We introduce a method for learning to find documents on the Web that contain answers to a given natural language question. In our approach, questions are transformed into new queries aimed at maximizing the probability of retrieving answers from existing information retrieval systems. The method involves automatically learning phrase features for classifying questions into different types, automatically generating candidate query transformations from a training set of question/answer pairs, and automatically evaluating the candidate transformations on target information retrieval systems such as real-world general purpose search engines. At run-time, questions are transformed into a set of queries, and reranking is performed on the documents retrieved. We present a prototype search engine,Tritus, that applies the method to Web search engines. Blind evaluation on a set of real queries from a Web search engine log shows that the method significantly outperforms the underlying search engines, and outperforms a commercial search engine specializing in question answering. Our methodology cleanly supports combining documents retrieved from different search engines, resulting in additional improvement with a system that combines search results from multiple Web search engines.

A major obstacle to fully integrated deployment of many data mining algorithms is the assumption that data sits in a single table, even though most real-world databases have complex relational structures. We propose an integrated approach to statistical modelling from relational databases. We structure the search space based on "refinement graphs", which are widely used in inductive logic programming for learning logic descriptions. The use of statistics allows us to extend the search space to include richer set of features, including many which are not Boolean. Search and model selection are integrated into a single process, allowing information criteria native to the statistical model, for example logistic regression, to make feature selection decisions in a step-wise manner. We present experimental results for the task of predicting where scientific papers will be published based on relational data taken from CiteSeer. Our approach results in classification accuracies superior to those achieved when using classical "flat" features. The resulting classifier can be used to recommend where to publish articles.

This paper describes a method of mapping music into a semantic space that can be used for similarity measurement, classification, and music information retrieval. The value along each dimension of this anchor space is computed as the output from a pattern classifier which is trained to measure a particular semantic feature. In anchor space, distributions that represent objects such as artists or songs are modeled with Gaussian mixture models, and several similarity measures are defined by computing approximations to the Kullback-Leibler divergence between distributions. Similarity measures are evaluated against human similarity judgements. The models are also used for artist classification to achieve 62% accuracy on a 25-artist set, and 38% on a 404-artist set (random guessing achieves 0.25%). Finally, we describe a music similarity browsing application that makes use of the fact that anchor space dimensions are meaningful to users.

The susceptibility of the Internet to random faults, malicious attacks, and mixtures of faults and attacks are analyzed. We analyze actual Internet data, as well as simulated data created with network models. The network models generalize previous research, and allow generation of graphs ranging from uniform to preferential, and from static to dynamic. We introduce new metrics for analyzing the connectivity and performance of networks which improve upon metrics used in earlier research. Previous research has shown that preferential networks like the Internet are more robust to random failures compared to uniform networks. We find that preferential networks, including the Internet, are more robust only when more than 95% of failures are random faults, and robustness is measured with average diameter. The advantage of preferential networks disappears with alternative metrics, and when a small fraction of faults are attacks. We also identify dynamic characteristics of the Internet which can be used to create improved network models. This model should allow more accurate analysis for the future Internet, for example facilitating the design of network protocols with optimal performance in the future, or predicting future attack and fault tolerance. We find that the Internet is becoming more preferential as it evolves. The average diameter has been stable or even decreasing as the number of nodes has been increasing. The Internet is becoming more robust to random failures over time, but has also become more vulnerable to attacks.

We analyze data from $52$ online in-game sports betting markets (where betting is allowed continuously throughout a game), including 34 markets based on soccer (European football) games from the 2002 World Cup, and 18 basketball games from the 2002 USA National Basketball Association (NBA) championship. We show that prices on average approach the correct outcome over time, and the price dynamics in the markets are closely coupled with game events, agreeing with efficient market assumptions. We also examine qualitative distinctions between the two types of games.

Niche Search Engines offer an efficient alternative to traditional search engines when the results returned by general-purpose search engines do not provide a sufficient degree of relevance. By taking advantage of their domain of concentration they achieve higher relevance and offer enhanced features. We discuss a new niche search engine, eBizSearch, based on the technology of CiteSeer and dedicated to e-business and e-business documents. We present the integration of CiteSeer in the framework of eBizSearch and the process necessary to tune the whole system towards the specific area of e-business. We also discuss how using machine learning algorithms we generate metadata to make eBizSearch Open Archives compliant. eBizSearch is a publicly available service and can be reached at [3].

