Andrew McCallum

Text Classification from Labeled and Unlabeled Documents using EM

This paper shows that the accuracy of learned text classifiers can be improved by augmenting a small number of labeled training documents with a large pool of unlabeled documents. This is important because in many text classification problems obtaining training labels is expensive, while large quantities of unlabeled documents are readily available.We introduce an algorithm for learning from labeled and unlabeled documents based on the combination of Expectation-Maximization (EM) and a naive Bayes classifier. The algorithm first trains a classifier using the available labeled documents, and probabilistically labels the unlabeled documents. It then trains a new classifier using the labels for all the documents, and iterates to convergence. This basic EM procedure works well when the data conform to the generative assumptions of the model. However these assumptions are often violated in practice, and poor performance can result. We present two extensions to the algorithm that improve classification accuracy under these conditions: (1) a weighting factor to modulate the contribution of the unlabeled data, and (2) the use of multiple mixture components per class. Experimental results, obtained using text from three different real-world tasks, show that the use of unlabeled data reduces classification error by up to 30%.

Efficient clustering of high-dimensional data sets with application to reference matching

important problems involve clustering large datasets. Although naive implementations of clustering are computa- tionally expensive, there are established ecient techniques for clustering when the dataset has either (1) a limited num- ber of clusters, (2) a low feature dimensionality, or (3) a small number of data points. However, there has been much less work on methods of eciently clustering datasets that are large in all three ways at once|for example, having millions of data points that exist in many thousands of di- mensions representing many thousands of clusters. We present a new technique for clustering these large, high- dimensional datasets. The key idea involves using a cheap, approximate distance measure to eciently divide the data into overlapping subsets we call canopies .T hen cluster- ing is performed by measuring exact distances only between points that occur in a common canopy. Using canopies, large clustering problems that were formerly impossible become practical. Under reasonable assumptions about the cheap distance metric, this reduction in computational cost comes without any loss in clustering accuracy. Canopies can be applied to many domains and used with a variety of cluster- ing approaches, including Greedy Agglomerative Clustering, K-means and Expectation-Maximization. We present ex- perimental results on grouping bibliographic citations from the reference sections of research papers. Here the canopy approach reduces computation time over a traditional clus- tering approach by more than an order of magnitude and decreases error in comparison to a previously used algorithm by 25%.

Early results for named entity recognition with conditional random fields, feature induction and web-enhanced lexicons

Models for many natural language tasks benefit from the flexibility to use overlapping, non-independent features. For example, the need for labeled data can be drastically reduced by taking advantage of domain knowledge in the form of word lists, part-of-speech tags, character n-grams, and capitalization patterns. While it is difficult to capture such inter-dependent features with a generative probabilistic model, conditionally-trained models, such as conditional maximum entropy models, handle them well. There has been significant work with such models for greedy sequence modeling in NLP (Ratnaparkhi, 1996; Borthwick et al., 1998).

Dynamic conditional random fields: factorized probabilistic models for labeling and segmenting sequence data

In sequence modeling, we often wish to represent complex interaction between labels, such as when performing multiple, cascaded labeling tasks on the same sequence, or when long-range dependencies exist. We present dynamic conditional random fields (DCRFs), a generalization of linear-chain conditional random fields (CRFs) in which each time slice contains a set of state variables and edges---a distributed state representation as in dynamic Bayesian networks (DBNs)---and parameters are tied across slices. Since exact inference can be intractable in such models, we perform approximate inference using several schedules for belief propagation, including tree-based reparameterization (TRP). On a natural-language chunking task, we show that a DCRF performs better than a series of linear-chain CRFs, achieving comparable performance using only half the training data.

Distributional clustering of words for text classification

This paper describes the application of Distributional Clustering [20] to document classification. This approach clusters words into groups based on the distribution of class labels associated with each word. Thus, unlike some other unsupervised dimensionalityreduction techniques, such as Latent Semantic Indexing, we are able to compress the feature space much more aggressively, while still maintaining high document classification accuracy. Experimental results obtained on three real-world data sets show that we can reduce the feature dimensional&y by three orders of magnitude and lose only 2% accuracy-significantly better than Latent Semantic Indexing [6], class-based clustering [l], feature selection by mutual information [23], or Markov-blanket-based feature selection [13]. We also show that less aggressive clustering sometimes results in improved classification accuracy over classification without clustering.

