We consider a model composed of nonlinear perceptrons and analytically investigate its generalization performance using correlated examples in the framework of on-line learning by a statistical mechanical method. In Hebbian and AdaTron learning, the larger the number of examples used in an update, the slower the learning. In contrast, Perceptron learning does not exhibit such behaviors, and the learning becomes fast in some time region.

This paper proposes a metric for example matching under the example-based machine translation. Our metric served as similarity measure is employed to retrieve the most similar examples to a given query. Basically it makes use of simple information such as lemma and part-of-speech information of typographically mismatched words. In addition, it uses the contiguity information of matched word units to catch the full context. Finally we show the results for the correctness of the proposed metric.

This paper proposes a new method of example-based query generation for spontaneous speech. Along with modeling the information flows of human dialogues, the authors have designed a system that allows users to retrieve information while driving a car. The system refers to the dialogue corpus to find an example that is similar to input speech, and it generates a query from the example. The experimental results for the prototype system show that 1) for transcribed text input, it provides the correct query in about 64% of cases and the partially collect query in about 88% 2) it has the ability to create correct queries for the utterances not including keywords, compared with the conventional keyword extraction method.

Usually, paintings are more appealing than photographic images. This is because paintings can incorporate styles based on the artist's subjective view of motif. This style can be distinguished by looking at elements such as motif, color, shape deformation and brush texture. In our work, we focus on the effect of "color" element and devise a method for transforming the color of an input photograph according to a reference painting. To do this, we consider basic color category concepts in the color transformation process. We assume that color transformations from one basic color category to another may cause peculiar feelings. Therefore, we restrict each color transformation within the same basic color category. For this, our algorithm first categorizes each pixel color of a photograph into one of eleven basic color categories. Next, for every pixel color of the photograph, the algorithm finds its corresponding color in the same category of a reference painting. Finally, the algorithm substitutes the pixel color with its corresponding color. In this way, we achieve large but natural color transformations of an image.

We developed a method for extracting negative examples when only positive examples are given as supervised data. This method calculates the probability of occurrence of an input example, which should be judged to be positive or negative. It considers an input example that has a high probability of occurrence but does not appear in the set of positive examples as a negative example. We used this method for one of important tasks in natural language processing: automatic detection of misspelled Japanese expressions. The results showed that the method is effective. In this study, we also described two other methods we developed for the detection of misspelled expressions: a combined method and a "leaving-one-out" method. In our experiments, we found that these methods are also effective.

Verb phrases are sometimes omitted in natural language (ellipsis). It is necessary to resolve the verb phrase ellipses in language understanding, machine translation, and dialogue processing. This paper describes a practical way to resolve verb phrase ellipses by using surface expressions and examples. To make heuristic rules for ellipsis resolution we classified verb phrase ellipses by checking whether the referent of a verb phrase ellipsis appears in the surrounding sentences or not. We experimented with the resolution of verb phrase elipses on a novel and obtained a recall rate of 73% and a precision rate of 66% on test sentences. In the case when the referent of a verb phrase ellipsis appeared in the surrounding sentences, the accuracy rate was high. But, in the case when the referent of a verb phrase ellipsis did not appear in the surrounding sentences, the accuracy rate was not so high. Since the analysis of this phenomena is very difficult, it is valuable to propose a way of solving the problem to a certain extent. When the size of corpus becomes larger and the machine performance becomes greater, the method of using corpus will become effective.

We apply two acceleration techniques for the backpropagation algorithm to an iterative gradient descent algorithm called the network inversion algorithm. Experimental results show that these techniques are also quite effective to decrease the number of iterations required for the detection of input vectors on the classification boundary of a multilayer perceptron.

We present a method for learning classification functions from pre-classified training examples and hypotheses written roughly by experts. The goal is to produce a classification function that has higher accuracy than either the expert's hypotheses or the classification function inductively learned from the training examples alone. The key idea in our proposed approach is to let the expert's hypotheses influence the process of learning inductively from the training examples. Experimental results are presented demonstrating the power of our approach in a variety of domains.

