D. Lewis and W. Gale. Training text classifiers by uncertainty sampling. In Proceedings of International ACM Conference on Research and Development in Information Retrieval, pages 3--12, 1994.  Home/Search   Document Not in Database   Summary   Related Articles   Check

....to obtain labeled training data, while unlabeled data are cheaply available. Several methods have been used for coping with the problem which comes from insufficient labeled data, such as Expectation Maximization (EM) 12] 15] selective sampling [3] sub sampling and uncertainty sampling [9]. The proposed sampling method, selfconfident sampling, picks out more promising data from unlabeled data sets to improve a classifier s performance. It is similar to uncertainty sampling in that it predicts the label of unlabeled data on the basis of the learner s confidence which is acquired ....

.... fiwk;j if e j 6= c(d i ) and fcpk ;j 2 e j Figure 3: Domain Experts algorithm. the category. But Sleeping expert did not allow a classifier to have the inconsistent hypotheses. The self confident sampling method which we have proposed in figure 2 shares a property of the uncertainty sampling [9], in that it predicts the label of an unlabeled data on the basis of the learner s confidence which is obtained through the training phase. The examples that are categorized with the least uncertain will be added to the training set in the next iteration. Unlike uncertainty sampling, our method ....

D. D. Lewis and W. Gale. Training text classifiers by uncertainty sampling. In Proc. of Int'l ACM Conf. on Research and Development in Information Retrieval, pages 3--12, 1994.

....relatively straightforward to take a large set of Web pages labeled as positive and negative examples of the concept home page and derive a classifier that predicts whether any given Web page is a home page or not; unfortunately, Web pages are unlabeled. Techniques such as uncertainty sampling [9] reduce the amount of labeled data needed, but do not eliminate the labeling problem. Clustering techniques do not require labeled inputs, and have been applied successfully to large collections of documents (e.g, 2] Indeed, the Web offers fertile ground for document clustering research. ....

Lewis, D. and Gale, W. Training text classifiers by uncertainty sampling. In Proceedings of the Seventeenth Annual International ACMSIGIR Conference on Research and Development in Information Retrieval, 1994.

....Furthermore, the acquired knowledge may be incomplete. Classification knowledge automatically learned from training document is efficient in time and cost, but its accuracy is limited by the employed learning model. There are many text categorization studies in the information retrieval discipline [4, 10, 20, 21, 22, 31]. In this paper, we use document classification instead of text categorization , since we focus on the Internet HTML documents rather than general texts. Document classification is the problem 5 of automatic documents grouping. Many studies also deal with the problem of document retrieval, ....

D. Lewis and William Gale, "Training Text Classifiers by Uncertainty Sampling", ACM SIGIR'94, 1994.

....applied learning methods. We view this result as a confirmation of the usefulness of classifiers that represent contextual information. 1 Introduction Learning methods are frequently used to automatically construct classifiers from labeled documents [Lewis, 1992; Lewis and Ringuette, 1994; Lewis and Gale, 1994; Apt e et al. 1994; Wiener et al. 1995; Cohen, 1995b] In this paper, we will investigate the performance of two recently implemented machine learning algorithms on a number of large text categorization problems. The two algorithms considered are set valued Ripper, a recent rule learning ....

....set for a specific loss ratio. 3 Experimental results 3.1 The AP titles corpus The first benchmark we will use is a corpus of AP newswire headlines, tagged as being relevant or irrelevant to topics like federal budget and Nielsens ratings . This dataset is described in more detail elsewhere [Lewis and Gale, 1994; Lewis and Catlett, 1994] The corpus contains 319,463 documents in the training set and 51,991 documents in the test set. The headlines are an average of nine words long, with a total vocabulary is 67,331 words. No preprocessing of the text was done, other than to convert all words to lower case ....

[Article contains additional citation context not shown here]

David Lewis and William Gale. Training text classifiers by uncertainty sampling. In Seventeenth Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, 1994.

