Benjamin Taskar

Alignment by Agreement

We present an unsupervised approach to symmetric word alignment in which two simple asymmetric models are trained jointly to maximize a combination of data likelihood and agreement between the models. Compared to the standard practice of intersecting predictions of independently-trained models, joint training provides a 32% reduction in AER. Moreover, a simple and efficient pair of HMM aligners provides a 29% reduction in AER over symmetrized IBM model 4 predictions.

Learning structured prediction models: a large margin approach

We consider large margin estimation in a broad range of prediction models where inference involves solving combinatorial optimization problems, for example, weighted graph-cuts or matchings. Our goal is to learn parameters such that inference using the model reproduces correct answers on the training data. Our method relies on the expressive power of convex optimization problems to compactly capture inference or solution optimality in structured prediction models. Directly embedding this structure within the learning formulation produces concise convex problems for efficient estimation of very complex and diverse models. We describe experimental results on a matching task, disulfide connectivity prediction, showing significant improvements over state-of-the-art methods.

An End-to-End Discriminative Approach to Machine Translation

We present a perceptron-style discriminative approach to machine translation in which large feature sets can be exploited. Unlike discriminative reranking approaches, our system can take advantage of learned features in all stages of decoding. We first discuss several challenges to error-driven discriminative approaches. In particular, we explore different ways of updating parameters given a training example. We find that making frequent but smaller updates is preferable to making fewer but larger updates. Then, we discuss an array of features and show both how they quantitatively increase BLEU score and how they qualitatively interact on specific examples. One particular feature we investigate is a novel way to introduce learning into the initial phrase extraction process, which has previously been entirely heuristic.

Selectivity estimation using probabilistic models

Estimating the result size of complex queries that involve selection on multiple attributes and the join of several relations is a difficult but fundamental task in database query processing. It arises in cost-based query optimization, query profiling, and approximate query answering. In this paper, we show how probabilistic graphical models can be effectively used for this task as an accurate and compact approximation of the joint frequency distribution of multiple attributes across multiple relations. Probabilistic Relational Models (PRMs) are a recent development that extends graphical statistical models such as Bayesian Networks to relational domains. They represent the statistical dependencies between attributes within a table, and between attributes across foreign-key joins. We provide an efficient algorithm for constructing a PRM front a database, and show how a PRM can be used to compute selectivity estimates for a broad class of queries. One of the major contributions of this work is a unified framework for the estimation of queries involving both select and foreign-key join operations. Furthermore, our approach is not limited to answering a small set of predetermined queries; a single model can be used to effectively estimate the sizes of a wide collection of potential queries across multiple tables. We present results for our approach on several real-world databases. For both single-table multi-attribute queries and a general class of select-join queries, our approach produces more accurate estimates than standard approaches to selectivity estimation, using comparable space and time.

Learning from ambiguously labeled images

In many image and video collections, we have access only to partially labeled data. For example, personal photo collections often contain several faces per image and a caption that only specifies who is in the picture, but not which name matches which face. Similarly, movie screenplays can tell us who is in the scene, but not when and where they are on the screen. We formulate the learning problem in this setting as partially-supervised multiclass classification where each instance is labeled ambiguously with more than one label. We show theoretically that effective learning is possible under reasonable assumptions even when all the data is weakly labeled. Motivated by the analysis, we propose a general convex learning formulation based on minimization of a surrogate loss appropriate for the ambiguous label setting. We apply our framework to identifying faces culled from Web news sources and to naming characters in TV series and movies. We experiment on a very large dataset consisting of 100 hours of video, and in particular achieve 6% error for character naming on 16 episodes of LOST.

Word Alignment via Quadratic Assignment

Recently, discriminative word alignment methods have achieved state-of-the-art accuracies by extending the range of information sources that can be easily incorporated into aligners. The chief advantage of a discriminative framework is the ability to score alignments based on arbitrary features of the matching word tokens, including orthographic form, predictions of other models, lexical context and so on. However, the proposed bipartite matching model of Taskar et al. (2005), despite being tractable and effective, has two important limitations. First, it is limited by the restriction that words have fertility of at most one. More importantly, first order correlations between consecutive words cannot be directly captured by the model. In this work, we address these limitations by enriching the model form. We give estimation and inference algorithms for these enhancements. Our best model achieves a relative AER reduction of 25% over the basic matching formulation, outperforming intersected IBM Model 4 without using any overly compute-intensive features. By including predictions of other models as features, we achieve AER of 3.8 on the standard Hansards dataset.

Controlling complexity in part-of-speech induction

We consider the problem of fully unsupervised learning of grammatical (part-of-speech) categories from unlabeled text. The standard maximum-likelihood hidden Markov model for this task performs poorly, because of its weak inductive bias and large model capacity. We address this problem by refining the model and modifying the learning objective to control its capacity via parametric and non-parametric constraints. Our approach enforces word-category association sparsity, adds morphological and orthographic features, and eliminates hard-to-estimate parameters for rare words. We develop an efficient learning algorithm that is not much more computationally intensive than standard training. We also provide an open-source implementation of the algorithm. Our experiments on five diverse languages (Bulgarian, Danish, English, Portuguese, Spanish) achieve significant improvements compared with previous methods for the same task.

A permutation-augmented sampler for DP mixture models

We introduce a new inference algorithm for Dirichlet process mixture models. While Gibbs sampling and variational methods focus on local moves, the new algorithm makes more global moves. This is done by introducing a permutation of the data points as an auxiliary variable. The algorithm is a blocked sampler which alternates between sampling the clustering and sampling the permutation. The key to the efficiency of this approach is that it is possible to use dynamic programming to consider all exponentially many clusterings consistent with a given permutation. We also show that random projections can be used to effectively sample the permutation. The result is a stochastic hill-climbing algorithm that yields burn-in times significantly smaller than those of collapsed Gibbs sampling.