2008
|
Learning from Labeled Features using Generalized Expectation Criteria.
Gregory Druck, Gideon Mann, Andrew McCallum.
To appear in Proceedings of SIGIR.
A version of this paper is also in the Proceedings of NESCAI.
We use generalized expectation (GE) criteria to train discriminative probabilistic models with "labeled features" and unlabeled instances.
Experimental results on text classification data sets show that this method outperforms heuristic approaches to training classifiers with labeled features.
Experiments with human annotators show that it is more beneficial to spend limited annotation time labeling features rather than labeling instances.
|
|
|
|
Learning to Predict the Quality of Contributions to Wikipedia.
Gregory Druck, Gerome Miklau, Andrew McCallum.
To appear in AAAI Workshop on Wikipedia and AI.
We propose metrics that quantify the quality of contributions to Wikipedia through implicit feedback from the community and learn discriminative
probabilistic models that predict the quality of a new edit using features of the changes made, the author of the edit, and the article being edited. Through estimating parameters for these models, we
learn about factors that influence quality. We advocate using this model to
alert users to potential quality problems.
|
|
|
2007
|
Leveraging Existing Resources using Generalized Expectation Criteria.Gregory Druck, Gideon Mann, Andrew
McCallum. In NIPS Workshop on Learning Problem Design
Updated: 12/17/07
We use generalized expectation (GE) criteria to train a sliding window named entity recognizer using only lexicons (with known class associations) and unlabeled data.
|
|
|
|
Reducing Annotation Effort using Generalized Expectation Criteria.Gregory
Druck, Gideon Mann, Andrew McCallum.
University of Massachusetts Amherst Tech Report 07-62
This technical report evolved into our SIGIR 2008 paper (above).
|
|
|
|
Generalized Expectation Criteria. (DRAFT)Andrew McCallum, Gideon Mann, Gregory Druck.
University of Massachusetts Amherst Tech Report 07-60
This note introduces and motivates generalized expectation (GE) criteria, terms in a parameter estimation objective function that express preferences about model expectations.
|
|
|
|
Semi-Supervised Classification with Hybrid Generative/Discriminative Methods.Gregory Druck, Chris Pal, Xiaojin Zhu, Andrew
McCallum. In Proceedings of KDD.
We compare two recently proposed frameworks for combining generative and discriminative probabilistic classifiers and apply them to semi-supervised classification.
While prominent semi-supervised learning methods assume low density regions between classes or are subject to generative modeling assumptions, we conjecture that hybrid generative/discriminative methods allow
semi-supervised learning in the presence of strongly overlapping classes and reduce the risk of modeling structure in the unlabeled data that is irrelevant for the specific classification task of interest. We apply both hybrid approaches within naively structured
Markov random field models and provide a thorough empirical comparison with two well-known semi-supervised learning methods on six text classification tasks.
|
|
|
|
Learning A* Underestimates: Using Inference to Guide Inference. Gregory Druck, Mukund Narasimhan, Paul Viola.
In Proceedings of AISTATS
It is difficult to use A* search in practice because an estimate of the cost to the goal that is optimistic (to ensure an optimal solution) and tight (to prune away enough of the search space to overcome the added overhead) is required.
We introduce the notion of a probable approximate overestimate, and show that we can learn such a function from data.
The overestimate function takes the functional form of a simpler model in which inference is faster.
We apply this idea to semi-Markov conditional random fields and show that inference using this approximate A* algorithm is about three times as fast as the conventional Viterbi algorithm, with only a
minor degradation in accuracy.
|
|
|
2006
|
Multi-Conditional Learning: Generative/Discriminative Training for Clustering and Classification. Andrew McCallum, Chris Pal, Gregory Druck, Xuerui Wang.
In Proceedings of AAAI.
We introduce a class of training objective functions for graphical models in which a single set of parameters is maximized with respect to the product of multiple weighted conditional likelihoods. Generative/discriminative hybrid models, semi-supervised learning, and transfer or multi-task learning can all easily be described using this framework. In this paper, we apply a multi-conditional objective function combining discriminative and generative likelihood components in naively-structured Markov random field and
harmonium models. Experimental results show improvements in accuracy for text classification and the ability to discover quantitatively improved latent spaces.
|
|
|
