Reuters Corpus Volume I (RCV1) is an archive of over 800,000 manually categorized newswire stories recently made available by Reuters, Ltd. for research purposes. Use of this data for research on text categorization requires a detailed understanding of the real world constraints under which the data was produced. Drawing on interviews with Reuters personnel and access to Reuters documentation, we describe the coding policy and quality control procedures used in producing the RCV1 data, the intended semantics of the hierarchical category taxonomies, and the corrections necessary to remove errorful data. We refer to the original data as RCV1-v1, and the corrected data as RCV1-v2. We benchmark several widely used supervised learning methods on RCV1-v2, illustrating the collection’s properties, suggesting new directions for research, and providing baseline results for future studies. We make available detailed, per-category experimental results, as well as

The naive Bayes classifier, currently experiencing a renaissance in machine learning, has long been a core technique in information retrieval. We review some of the variations of naive Bayes models used for text retrieval and classification, focusing on the distributional assump- tions made about word occurrences in documents.

Systems for text retrieval, routing, categorization and other IR tasks rely heavily on linear classifiers. We propose that two machine learning algorithms, the Widrow-Hoff and EG algorithms, be used in training linear text classifiers. In contrast to most IR methods, theoretical analysis provides performance guarantees and guidance on parameter settings for these algorithms. Experimental data is presented showing Widrow-Hoff and EG to be more effective than the widely used Rocchio algorithm on several categorization and routing tasks. 1 Introduction Document retrieval, categorization, routing, and filtering systems often are based on classification. That is, the IR system decides for each document which of two or more classes it belongs to, or how strongly it belongs to a class, in order to accomplish the IR task of interest. For instance, the two classes may be the documents relevant to and not relevant to a particular user, and the system may rank documents based on how likely it i...

This paper presents a Support Vector Method for optimizing multivariate nonlinear performance measures like the F1score. Taking a multivariate prediction approach, we give an algorithm with which such multivariate SVMs can be trained in polynomial time for large classes of potentially non-linear performance measures, in particular ROCArea and all measures that can be computed from the contingency table. The conventional classification SVM arises as a special case of our method. 1.

This paper presents a novel way of examining the accuracy of the evaluation measures commonly used in information retrieval experiments. It validates several of the rules-of-thumb experimenters use, such as the number of queries needed for a good experiment is at least 25 and 50 is better, while challenging other beliefs, such as the common evaluation measures are equally reliable. As an example, we show that Precision at 30 documents has about twice the average error rate as Average Precision has. These results can help information retrieval researchers design experiments that provide a desired level of confidence in their results. In particular, we suggest researchers using Web measures such as Precision at 10 documents will need to use many more than 50 queries or will have to require two methods to have a very large difference in evaluation scores before concluding that the two methods are actually different.

Three different types of classifiers were investigated in the context of a text categorization problem in the medical domain: the automatic assignment of ICD9 codes to dictated inpatient discharge summaries. K-nearest-neighbor, relevance feedback, and Bayesian independence classifers were applied individually and in combination. A combination of different classifiers produced better results than any single type of classifier. For this specific medical categorization problem, new query formulation and weighting methods used in the k-nearest-neighbor classifier improved performance. 1 Introduction Past research in information retrieval has shown that one can improve retrieval effectiveness by using multiple representations in indexing and query formulation [27] [19] [3] [11] and by using multiple search strategies [5] [24] [7]. In this work, we investigate whether we can attain similar improvements in the domain of text categorization by combining different representations and classif...

In many practical learning scenarios, there is a small amount of labeled data along with a large pool of unlabeled data. Many supervised learning algorithms have been developed and extensively studied. We present a new "co-training" strategy for using unlabeled data to improve the performance of standard supervised learning algorithms. Unlike much of the prior work, such as the co-training procedure of Blum and Mitchell (1998), we do not assume there are two redundant views both of which are sufficient for perfect classification. The only requirement our co-training strategy places on each supervised learning algorithm is that its hypothesis partitions the example space into a set of equivalence classes (e.g. for a decision tree each leaf defines an equivalence class). We evaluate our co-training strategy via experiments using data from the UCI repository. 1. Introduction In many practical learning scenarios, there is a small amount of labeled data along with a lar...

We discuss two learning algorithms for text filtering: modified Rocchio and a boosting algorithm called AdaBoost. We show how both algorithms can be adapted to maximize any general utility matrix that associates cost (or gain) for each pair of machine prediction and correct label. We first show that AdaBoost significantly outperforms another highly effective text filtering algorithm. We then compare AdaBoost and Rocchio over three large text filtering tasks. Overall both algorithms are comparable and are quite effective. AdaBoost produces better classifiers than Rocchio when the training collection contains a very large number of relevant documents. However, on these tasks, Rocchio runs much faster than AdaBoost. 1

Most research in text classification has used the “bag of words ” representation of text. This paper examines some alternative ways to represent text based on syntactic and semantic relationships between words (phrases, synonyms and hypernyms). We describe the new representations and try to justify our suspicions that they could have improved the performance of a rule-based learner. The representations are evaluated using the RIPPER rule-based learner on the Reuters-21578 and DigiTrad test corpora, but on their own the new representations are not found to produce a significant performance improvement. Finally, we try combining classifiers based on different representations using a majority voting technique. This step does produce some performance improvement on both test collections. In general, our work supports the emerging consensus in the information retrieval community that more sophisticated Natural Language Processing techniques need to be developed before better text representations can be produced. We conclude that for now, research into new learning algorithms and methods for combining existing learners holds the most promise.

This paper describes a set of comparative experiments for the problem of automatically ltering unwanted electronic mail messages. Several variants of the AdaBoost algorithm with con dence{ rated predictions (Schapire & Singer 99) have been applied, which di er in the complexity of the base learners considered. Two main conclusions can be drawn from our experiments: a) The boosting{based methods clearly outperform the baseline learning algorithms (Naive Bayes and Induction of Decision Trees) on the PU1 corpus, achieving very high levels of the F1 measure � b) Increasing the complexity of the base learners allows to obtain better \high{precision " classi ers, which isavery important issue when misclassication costs are considered. 1