The web contains a wealth of product reviews, but sifting through them is a daunting task. Ideally, an opinion mining tool would process a set of search results for a given item, generating a list of product attributes (quality, features, etc.) and aggregating opinions about each of them (poor, mixed, good). We begin by identifying the unique properties of this problem and develop a method for automatically distinguishing between positive and negative reviews. Our classifier draws on information retrieval techniques for feature extraction and scoring, and the results for various metrics and heuristics vary depending on the testing situation. The best methods work as well as or better than traditional machine learning. When operating on individual sentences collected from web searches, performance is limited due to noise and ambiguity. But in the context of a complete web-based tool and aided by a simple method for grouping sentences into attributes, the results are qualitatively quite useful.

We create a statistical model for inferring hierarchical term relationships about a topic, given only a small set of example web pages on the topic, without prior knowledge of any hierarchical information. The model can utilize either the full text of the pages in the cluster or the context of links to the pages. To support the model, we use "ground truth" data taken from the category labels in the Open Directory. We show that the model accurately separates terms in the following classes: self terms describing the cluster, parent terms describing more general concepts, and child terms describing specializations of the cluster. For example, for a set of biology pages, sample parent, self, and child terms are science, biology, and genetics respectively. We create an algorithm to predict parent, self, and child terms using the new model, and compare the predictions to the ground truth data. The algorithm accurately ranks a majority of the ground truth terms highly, and identifies additional complementary terms missing in the Open Directory.

When searching the WWW, users often desire results restricted to a particular document category. Ideally, a user would be able to filter results with a text classifier to minimize false positive results; however, current search engines allow only simple query modifications. To automate the process of generating effective query modifications, we introduce a sensitivity analysis-based method for extracting rules from nonlinear support vector machines. The proposed method allows the user to specify a desired precision while attempting to maximize the recall. Our method performs several levels of dimensionality reduction and is vastly faster than searching the combination feature space; moreover, it is very effective on real-world data.

The structure of the web is increasingly being used to improve organization, search, and analysis of information on the web. For example, Google uses the text in citing documents (documents that link to the target document) for search. We analyze the relative utility of document text, and the text in citing documents near the citation, for classification and description. Results show that the text in citing documents, when available, often has greater discriminative and descriptive power than the text in the target document itself. The combination of evidence from a document and citing documents can improve on either information source alone. Moreover, by ranking words and phrases in the citing documents according to expected entropy loss, we are able to accurately name clusters of web pages, even with very few positive examples. Our results confirm, quantify, and extend previous research using web structure in these areas, introducing new methods for classification and description of pages.

ResearchIndex (also known as CiteSeer) is a digital library of scientific literature that aims to improve communication and progress in science. This talk covers the design, implementation, and operation of ResearchIndex.

Game sites on the World Wide Web draw people from around the world with specialized interests, skills, and knowledge. Data from the games often reflects the players' expertise and will to win. We extract probabilistic forecasts from data obtained from three online games: the Hollywood Stock Exchange (HSX), the Foresight Exchange (FX), and the Formula One Pick Six (F1P6) competition. We find that all three yield accurate forecasts of uncertain future events. In particular, prices of so-called "movie stocks" on HSX are good indicators of actual box office returns. Prices of HSX securities in Oscar, Emmy, and Grammy awards correlate well with observed frequencies of winning. FX prices are reliable indicators of future developments in science and technology. Collective predictions from players in the F1 competition serve as good forecasts of true race outcomes. In some cases, forecasts induced from game data are more reliable than expert opinions. We argue that web games naturally attract well-informed and well-motivated players, and thus offer a valuable and oft-overlooked source of high-quality data with significant predictive value.

We introduce a method for learning query transformations that improves the ability to retrieve answers to questions from an information retrieval system. During the training stage the method involves automatically learning phrase features for classifying questions into different types, automatically generating candidate query transformations from a training set of question/answer pairs, and automatically evaluating the candidate transforms on target information retrieval systems such as real-world general purpose search engines. At run time, questions are transformed into a set of queries, and re-ranking is performed on the documents retrieved. We present a prototype search engine, Tritus, that applies the method to web search engines. Blind evaluation on a set of real queries from a web search engine log shows that the method significantly outperforms the underlying web search engines as well as a commercial search engine specializing in question answering. Keywords Web search, quer...

