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. 1

We present a simple and effective semisupervised method for training dependency parsers. We focus on the problem of lexical representation, introducing features that incorporate word clusters derived from a large unannotated corpus. We demonstrate the effectiveness of the approach in a series of dependency parsing experiments on the Penn Treebank and Prague Dependency Treebank, and we show that the cluster-based features yield substantial gains in performance across a wide range of conditions. For example, in the case of English unlabeled second-order parsing, we improve from a baseline accuracy of 92.02 % to 93.16%, and in the case of Czech unlabeled second-order parsing, we improve from a baseline accuracy of 86.13% to 87.13%. In addition, we demonstrate that our method also improves performance when small amounts of training data are available, and can roughly halve the amount of supervised data required to reach a desired level of performance. 1

We demonstrate that log-linear grammars with latent variables can be practically trained using discriminative methods. Central to efficient discriminative training is a hierarchical pruning procedure which allows feature expectations to be efficiently approximated in a gradient-based procedure. We compare L1 and L2 regularization and show that L1 regularization is superior, requiring fewer iterations to converge, and yielding sparser solutions. On full-scale treebank parsing experiments, the discriminative latent models outperform both the comparable generative latent models as well as the discriminative non-latent baselines. 1

We propose a generative dependency parsing model which uses binary latent variables to induce conditioning features. To define this model we use a recently proposed class of Bayesian Networks for structured prediction, Incremental Sigmoid Belief Networks. We demonstrate that the proposed model achieves state-of-the-art results on three different languages. We also demonstrate that the features induced by the ISBN’s latent variables are crucial to this success, and show that the proposed model is particularly good on long dependencies. 1

This thesis is about estimating probabilistic models to uncover useful hidden structure in data; specifically, we address the problem of discovering syntactic structure in natural language text. We present three new parameter estimation techniques that generalize the standard approach, maximum likelihood estimation, in different ways. Contrastive estimation maximizes the conditional probability of the observed data given a “neighborhood” of implicit negative examples. Skewed deterministic annealing locally maximizes likelihood using a cautious parameter search strategy that starts with an easier optimization problem than likelihood, and iteratively moves to harder problems, culminating in likelihood. Structural annealing is similar, but starts with a heavy bias toward simple syntactic structures and gradually relaxes the bias. Our estimation methods do not make use of annotated examples. We consider their performance in both an unsupervised model selection setting, where models trained under different initialization and regularization settings are compared by evaluating the training objective on a small set of unseen, unannotated development data, and supervised model selection, where the most accurate model on the development set (now with annotations)

A notable gap in research on statistical dependency parsing is a proper conditional probability distribution over nonprojective dependency trees for a given sentence. We exploit the Matrix Tree Theorem (Tutte, 1984) to derive an algorithm that efficiently sums the scores of all nonprojective trees in a sentence, permitting the definition of a conditional log-linear model over trees. While discriminative methods, such as those presented in McDonald et al. (2005b), obtain very high accuracy on standard dependency parsing tasks and can be trained and applied without marginalization, “summing trees ” permits some alternative techniques of interest. Using the summing algorithm, we present competitive experimental results on four nonprojective languages, for maximum conditional likelihood estimation, minimum Bayes-risk parsing, and hidden variable training. 1

In this paper we approach word sense disambiguation and information extraction as a unified tagging problem. The task consists of annotating text with the tagset defined by the 41 Wordnet supersense classes for nouns and verbs. Since the tagset is directly related to Wordnet synsets, the tagger returns partial word sense disambiguation. Furthermore, since the noun tags include the standard named entity detection classes – person, location, organization, time, etc. – the tagger, as a by-product, returns extended named entity information. We cast the problem of supersense tagging as a sequential labeling task and investigate it empirically with a discriminatively-trained Hidden Markov Model. Experimental evaluation on the main sense-annotated datasets available, i.e., Semcor and Senseval, shows considerable improvements over the best known “first-sense ” baseline. 1

Creating large amounts of annotated data to train statistical PCFG parsers is expensive, and the performance of such parsers declines when training and test data are taken from different domains. In this paper we use selftraining in order to improve the quality of a parser and to adapt it to a different domain, using only small amounts of manually annotated seed data. We report significant improvement both when the seed and test data are in the same domain and in the outof-domain adaptation scenario. In particular, we achieve 50 % reduction in annotation cost for the in-domain case, yielding an improvement of 66 % over previous work, and a 20-33 % reduction for the domain adaptation case. This is the first time that self-training with small labeled datasets is applied successfully to these tasks. We were also able to formulate a characterization of when selftraining is valuable.

We present a discriminative, latent variable approach to syntactic parsing in which rules exist at multiple scales of refinement. The model is formally a latent variable CRF grammar over trees, learned by iteratively splitting grammar productions (not categories). Different regions of the grammar are refined to different degrees, yielding grammars which are three orders of magnitude smaller than the single-scale baseline and 20 times smaller than the split-and-merge grammars of Petrov et al. (2006). In addition, our discriminative approach integrally admits features beyond local tree configurations. We present a multiscale training method along with an efficient CKY-style dynamic program. On a variety of domains and languages, this method produces the best published parsing accuracies with the smallest reported grammars. 1

While the average performance of statistical parsers gradually improves, they still attach to many sentences annotations of rather low quality. The number of such sentences grows when the training and test data are taken from different domains, which is the case for major web applications such as information retrieval and question answering. In this paper we present a Sample Ensemble Parse Assessment (SEPA) algorithm for detecting parse quality. We use a function of the agreement among several copies of a parser, each of which trained on a different sample from the training data, to assess parse quality. We experimented with both generative and reranking parsers (Collins, Charniak and Johnson respectively). We show superior results over several baselines, both when the training and test data are from the same domain and when they are from different domains. For a test setting used by previous work, we show an error reduction of 31 % as opposed to their 20%. 1