Automatic alignment of bilingual texts is useful to example-based machine translation by facilitating the creation of example pairs of translation for the machine. Two main approaches to automatic alignment have been reported in the literature. They are lexical approach and statistical approach. The former looks for relationships between lexical contents of the bilingual texts in order to find alignment pairs, while the latter uses statistical correlation between sentence lengths of the bilingual texts as the basis of matching. This paper describes a combination of the two approaches in aligning Japanese-Cinese bilingual texts by allowing kanji contents and sentence lengths in the texts to work together in achieving an alignment process. Because of the sentential structure differences between Japanese and Chinese, matching at the sentence level may result in frequent matching between a number of sentences en masses. In view of this, the current work also attempts to create shorter alignment pairs by permitting sentences to be matched with clauses or phrases of the other text if possible. While such matching is more difficult and error-prone, the reliance on kanji contents has proven to be very useful in minimizing the errors. The current research has thus found solutions to problems that are unique to the present work.

Learning correctly from queries" is a formal learning model proposed by Angluin. In this model, for a class Γ of language representations, a learner asks queries to a teacher of an unknown language Lq which can be represented by some GqΓ, and eventually outputs a language representation GΓ which represents Lq and halts. An algorithm (leaner) A is said to learn a class of languages represented by Γ in the weak definition if the time complexity of A is some polynomial of n and m, where n is the minimum size of the lagunage representations in Γ which represent Lq, and m is the maximum length of the counterexamples returned in an execution. On the other band, A is said to learn represented by Γ in the strong definition if at any point τ of the execution, the time consumed up to τ is some polynomial of n and m, where n is the same as above, and m is the maximum length of the counterexamples returned up to τ. In this paper, adequacy of the model is examined, and it is shown that both in the weak and strong definitions, there exist learners which extract a long counterexample, and identify Lq by using equivalence queries exhaustively. For example, there exists a learner which learns the class CFL of context-free languages represented by the class CFG of context-free grammars in the weak definition using only equivalence queries. Next, two restrictions concerning with learnability criteria are introduced. Proper termination condition is that when a teacher replies with yes" to an equivalence query, then the learner must halt immediately. The other condition, called LBC-condition, is that in the weak/strong definition, the time complexity must be some polynomial of n and log m. In this paper, it is shown that under these conditions, there still exist learners which execute exhaustive search. For instance, there exists a learner which learns CFL represented by CFG in the weak definition using membership queries and equivalence queries under the proper termination condition, and there also exists a learner that learns CFL represented by CFG in the strong definition using subset queries and superset queries under LBC-condition. These results suggest that the weak definition is not an adequate learning model even if the proper termination condition is assumed. Also, the model becomes inadequate in the strong definition if some combination of queries, such as subset queries and superset queries, is used instead of equivalence queries. Many classes of languages become learnable by our extracting long counterexample" technique. However, it is still open whether or not CFL represented by CFG is learnable in the strong definition from membership queries and equivalence queries, although the answer is known to be negative if at least one of (1) quadratic residues modulo a composite, (2) inverting RSA encryption, or (3) factoring Blum integers, is intractable.

A case structure expression is one of the most important forms to represent the meaning of the sentence. Case structure analysis is usually performed by consulting case frame information in a verb dictionary. However, this analysis is very difficult because of several problems, such as word sense ambiguity and structural ambiguity. A conventional method for solving these problems is to use the method of selectional restriction, but this method has a drawback in the semantic marker (SM) method --the trade-off between descriptive power and construction cost. In this paper, we propose a method of case structure analysis based on examples in case frame dictionary This method uses the case frame dictionary which has some typical example sentences for each case frame, and it selects a proper case frame for an input sentence by matching the input sentence with the examples in the case frame dictionary. The best matching score, which is utilized for selecting a proper case frame for a predicate, can be considered as the score for the case structure of the predicate. Therefore, when there are two or more readings for a sentence because of structural ambiguity, the best reading of a sentence can be selected by evaluating the sum of the scores for the case structures of all predicates in a sentence. We report on experiments which shows that this method is superior to the conventional, coarse-grained SM method, and also describe the superiority of the example-based method over the SM method.