....previously applied learning methods. We view this result as a confirmation of the usefulness of classifiers that represent contextual information. 1 Introduction Learning methods are frequently used to automatically construct classifiers from labeled documents [Lewis and Ringuette, 1994; Lewis and Gale, 1994; Apt e et al. 1994; Wiener et al. 1995; Cohen, 1995b] In this paper, we will investigate the performance of two recently implemented machine learning algorithms on a number of large text categorization problems. The two algorithms considered are set valued Ripper, a recent rule learning ....

....set for a specific loss ratio. 3 Experimental results 3.1 The AP titles corpus The first benchmark we will use is a corpus of AP newswire headlines, tagged as being relevant or irrelevant to topics like federal budget and Nielsens ratings . This dataset is described in more detail elsewhere [Lewis and Gale, 1994; Lewis and Catlett, 1994] The corpus contains 319,463 documents in the training set and 51,991 documents in the test set. The headlines are an average of nine words long, with a total vocabulary is 67,331 words. No preprocessing of the text was done, other than to convert all words to lower case ....

[Article contains additional citation context not shown here]

David Lewis and William Gale. Training text classifiers by uncertainty sampling. In Seventeenth Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, 1994.

....training pages for text categorization problems. However, significant effort is required to assign semantic categories to these pages. Not surprisingly, much of the recent work in active learning has concentrated on text categorization problems. Closely related to windowing is uncertainty sampling (Lewis Gale, 1994; Lewis Catlett, 1994) The difference is that the window is not adjusted on the basis of misclassified examples, but on the basis of the learner s confidence in its own predictions. The examples that are classified with the least confidence will be added to the training set in the next ....

Lewis, D. D., & Gale, W. (1994). Training text classifiers by uncertainty sampling. In Proceedings of the 17th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR-94), pp. 3--12.

....to navigate and sift through a burgeoning growth of the Web documents. Subsequent generations of text search robots have been built on this humble foundation[63] Machine learning approaches (see Knoblock[43] faced the problem of labeling the samples. Techniques such as uncertainty sampling[45] reduce the amount of labeled data needed, but do not eliminate the labeling problem. Clustering techniques do not require labeled inputs, and have been applied successfully to large collections of documents[15] Indeed, the Web offers fertile ground for document clustering research. However, ....

Lewis, D. and Gale, W. "Training text classifiers by uncertainty sampling", Proceedings of the Seventeenth Annual International ACMSIGIR Conference on Research and Development in Information Retrieval, 1994.

....relatively straight forward to take a large set of Web pages labeled as positive and negative examples of the concept home page and derive a classifier that predicts whether any given Web page is a home page or not; unfortunately, Web pages are unlabeled. Techniques such as uncertainty sampling [9] reduce the amount of labeled data needed, but do not eliminate the labeling problem. Clustering techniques do not require labeled inputs, and have been applied successfully to large collections of documents (e.g, 2] Indeed, the Web offers fertile ground for document clustering research. ....

D. Lewis and W. Gale. Training text classifiers by uncertainty sampling. In 17th Annual Int'l ACM SIGIR Conference on Research and Development in Information Retrieval, 1994.

....led us to tackle an important impediment to the use of machine learning on the Web. Data is abundant on the Web, but it is unlabeled. Most concept learning techniques require training data labeled as positive (or negative) examples of some concept. Techniques such as uncertainty sampling (Lewis Gale 1994) reduce the amount of labeled data needed, but do not eliminate the problem. Instead, Ahoy attempts to harness the Web s interactive nature to solve the labeling problem. Ahoy relies on its initial power to draw numerous users to it and to solicit their feedback; it then uses this feedback to ....

Lewis, D., and Gale, W. 1994. Training text classifiers by uncertainty sampling. In 17th Annual Int'l ACM SIGIR Conference on Research and Development in Information Retrieval.

.... In this case the second type of active learning, selective sampling, can be applied: The learner examines many unlabeled examples, and selects only the most informative for learning (Seung, Opper, and Sompolinsky, 1992; Freund et al. 1993; Cohn, Atlas, and Ladner, 1994; Lewis and Catlett, 1994; Lewis and Gale, 1994). In this paper, we address the problem of selective sampling for training a probabilistic classifier. Classification in this framework is performed by a probabilistic model which, given an input example, assigns a probability to each possible classification and selects the most probable one. Our ....

