Search

In this paper we present our approach towards legal citation classification using incremental knowledge acquisition. This forms a part of our more ambitious goal of automatic legal text summarization. We created a large training and test corpus from court decision reports in Australia. We showed that, within less than a week, it is possible to develop a good quality knowledge base which considerably outperforms a baseline Machine Learning approach. We note that the problem of legal citation classification allows the use of Machine Learning as classified training data is available. For other subproblems of legal text summarization this is unlikely to be the case. [ABSTRACT FROM AUTHOR]

We present a new interactive workbench RDRCE (RDR Case Explorer) to facilitate the combination of Machine Learning and manual Knowledge Acquisition for Natural Language Processing problems. We show how to use Brill΄s well regarded transformational learning approach and convert its results into an RDR tree. RDRCE then strongly guides the systematic inspection of the generated RDR tree in order to further refine and improve it by manually adding more rules. Furthermore, RDRCE also helps in quickly recognising potential noise in the training data and allows to deal with noise effectively. Finally, we present a first study using RDRCE to build a high-quality Part-of-Speech tagger for English. After some 60 hours of manual knowledge acquisition, we already exceed slightly the state-of-the art performance on unseen benchmark test data and the fruits of some 15 years of further research in learning methods for Part-of-Speech taggers. [ABSTRACT FROM AUTHOR]

Word Sense Disambiguation(WSD) is one of the traditionally most difficult problems in natural language processing and has broad theoretical and practical implications. One of the main difficulties for WSD systems is the lack of relevant knowledge-commonly known as the knowledge acquisition bottleneck problem. We present in this paper a novel method that utilizes dynamic Web data obtained through Web search engines to effectively enrich the semantic knowledge for WSD systems. We demonstrated through a word sense disambiguation system the large quantity and good quality of the extracted knowledge. [ABSTRACT FROM AUTHOR]

We present a novel Web Pages Summarizer ContextSummarizer that groups the given Web pages into ‘sense-clusters' respecting a user's topical interests. ContextSummarizer constructs then an extractive summary for each sense-cluster. A user's topical interest is described by the user who selects and refines some of the word senses disambiguated within the content contexts of the given Web pages. The semantic similarity measures between the contents of Web pages/segments/sentences and the user-selected word senses were used to choose the most topically relevant sentences as the extractive summaries referring to a user's topical interest. ContextSummarizer addresses the semantic-alignment problem between the content of a Web page, the user's topical interest, and the extractive summary of the Web page. Our case studies and experimental results showed that our query-topic focused extractive summaries returns more topically relevant sentences for an extractive summary than those produced by existing summarization systems. [ABSTRACT FROM AUTHOR]

Most traditional relevance feedback systems simply choose the top ranked Web pages as the source of providing the weights of candidate query expansion terms. However, the contents of such top-ranked Web pages is often composed of heterogeneous sub-topics which can be and should be recognized and distinguished. However, current approaches treat retrieved Web pages as one unit and often fail to extract good quality candidate query expansion terms. In this paper, our basic idea is that the Web pages properly clustered into a sub-topic cluster can be used as a better source than whole given Web pages, to provide more topically coherent relevance feedback for that specific sub-topic. Thus, we propose Clustering-Based Relevance Feedback for Web Pages, which utilizes three methods to cluster retrieved Web pages into several subtopic-clusters. These three methods cooperate to construct good quality clusters by respectively supporting Web page Segmentation, Term Selection, k Seed Centroid Selection. Here, the automatically selected terms indicate the relevance feedback to construct all sub-topic clusters and assign the given Web pages to proper clusters. Each subset of the selected terms, which occurs in the Web pages assigned into a sub-topic cluster, indicates the relevance feedback to expand a query over that sub-topic cluster. Our experimental results showed that the clustering performances based on two traditional term-weighting methods (i.e., an unsupervised method and a supervised method) can be significantly improved with our methods. [ABSTRACT FROM AUTHOR]

Literature discovery can be characterized as a goal directed search for previously unknown implicit knowledge captured within a collection of scientific articles. Swanson's serendipitous discovery of a treatment for Raynaud's disease by dietary fish-oil while browsing Medline, an online collection of biomedical literature, exemplifies such a discovery. By means of a series of experiments, the impact of stop words, various weighting schemes, discovery mechanisms, and contextual reduction are studied in relation to replicating the Raynaud/fish-oil and migraine-magnesium discoveries by operational means. Two aspects of discovery were brought under focus: (i) the discovery of intermediate, or B -terms, and (ii) the discovery of indirect A - C connections via the B-terms. A semantic space representation of the underlying corpus is computed and discoveries automated by computing associations between words in both higher and contextually reduced spaces. It was found that the discovery of B-terms and A - C connections can be achieved to an encouraging degree with a standard stop word list. In addition, no single weighting scheme seems to suffice. Log-likelihood appears to be potentially effective for leading to the discovery of B-terms, whereas both odds ratio and simple co-occurrence frequencies both facilitate the discovery of A - C connections. With regard to discovery mechanism, both semantic similarity (via cosine) and information flow computation seem promising for computing A - C connections, but more research is needed to understand their relative strengths and weaknesses. Discovery in a contextually reduced semantic space revealed mixed results. [ABSTRACT FROM AUTHOR]