This paper presents a new method for resolving lexical (word sense) ambiguities inherent in natural language sentences. The Sentence Analyzer (SENA) was developed to resolve such ambiguities by using constraints and example-based preferences. The ambiguities are packed into a single dependency structure, and grammatical and lexical constraints are applied to it in order to reduce the degree of ambiguity. The application of constraints is realized by a very effective constraint-satisfaction technique. Remaining ambiguities are resolved by the use of preferences calculated from an example-base, which is a set of fully parsed word-to-word dependencies acquired semi-automatically from on-line dictionaries.

This paper proposes a new type of transfer system, called Similarity-driven Transfer System (or SimTran), which uses an example-based approach to the transfer phase of MT. In this paper, we describe a method for calculating similarity, a method for searching the most appropriate set of translation rules, and a method for constructing an output structure from those selected rules. Further, we show that SimTran can use not only translation examples but also syntax-based translation rules used in conventional transfer systems at the same time.

This paper presents a system that automatically refines the theory expressed in the function-free first-order logic. Our system can efficiently correct multiple faults in both the concept and subconcepts of the theory, given only the classified examples of the concept. It can refine larger classes of theory than existing systems can since it has overcome many of their limitations. Our system is based on a new combination of an inductive and an explanation-based learning algorithms, which we call the biggest-first multiple-example EBL (BM-EBL). From a learning perspective, our system is an improvement over the FOIL learning system in that our system can accept a theory as well as examples. An experiment shows that when our system is given a theory that has the classification error rate as high as 50%, it can still learn faster and with more accuracy than when it is not given any theory.

This paper deals with the problem of translating Japanese adnominal particles into English according to the idea of Example-Based Machine Translation (EBMT) proposed by Nagao. Japanese adnominal particles are important because: (1) they are frequent function words; (2) to translate them into English is difficult because their translations are diversified; (3) EBMT's effectiveness for adnominal particles suggests that EBMT is effective for other function words, e. g., prepositions of European languages. In EBMT, (1) a database which consists of examples (pairs of a source language expression and its target language translation) is prepared as knowledge for translation; (2) an example whose source expression is similar to the input phrase or sentence is retrieved from the example database; (3) by replacements of corresponding words in the target expression of the retrieved example, the translation is obtained. The similarity in EBMT is computed by the summation of the distance between words multiplied by the weight of each word. The authors' method differs from preceding research in two important points: (1) the authors utilize a general thesaurus to compute the distance between words; (2) the authors propose a weight which changes for every input. The feasibility of our approach has been proven through experiments concerning success rate.

This paper proposes a model for learning non-parametric densities using finite-dimensional parametric densities by applying Yamanishi's stochastic analogue of Valiant's probably approximately correct learning model to density estimation. The goal of our learning model is to find, with high probability, a good parametric approximation of the non-parametric target density with sample size and computation time polynomial in parameters of interest. We use a learning algorithm based on the minimum description length (MDL) principle and derive a new general upper bound on the rate of convergence of the MDL estimator to a true non-parametric density. On the basis of this result, we demonstrate polynomial-sample-size learnability of classes of non-parametric densities (defined under some smoothness conditions) in terms of exponential families with polynomial bases, and we prove that under some appropriate conditions, the sample complexity of learning them is bounded as O((1/ε)(2r1)/2r1n(2r1)/2r(1/ε)(1/ε)1n(1/δ) for a smoothness parameter r (a positive integer), where ε and δ are respectively accuracy and confidence parameters. Futher, we demonstrate polynomial-time learnability of classes of non-parametric densities (defined under some smoothness conditions) in terms of histogram densities with equal-length cells, and we prove that under some appropriate condition, the sample complexity of learning them is bounded as O((1/ε)3/21n3/2(1/ε)(1/ε)1n(1/δ)).