We p r e s e n t DEADLINER, a search engine that catalogs conference and workshop announcements, and ultimately will monitor and extract a wide range of academic convocation material from the web.The system currently extracts speakers, locations, dates, paper submission (and other) deadlines, topics, program committees, abstracts, and aÆliations.A user or user agent can perform detailed searches on these elds.DEADLINER was constructed using a methodology for rapid implementation of specialized search engines.This methodology avoids complex hand-tuned text extraction solutions, or natural language processing, by B a yesian integration of simple extractors that exploit loose formatting and keyw ord con ventions.The Bayesian framework further produces a search engine where each user can control the false alarm rate on a eld in an intuitive y et rigorous fashion.

We analyze the persistence of information on the web, looking at the percentage of invalid URLs contained in academic articles within the CiteSeer (ResearchIndex) database.The number of URLs contained in the papers has increased from an average of 0.06 in 1993 to 1.6 in 1999.We found that a significant percentage of URLs are now invalid, ranging from 23% for 1999 articles, to 53% for 1994.We also found that for almost all of the invalid URLs, it was possible to locate the information (or highly related information) in an alternate location, primarily with the use of search engines.However, the ability to relocate missing information varied according to search experience and effort expended.Citation practices suggest that more information may be lost in the future unless these practices are improved.We discuss persistent URL standards and their usage, and give recommendations for citing URLs in research articles as well as for finding the new location of invalid URLs.

Methods for controlling the bias/variance tradeoff typically assume that overfitting or overtraining is a global phenomenon. For multi-layer perceptron (MLP) neural networks, global parameters such as the training time, network size, or the amount of weight decay are commonly used to control the bias/variance tradeoff. However, the degree of overfitting can vary significantly throughout the input space of the model. We show that overselection of the degrees of freedom for an MLP trained with backpropagation can improve the approximation in regions of underfitting, while not significantly overfitting in other regions. This can be a significant advantage over other models. Furthermore, we show that "better" learning algorithms such as conjugate gradient can in fact lead to worse generalization, because they can be more prone to creating varying degrees of overfitting in different regions of the input space. While experimental results can not cover all practical situations, our results do help to explain common behavior that does not agree with theoretical expectations. Our results suggest one important reason for the relative success of MLPs, bring into question common beliefs about neural network training regarding training algorithms, overfitting, and optimal network size, suggest alternate guidelines for practical use, and help to direct future work.

We de ne a communit y on the web as a set of sites that have more links (in either direction) to members of the community than to non-members.Members of such a c o m m unity can be eÆciently iden ti ed in a maxim um ow / m i n im um cut framew ork, where the source is composed of known members, and the sink consists of well-kno wn non-members.A focused crawler that crawls to a xed depth can approximate community m e m bership by augmenting the graph induced by the cra wl with links to a virtual sink node.The effectiveness of the approximation algorithm is demonstrated with several crawl results that iden tify hubs, authorities, w eb rings, and other link topologies that are useful but not easily categorized.Applications of our approach include focused cra wlers and searc h engines, automatic population of portal categories, and improved ltering.

The conventional wisdom is that backprop nets with excess hidden units generalize poorly. We show that nets with excess capacity generalize well when trained with backprop and early stopping. Experiments sug(cid:173) gest two reasons for this: 1) Overfitting can vary significantly in different regions of the model. Excess capacity allows better fit to regions of high non-linearity, and backprop often avoids overfitting the regions of low non-linearity. 2) Regardless of size, nets learn task subcomponents in similar sequence. Big nets pass through stages similar to those learned by smaller nets. Early stopping can stop training the large net when it generalizes comparably to a smaller net. We also show that conjugate gradient can yield worse generalization because it overfits regions of low non-linearity when learning to fit regions of high non-linearity.

This paper examines the inductive inference of a complex grammar with neural networks and specifically, the task considered is that of training a network to classify natural language sentences as grammatical or ungrammatical, thereby exhibiting the same kind of discriminatory power provided by the Principles and Parameters linguistic framework, or Government-and-Binding theory. Neural networks are trained, without the division into learned vs. innate components assumed by Chomsky (1956), in an attempt to produce the same judgments as native speakers on sharply grammatical/ungrammatical data. How a recurrent neural network could possess linguistic capability and the properties of various common recurrent neural network architectures are discussed. The problem exhibits training behavior which is often not present with smaller grammars and training was initially difficult. However, after implementing several techniques aimed at improving the convergence of the gradient descent backpropagation-through-time training algorithm, significant learning was possible. It was found that certain architectures are better able to learn an appropriate grammar. The operation of the networks and their training is analyzed. Finally, the extraction of rules in the form of deterministic finite state automata is investigated.