We are interested in the automation of science for both philosophical and technological reasons. To this end we have built the first automated system that is capable of automatically: originating hypotheses to explain data, devising experiments to test these hypotheses, physically running these experiments using a laboratory robot, interpreting the results, and then repeat the cycle. We call such automated systems "Robot Scientists". We applied our first Robot Scientist to predicting the function of genes in a well-understood part of the metabolism of the yeast S. cerevisiae. For background knowledge, we built a logical model of metabolism in Prolog. The experiments consisted of growing mutant yeast strains with known genes knocked out on specified growth media. The results of these experiments allowed the Robot Scientist to test hypotheses it had abductively inferred from the logical model. In empirical tests, the Robot Scientist experiment selection methodology outperformed both randomly selecting experiments, and a greedy strategy of always choosing the experiment of lowest cost; it was also as good as the best humans tested at the task. To extend this proof of principle result to the discovery of novel knowledge we require new hardware that is fully automated, a model of all of the known metabolism of yeast, and an efficient way of inferring probable hypotheses. We have made progress in all of these areas, and we are currently 6building a new Robot Scientist that we hope will be able to automatically discover new biological knowledge. [ABSTRACT FROM AUTHOR]

This paper presents an algorithm to prune a tree classifier with a set of rules which are converted from a C4.5 classifier, where rule information is used as a pruning criterion. Rule information measures the goodness of a rule when discriminating labeled instances. Empirical results demonstrate that the proposed pruning algorithm has high predictive accuracyThe work is supported by the Scientific Research Foundation for the Returned Overseas Chinese Scholars, State Education Ministry, and the Project (No.2004D006) from Hubei Provincial Department of Education, P. R. China.. [ABSTRACT FROM AUTHOR]

This paper proposes a method for detecting misclassifications of a classification rule and then revising them. Given a rule and a set of examples, the method divides misclassifications by the rule into miscovered examples and uncovered examples, and then, separately, learns to detect them using Inductive Logic Programming (ILP). The method then combines the acquired rules with the initial rule and revises the labels of misclassified examples. The paper shows the effectiveness of the proposed method by theoretical analysis. In addition, it presents experimental results, using the Brill tagger for Part-Of-Speech (POS) tagging. [ABSTRACT FROM AUTHOR]

A novel inductive logic programming system, called Classic'cl is presented. Classic'cl integrates several settings for learning, in particular learning from interpretations and learning from satisfiability. Within these settings, it addresses descriptive and probabilistic modeling tasks. As such, Classic'cl (C-armr, cLAudien, icl-S(S)at, ICl, and CLlpad) integrates several well-known inductive logic programming systems such as Claudien, Warmr (and its extension C-armr), ICL, ICL-SAT, and LLPAD. We report on the implementation, the integration issues as well as on some experiments that compare Classic'cl with some of its predecessors. [ABSTRACT FROM AUTHOR]

In this paper, we discuss a method of finding useful clusters of web pages which are significant in the sense that their contents are similar or closely related to ones of higher-ranked pages. Since we are usually careless of pages with lower ranks, they are unconditionally discarded even if their contents are similar to some pages with high ranks. We try to extract such hidden pages together with significant higher-ranked pages as a cluster. In order to obtain such clusters, we first extract semantic correlations among terms by applying Singular Value Decomposition(SVD) to the term-document matrix generated from a corpus w.r.t. a specific topic. Based on the correlations, we can evaluate potential similarities among web pages from which we try to obtain clusters. The set of web pages is represented as a weighted graph G based on the similarities and their ranks. Our clusters can be found as pseudo-cliques in G. We present an algorithm for finding Top-N weighted pseudo-cliques. Our experimental result shows that quite valuable clusters can be actually extracted according to our method. [ABSTRACT FROM AUTHOR]

Extracting useful knowledge from numerous distributed data repositories can be a very hard task when such data cannot be directly centralized or unified as a single file or database. This paper suggests practical distributed clustering algorithms without accessing the raw data to overcome the inefficiency of centralized data clustering methods. The aim of this research is to generate unit volume based probabilistic mixture model from local clustering results without moving original data. It has been shown that our method is appropriate for distributed clustering when real data cannot be accessed or centralized. [ABSTRACT FROM AUTHOR]