Automating the Construction of Internet Portals with Machine Learning

Domain-specific internet portals are growing in popularity because they gather content from the Web and organize it for easy access, retrieval and search. For example, www.campsearch.com allows complex queries by age, location, cost and specialty over summer camps. This functionality is not possible with general, Web-wide search engines. Unfortunately these portals are difficult and time-consuming to maintain. This paper advocates the use of machine learning techniques to greatly automate the creation and maintenance of domain-specific Internet portals. We describe new research in reinforcement learning, information extraction and text classification that enables efficient spidering, the identification of informative text segments, and the population of topic hierarchies. Using these techniques, we have built a demonstration system: a portal for computer science research papers. It already contains over 50,000 papers and is publicly available at www.cora.justresearch.com. These techniques are widely applicable to portal creation in other domains.

Pachinko allocation: DAG-structured mixture models of topic correlations

Latent Dirichlet allocation (LDA) and other related topic models are increasingly popular tools for summarization and manifold discovery in discrete data. However, LDA does not capture correlations between topics. In this paper, we introduce the pachinko allocation model (PAM), which captures arbitrary, nested, and possibly sparse correlations between topics using a directed acyclic graph (DAG). The leaves of the DAG represent individual words in the vocabulary, while each interior node represents a correlation among its children, which may be words or other interior nodes (topics). PAM provides a flexible alternative to recent work by Blei and Lafferty (2006), which captures correlations only between pairs of topics. Using text data from newsgroups, historic NIPS proceedings and other research paper corpora, we show improved performance of PAM in document classification, likelihood of held-out data, the ability to support finer-grained topics, and topical keyword coherence.

Learning to construct knowledge bases from the World Wide Web

Abstract The World Wide Web is a vast source of information accessible to computers, but understandable only to humans. The goal of the research described here is to automatically create a computer understandable knowledge base whose content mirrors that of the World Wide Web. Such a knowledge base would enable much more effective retrieval of Web information, and promote new uses of the Web to support knowledge-based inference and problem solving. Our approach is to develop a trainable information extraction system that takes two inputs. The first is an ontology that defines the classes (e.g., company , person , employee , product ) and relations (e.g., employed_by , produced_by ) of interest when creating the knowledge base. The second is a set of training data consisting of labeled regions of hypertext that represent instances of these classes and relations. Given these inputs, the system learns to extract information from other pages and hyperlinks on the Web. This article describes our general approach, several machine learning algorithms for this task, and promising initial results with a prototype system that has created a knowledge base describing university people, courses, and research projects.

Chinese segmentation and new word detection using conditional random fields

Chinese word segmentation is a difficult, important and widely-studied sequence modeling problem. This paper demonstrates the ability of linear-chain conditional random fields (CRFs) to perform robust and accurate Chinese word segmentation by providing a principled framework that easily supports the integration of domain knowledge in the form of multiple lexicons of characters and words. We also present a probabilistic new word detection method, which further improves performance. Our system is evaluated on four datasets used in a recent comprehensive Chinese word segmentation competition. State-of-the-art performance is obtained.

Topic and role discovery in social networks with experiments on enron and academic email

Previous work in social network analysis (SNA) has modeled the existence of links from one entity to another, but not the attributes such as language content or topics on those links. We present the Author-Recipient-Topic (ART) model for social network analysis, which learns topic distributions based on the direction-sensitive messages sent between entities. The model builds on Latent Dirichlet Allocation (LDA) and the Author-Topic (AT) model, adding the key attribute that distribution over topics is conditioned distinctly on both the sender and recipient--steering the discovery of topics according to the relationships between people. We give results on both the Enron email corpus and a researchers email archive, providing evidence not only that clearly relevant topics are discovered, but that the ART model better predicts peoples roles and gives lower perplexity on previously unseen messages. We also present the Role-Author-Recipient-Topic (RART) model, an extension to ART that explicitly represents peoples roles.