When a query is submitted to a metasearch engine, decisions are made with respect to the underlying search engines to be used, what modifications will be made to the query, and how to score the results. These decisions are typically made by considering only the user's keyword query, neglecting the larger information need. Users with specific needs, such as “research papers” or “homepages,” are not able to express these needs in a way that affects the decisions made by the metasearch engine. In this paper, we describe a metasearch engine architecture that considers the user's information need for each decision. Users with different needs, but the same keyword query, may search different sub-search engines, have different modifications made to their query, and have results ordered differently. Our architecture combines several powerful approaches together in a single general purpose metasearch engine.

The web has greatly improved access to scientific literature. However, scientific articles on the web are largely disorganized, with research articles being spread across archive sites, institution sites, journal sites, and researcher homepages. No index covers all of the available literature, and the major web search engines typically do not index the content of Postscript/PDF documents at all. This paper discusses the creation of digital libraries of scientific literature on the web, including the efficient location of articles, full-text indexing of the articles, autonomous citation indexing, information extraction, display of query-sensitive summaries and citation context, hubs and authorities computation, similar document detection, user profiling, distributed error correction, graph analysis, and detection of overlapping documents. The software for the system is available at no cost for non-commercial use.

The paper considers the prediction of noisy time series data, specifically, the prediction of foreign exchange rate data. A novel hybrid neural network algorithm for noisy time series prediction is presented which exhibits excellent performance on the problem. The method is motivated by consideration of how neural networks work, and by fundamental difficulties with random correlations when dealing with small sample sizes and high noise data. The method permits the inference and extraction of rules. One of the greatest complaints against neural networks is that it is hard to figure out exactly what they are doing-this work provides one answer for the internal workings of the network. Furthermore, these rules can be used to gain insight into both the real world system and the predictor. The paper focuses on noisy time series prediction and rule inference-use of the system in trading would typically involve the utilization of other financial indicators and domain knowledge.

The performance of neural network simulations is often reported in terms of the mean and standard deviation of a number of simulations performed with different starting conditions. However, in many cases, the distribution of the individual results does not approximate a Gaussian distribution, may not be symmetric, and may be multimodal. We present the distribution of results for practical problems and show that assuming Gaussian distributions can significantly affect the interpretation of results, especially those of comparison studies. For a controlled task which we consider, we find that the distribution of performance is skewed toward better performance for smoother target functions and skewed toward worse performance for more complex target functions. We propose new guidelines for reporting performance which provide more information about the actual distribution.

We present a hybrid neural-network for human face recognition which compares favourably with other methods. The system combines local image sampling, a self-organizing map (SOM) neural network, and a convolutional neural network. The SOM provides a quantization of the image samples into a topological space where inputs that are nearby in the original space are also nearby in the output space, thereby providing dimensionality reduction and invariance to minor changes in the image sample, and the convolutional neural network provides partial invariance to translation, rotation, scale, and deformation. The convolutional network extracts successively larger features in a hierarchical set of layers. We present results using the Karhunen-Loeve transform in place of the SOM, and a multilayer perceptron (MLP) in place of the convolutional network for comparison. We use a database of 400 images of 40 individuals which contains quite a high degree of variability in expression, pose, and facial details. We analyze the computational complexity and discuss how new classes could be added to the trained recognizer.

Faces represent complex, multidimensional, meaningful visual stimuli and developing a computational model for face recognition is difficult. We present a hybrid neural network solution which compares favorably with other methods. The system combines local image sampling, a self-organizing map neural network, and a convolutional neural network. The self-organizing map provides a quantization of the image samples into a topological space where inputs that are nearby in the original space are also nearby in the output space, thereby providing dimensionality reduction and invariance to minor changes in the image sample, and the convolutional neural network provides for partial invariance to translation, rotation, scale, and deformation. The method is capable of rapid classification, requires only fast, approximate normalization and preprocessing, and consistently exhibits better classification performance than the eigenfaces approach on the database considered as the number of images per person in the training database is varied from 1 to 5. With 5 images per person the proposed method and eigenfaces result in 3.8% and 10.5% error respectively. The recognizer provides a measure of confidence in its output and classification error approaches zero when rejecting as few as 10% of the examples. We use a database of 400 images of 40 individuals which contains quite a high degree of variability in expression, pose, and facial details.