....of the posterior distribution of the estimates is large, and hence there will be large differences in the values of the parameter picked for different committee members. Note that property 1 is not addressed when uncertainty in classification is only judged relative to a single model (as in, eg, (Lewis and Gale, 1994)) Such an approach captures uncertainty with respect to given parameter values, in the sense of property 2, but it does not model uncertainty about the choice of these values in the first place (the use of a single model is also criticized in (Cohn, Atlas, and Ladner, 1994) Property 2 is ....

Lewis, D. and W. Gale. 1994. Training text classifiers by uncertainty sampling. In Proceedings of ACMSIGIR Conference on Information Retrieval.

....redundancy of annotating many examples that contribute roughly the same information to the learner. The machine learning literature suggests several different approaches for selective sampling (Seung, Opper, Sompolinsky 1992; Freund et al. 1993; Cohn, Atlas, Ladner 1994; Lewis Catlett 1994; Lewis Gale 1994). In the first part of the paper, we analyze the different issues that need to be addressed when constructing a selective sampling algorithm. These include measuring the utility of an example for the learner, the number of models that will be used for the selection process, the method for ....

....of an example is evaluated with respect to models derived from the training data at each stage. The key question then is how many models to use to evaluate an example. One approach is to use a single, optimal 1 , model based on the training data seen so far. This approach is taken by Lewis and Gale (1994), for training a binary classifier. They select for training those examples whose classification probability is closest to 0.5, i.e, those examples for which the current model is most uncertain. There are some difficulties with the single model approach, however (Cohn, Atlas, Ladner 1994) First ....

[Article contains additional citation context not shown here]

Lewis, D., and Gale, W. 1994. Training text classifiers by uncertainty sampling. In Proceedings of ACMSIGIR Conference on Information Retrieval.

....instances to add to the training set at each iteration. The key difference is its assumption that the class labels of all training instances are known: it examines them in order to choose misclassified examples to add. A large scale test of uncertainty sampling with a single classifier approach [18] showed that uncertainty sampling could reduce by a factor of up to 500 the amount of data that had to be labeled to achieve a given level of accuracy. 3 Heterogeneous Uncertainty Sampling Uncertainty sampling requires the construction of large numbers (perhaps thousands) of classifiers which are ....

....Sampling with a Probabilistic Classifier Methods for efficient training of probabilistic classifiers from large, sparse data sets are widely used in information retrieval [14] We used this type of classifier to select instances in uncertainty sampling. The model is described in detail elsewhere [18], but in brief it gives the following estimate for the probability that an instance belongs to class C: P (Cjw) exp(a b P d i=1 log P (w i jC) P (w i j C) 1 exp(a b P d i=1 log P (w i jC) P (w i j C) 1) C indicates class membership, and w i is the ith of d attribute ....

David D. Lewis and William A. Gale. Training text classifiers by uncertainty sampling. In Seventeenth Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, 1994. To appear.

No context found.

D. Lewis and W. Gale. Training text classifiers by uncertainty sampling. In Proceedings of International ACM Conference on Research and Development in Information Retrieval, pages 3--12, 1994.

No context found.

D. Lewis and W. Gale. Training text classifiers by uncertainty sampling. In Proceedings of International ACM Conference on Research and Development in Information Retrieval, pages 3--12, 1994.

No context found.

D. Lewis and W. Gale. Training text classifiers by uncertainty sampling. In Proceedings of International ACM Conference on Research and Development in Information Retrieval, pages 3--12, 1994.

No context found.

D. Lewis and W. Gale. Training text classifiers by uncertainty sampling. In Proceedings of International ACM Conference on Research and Development in Information Retrieval, pages 3--12, 1994.

No context found.

Lewis D. and Gale W., 1994. Training text classifiers by uncertainty sampling. In Proceedings of ACM-SIGIR Conference on Information Retrieval.

No context found.

David Lewis and William Gale. Training text classifiers by uncertainty sampling. In Seventeenth Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, 1994.

No context found.

David Lewis and William Gale. Training text classifiers by uncertainty sampling. In Seventeenth Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, 1994.

No context found.

David Lewis and William Gale. Training text classifiers by uncertainty sampling. In Seventeenth Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, 1994.

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC