo Search with AMAF Heuristic

Nested Monte-Carlo Search, which calls MonteCarlo search in the nested call, has succeeded in the one-person game named Morpion Solitaire. The depth for the nest is called a level, and the runtime increases exponentially in the search for higher level. In the present study, All-Move-As-First heuristic is incorporated in Nested Monte-Carlo Search and the number of search is reduced to maintain a pseudo number of searches in order to achieve the higher level search. Our system generated a new world record 146 moves of the computer search in Morpion Solitaire touching version by this

Learning face marks for natural language dialogue systems

Face marks figures of faces which consist of characters such as (¿¿), and are effective for expressing emotions in a text-dialogue system. We usually determine face marks from history of emotional elements and actional elements. We propose a method of learning face marks for a natural language dialogue system from chat dialogue data in the Internet, etc. We use a back propagation error learning of a three-layer neural network to learn a model of face marks. In this neural network, the input neurons express emotional parameters and actional categories of texts, and the output neurons express parts of face marks: mouth, eyes, arms, and optional things. The experimental results showed that the learning error was 0.19, and we could get the performance approximately 93% permissible value for the learning set of dialogues and approximately 60% for the evaluation set of dialogues. It also showed that our system acquired the good information of relationship between parts of face marks and emotional and actional elements.

Combining final score with winning percentage by sigmoid function in Monte-Carlo simulations

Monte-Carlo method recently has produced good results in Go. Monte-Carlo Go uses a move which has the highest mean value of either winning percentage or final score. In a past research, winning percentage is superior to final score in Monte-Carlo Go. We investigated them in BlokusDuo, which is a relatively new game, and showed that Monte-Carlo using final score is superior to the one that uses winning percentage in cases where many random simulations are used. Besides, we showed that using final score is unfavorable for UCT, which is the most famous algorithm in Monte-Carlo Go. To introduce the effectivity of final score to UCT, we suggested a way to combine winning percentage and final score by using sigmoid function. We show the effectivity of the suggested method and show that the method improves a bias where Monte-Carlo Go plays very safe moves when it has advantage.

Classification of Japanese onomatopoeias using hierarchical clustering depending on contexts

Japanese has thousands of onomatopoeias and they have recently started to attract a lot of attention of researchers on natural language processing. Some onomatopoeias are semantically or phonologically similar each other and the choice of these onomatopoeias sometimes give a big difference among Japanese sentences. In this paper, the authors classify Japanese onomatopoeias using single link hierarchical clustering depending on contexts such as surrounding words, their part-of-speeches and their meanings, and show explicitly the relationships of them from the perspective of word senses.

Nested Monte-Carlo Search with simulation reduction

The execution time of Nested Monte-Carlo Search for Morpion Solitaire, a single-player game, increases exponentially with the level of the nested search. We investigated the use of two methods for reducing the execution time in order to enable a deeper nested search: simply reducing the number of lower level searches by a constant rate and using All-Moves-As-First heuristic to the reduction in the number of lower level searches. Testing showed the latter is more effective. Using it, we achieved a new world record of 146 moves for a computer search for the touching version of Morpion Solitaire.

Polynomial Time Probabilistic Learning of a Subclass of Linear Languages with Queries

We show a probabilistic learnability of a subclass of linear languages with queries. Learning via queries is an important problem in grammatical inference but the power of queries to probabilistic learnability is not clear yet. In probabilistic learning model, PAC (Probably Approximately Correct) criterion is an important one and many results have been shown in this model. Angluin has shown the ability of replacement from equivalence queries to random examples in PAC criterion but there are also many hardness results. We have shown that the class of simple deterministic languages is polynomial time learnable from membership queries and a representative sample. Also, we have shown that a representative sample can be constructed from polynomial number of random examples with the confidence probability. In this paper, we newly define a subclass of linear languages called strict deterministic linear languages and show the probabilistic learnability with membership queries in polynomial time. This learnability is derived from an exact learning algorithm for this subclass with membership queries, equivalence queries and a representative sample.

Handling time recognition for for friendly natural language interface: a case study about using temporal adverbs in Japanese

A formal method to determine the style of sentences is proposed. Sentences generated by a human are expressed in various ways, depending on his state of mind, even if he utters the same thoughts. Selecting one of the adverbs which is suitable for a situation is one of these problems. Especially in Japanese, what a speaker wants to tell a hearer is sometimes determined only by the kind of adverb. If natural language interfaces were to have the ability to handle styles of sentences or adverbs so as to adapt to a situation, they would become friendly interfaces, because they must surely understand the intention of user. In this paper, we describe the nature of the problem and propose the formalization of sentence expression by using modal logic with temporal information, then we evaluate it by comparing sentences generated by a human and by the system. As a result, it was found that the system could generate sentences which are the same as the human did. Intuitively understandable sentences for friendly natural language interfaces have thus been generated by considering information about the human mind.

Categorization of product pages depending on information on the Web

In this paper, the authors categorize product pages on the Web depending on their information. We used naive Bayes and the complement naive Bayes classifier, and tried four kinds of features to categorize them: all the words of the titles of the product pages, the nouns extracted from the titles, all the words of the titles and the descriptions of the product pages, and the nouns extracted from them. The experiments show that the product pages can be classified most correctly depending on only the nouns of the titles of the product pages. Moreover the complement naive Bayes classifier outperformed the naive Bayes classifier.