In recent years, the problem of finding the different aspects existing in a dataset has attracted many authors in the domain of knowledge quality in KDD. The discovery of knowledge in the form of association rules has become an important research. One of the most difficult issues is that an enormous number of association rules are discovered, so it is not easy to choose the best association rules or knowledge for a given dataset. Some methods are proposed for choosing the best rules with an interestingness measure or matching properties of interestingness measure for a given set of interestingness measures. In this paper, we propose a new approach to discover the clusters of interestingness measures existing in a dataset. Our approach is based on the evaluation of the distance computed between interestingness measures. We use two techniques: agglomerative hierarchical clustering (AHC) and partitioning around medoids (PAM) to help the user graphically evaluates the behavior of interestingness measures. [ABSTRACT FROM AUTHOR]

Text mining techniques have been proposed for extracting protein names and their interactions from biological text. First, we have made improvements on existing methods for handling single word protein names consisting of characters, special symbols, and numbers. Second, compound word protein names are also extracted using conditional probabilities of the occurrences of neighboring words. Third, interactions are extracted based on Bayes theorem over discriminating verbs that represent the interactions of proteins. Experimental results demonstrate the feasibility of our approach with improved performance in terms of accuracy and F-measure, requiring significantly less amount of computational time. [ABSTRACT FROM AUTHOR]

Many practical applications of machine learning in data-driven scientific discovery commonly call for the exploration of data from multiple points of view that correspond to explicitly specified ontologies. This paper formalizes a class of problems of learning from ontology and data, and explores the design space of learning classifiers from attribute value taxonomies (AVTs) and data. We introduce the notion of AVT-extended data sources and partially specified data. We propose a general framework for learning classifiers from such data sources. Two instantiations of this framework, AVT-based Decision Tree classifier and AVT-based Naïve Bayes classifier are presented. Experimental results show that the resulting algorithms are able to learn robust high accuracy classifiers with substantially more compact representations than those obtained by standard learners. [ABSTRACT FROM AUTHOR]

This work focuses on the active selection of pairwise constraints for spectral clustering. We develop and analyze a technique for Active Constrained Clustering by Examining Spectral eigenvectorS (ACCESS) derived from a similarity matrix. The ACCESS method uses an analysis based on the theoretical properties of spectral decomposition to identify data items that are likely to be located on the boundaries of clusters, and for which providing constraints can resolve ambiguity in the cluster descriptions. Empirical results on three synthetic and five real data sets show that ACCESS significantly outperforms constrained spectral clustering using randomly selected constraints. [ABSTRACT FROM AUTHOR]

Most technical and scientific terms are comprised of complex, multi-word noun phrases but certainly not all noun phrases are technical or scientific terms. The distinction of specific terminology from common non-specific noun phrases can be based on the observation that terms reveal a much lesser degree of distributional variation than non-specific noun phrases. We formalize the limited paradigmatic modifiability of terms and, subsequently, test the corresponding algorithm on bigram, trigram and quadgram noun phrases extracted from a 104-million-word biomedical text corpus. Using an already existing and community-wide curated biomedical terminology as an evaluation gold standard, we show that our algorithm significantly outperforms standard term identification measures and, therefore, qualifies as a high-performant building block for any terminology identification system. [ABSTRACT FROM AUTHOR]

In this paper, the semantic relationships between a predicate and its arguments in terms of semantic roles are employed to improve lexical-based named entity recognition (NER) in the molecular biology domain. The semantic roles were realized in various sets of syntactic features used by a machine learning model to explore what should be the efficient way in allowing this knowledge to provide the highest positive effect on the NER. The empirical results show that the best feature set consists of predicate's surface form, predicate's lemma, voice, and the united feature of subject-object head's lemma and transitive-intransitive sense. The performance improvement from using these features indicates the advantage of the predicate-argument semantic knowledge on NER. There are still rooms to enhance NER by using this semantic knowledge (e.g. to employ other semantic roles besides agent and theme and to extend the rules for efficient identification of an argument's boundary). [ABSTRACT FROM AUTHOR]

This paper proposes a novel system to discover simultaneous time differential law equations reflecting first principles underlying objective processes. The system has the power to discover equations containing hidden state variables and/or representing chaotic dynamics without using any detailed domain knowledge. These tasks have not been addressed in any mathematical and engineering domains in spite of their essential importance. Its promising performance is demonstrated through applications to both mathematical and engineering examples. [ABSTRACT FROM AUTHOR]

Previous sphere-based classification algorithms usually need a number of spheres in order to achieve good classification performance. In this paper, inspired by the support vector machines for classification and the support vector data description method, we present a new method for constructing single spheres that separate data with the maximum separation ratio. In contrast to previous methods that construct spheres in the input space, the new method constructs separating spheres in the feature space induced by the kernel. As a consequence, the new method is able to construct a single sphere in the feature space to separate patterns that would otherwise be inseparable when using a sphere in the input space. In addition, by adjusting the ratio of the radius of the sphere to the separation margin, it can provide a series of solutions ranging from spherical to linear decision boundaries, effectively encompassing both the support vector machines for classification and the support vector data description method. Experimental results show that the new method performs well on both artificial and real-world datasets. [ABSTRACT FROM AUTHOR]

Given a transaction database as a global set of transactions and its local database obtained by some conditioning to the global one, we consider a pair of itemsets whose degrees of correlations are higher in the local database than in the global one. A problem of finding paired itemsets with high correlation in one database is known as Discovery of Correlation, and some algorithms to search for such characteristic paired itemsets are already proposed. However, even non-characteristic paired itemsets in the local database are also meaningful, provided the degree of correlation increases much higher in the local database than in the global one. They can be an implicit and hidden evidence showing that something particular to the local database occurs even though they are not yet realized as characteristic ones in the local. From this viewpoint, we have already proposed to measure the significance of paired itemsets by the difference of two correlations before and after the conditioning to the local database, and define a notion of DC pairs whose degrees of differences of correlations are high. As DC pairs are regarded as compound itemsets consisting of two component itemsets, we can have two basic strategies for finding them. One strategy firstly examines the compound itemsets and then the components, while another one does the component itemsets and then the compound ones. According to the former strategy, which we have already proposed and tested for its effectiveness, we have to enumerate many number of candidate compound itemsets that cannot be decomposable to components. For this reason, this paper presents a new algorithm according to the second strategy. It firstly enumerate possible component itemsets based on a new pruning rule for cutting off useless components. Secondly it forms the compound itemsets by combining the components thus detected, while we also make use of a constraint for preventing our algorithm from checking meaningless combinations. [ABSTRACT FROM AUTHOR]

The classification of time-varying neurophysiological signals, e.g., electroencephalogram (EEG) signals, advances the requirement of adaptability for classifiers. In this paper we address the challenge of neurophysiological signal classification arising from brain-computer interface (BCI) applications and propose an on-line classifier designed via the decorrelated least mean square (LMS) algorithm. Based on a Bayesian classifier with Gaussian mixture models, we derive the general formulation of gradient descent algorithms under the criterion of LMS. Further, to accelerate convergence, the decorrelated gradient instead of the instantaneous gradient is adopted for updating the parameters of the classifier adaptively. Utilizing the presented classifier for the off-line analysis of practical classification tasks in brain-computer interface applications shows its effectiveness and robustness compared to the stochastic gradient descent classifier which uses the instantaneous gradient directly. [ABSTRACT FROM AUTHOR]

The paper focuses on the problem of classification by function decomposition within the frame of monotone classification. We propose a decomposition method for discrete functions which can be applied to monotone problems in order to generate a monotone classifier based on the extracted concept hierarchy. We formulate and prove a criterion for the existence of a positive extension of the scheme f=g(S0,h(S1)) in the context of discrete functions. We also propose a method for finding an assignment for the intermediate concept with a minimal number of values. [ABSTRACT FROM AUTHOR]

In this paper, we investigate the q-gram distance for ordered unlabeled trees (trees, for short). First, we formulate a q-gram as simply a tree with q nodes isomorphic to a line graph, and the q-gram distance between two trees as similar as one between two strings. Then, by using the depth sequence based on postorder, we design the algorithm EnumGram to enumerate all q-grams in a tree T with n nodes which runs in O(n2) time and in O(q) space. Furthermore, we improve it to the algorithm LinearEnumGram which runs in O(qn) time and in O(qd) space, where d is the depth of T. Hence, we can evaluate the q-gram distance Dq(T1,T2) between T1 and T2 in time and in space, where ni and di are the number of nodes in Ti and the depth of Ti, respectively. Finally, we show the relationship between the q-gram distance Dq(T1,T2) and the edit distanceE(T1,T2) that Dq(T1,T2)≤ (gl+1)E(T1,T2), where , , gi is the degree of Ti and li is the number of leaves in Ti. In particular, for the top-down tree edit distanceF(T1,T2), this result implies that . [ABSTRACT FROM AUTHOR]

We consider the problem of finding a set of patterns that best characterizes a set of strings. To this end, Arimura et. al. [3] considered the use of minimal multiple generalizations (mmg) for such characterizations. Given any sample set, the mmgs are, roughly speaking, the most (syntactically) specific set of languages containing the sample within a given class of languages. Takae et. al. [17] found the mmgs of the class of pattern languages [1] which includes so-called sort symbols to be fairly accurate as predictors for signal peptides. We first reproduce their results using updated data. Then, by using a measure for estimating the level of over-generalizations made by the mmgs, we show results that explain the high level of accuracies resulting from the use of sort symbols, and discuss how better results can be obtained. The measure that we suggests here can also be applied to other types of patterns, e.g. the PROSITE patterns [4]. [ABSTRACT FROM AUTHOR]