Conditional random fields (Lafferty et al., 2001) are quite effective at sequence labeling tasks like shallow parsing (Sha and Pereira, 2003) and namedentity extraction (McCallum and Li, 2003). CRFs are log-linear, allowing the incorporation of arbitrary features into the model. To train on unlabeled data, we require unsupervised estimation methods for log-linear models; few exist. We describe a novel approach, contrastive estimation. We show that the new technique can be intuitively understood as exploiting implicit negative evidence and is computationally efficient. Applied to a sequence labeling problem—POS tagging given a tagging dictionary and unlabeled text—contrastive estimation outperforms EM (with the same feature set), is more robust to degradations of the dictionary, and can largely recover by modeling additional features. 1

We first show how a structural locality bias can improve the accuracy of state-of-the-art dependency grammar induction models trained by EM from unannotated examples (Klein and Manning, 2004). Next, by annealing the free parameter that controls this bias, we achieve further improvements. We then describe an alternative kind of structural bias, toward “broken ” hypotheses consisting of partial structures over segmented sentences, and show a similar pattern of improvement. We relate this approach to contrastive estimation (Smith and Eisner, 2005a), apply the latter to grammar induction in six languages, and show that our new approach improves accuracy by 1–17 % (absolute) over CE (and 8–30% over EM), achieving to our knowledge the best results on this task to date. Our method, structural annealing, is a general technique with broad applicability to hidden-structure discovery problems. 1

We describe a novel training criterion for probabilistic grammar induction models, contrastive estimation [Smith and Eisner, 2005], which can be interpreted as exploiting implicit negative evidence and includes a wide class of likelihood-based objective functions. This criterion is a generalization of the function maximized by the Expectation-Maximization algorithm [Dempster et al., 1977]. CE is a natural fit for log-linear models, which can include arbitrary features but for which EM is computationally difficult. We show that, using the same features, log-linear dependency grammar models trained using CE can drastically outperform EMtrained generative models on the task of matching human linguistic annotations (the MATCHLIN-GUIST task). The selection of an implicit negative evidence class—a “neighborhood”—appropriate to a given task has strong implications, but a good neighborhood one can target the objective of grammar induction to a specific application. 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)

We present three approaches for unsupervised grammar induction that are sensitive to data complexity and apply them to Klein and Manning’s Dependency Model with Valence. The first, Baby Steps, bootstraps itself via iterated learning of increasingly longer sentences and requires no initialization. This method substantially exceeds Klein and Manning’s published scores and achieves 39.4 % accuracy on Section 23 (all sentences) of the Wall Street Journal corpus. The second, Less is More, uses a low-complexity subset of the available data: sentences up to length 15. Focusing on fewer but simpler examples trades off quantity against ambiguity; it attains 44.1% accuracy, using the standard linguisticallyinformed prior and batch training, beating state-of-the-art. Leapfrog, our third heuristic, combines Less is More with Baby Steps by mixing their models of shorter sentences, then rapidly ramping up exposure to the full training set, driving up accuracy to 45.0%. These trends generalize to the Brown corpus; awareness of data complexity may improve other parsing models and unsupervised algorithms.

Weighted deduction with aggregation is a powerful theoretical formalism that encompasses many NLP algorithms. This paper proposes a declarative specification language, Dyna; gives general agenda-based algorithms for computing weights and gradients; briefly discusses Dyna-to-Dyna program transformations; and shows that a first implementation of a Dyna-to-C++ compiler produces code that is efficient enough for real NLP research, though still several times slower than hand-crafted code. 1

We present a simple but very effective approach to identifying high-quality data in noisy data sets for structured problems like parsing, by greedily exploiting partial structures. We analyze our approach in an annotation projection framework for dependency trees, and show how dependency parsers from two different paradigms (graph-based and transition-based) can be trained on the resulting tree fragments. We train parsers for Dutch to evaluate our method and to investigate to which degree graph-based and transitionbased parsers can benefit from incomplete training data. We find that partial correspondence projection gives rise to parsers that outperform parsers trained on aggressively filtered data sets, and achieve unlabeled attachment scores that are only 5 % behind the average UAS for Dutch in the CoNLL-X Shared Task on supervised parsing (Buchholz and

Clustering is a central technique in NLP. Consequently, clustering evaluation is of great importance. Many clustering algorithms are evaluated by their success in tagging corpus tokens. In this paper we discuss type level evaluation, which reflects class membership only and is independent of the token statistics of a particular reference corpus. Type level evaluation casts light on the merits of algorithms, and for some applications is a more natural measure of the algorithm’s quality. We propose new type level evaluation measures that, contrary to existing measures, are applicable when items are polysemous, the common case in NLP. We demonstrate the benefits of our measures using a detailed case study, POS induction. We experiment with seven leading algorithms, obtaining useful insights and showing that token and type level measures can weakly or even negatively correlate, which underscores the fact that these two approaches reveal different aspects of clustering quality. 1

We consider a new subproblem of unsupervised parsing from raw text, unsupervised partial parsing—the unsupervised version of text chunking. We show that addressing this task directly, using probabilistic finite-state methods, produces better results than relying on the local predictions of a current best unsupervised parser, Seginer’s (2007) CCL. These finite-state models are combined in a cascade to produce more general (full-sentence) constituent structures; doing so outperforms CCL by a wide margin in unlabeled PARSEVAL scores for English, German and Chinese. Finally, we address the use of phrasal punctuation

We present new training methods that aim to mitigate local optima and slow convergence in unsupervised training by using additional imperfect objectives. In its simplest form, lateen EM alternates between the two objectives of ordinary “soft ” and “hard ” expectation maximization (EM) algorithms. Switching objectives when stuck can help escape local optima. We find that applying a single such alternation already yields state-of-the-art results for English dependency grammar induction. More elaborate lateen strategies track both objectives, with each validating the moves proposed by the other. Disagreements can signal earlier opportunities to switch or terminate, saving iterations. De-emphasizing fixed points in these ways eliminates some guesswork from tuning EM. An evaluation against a suite of unsupervised dependency parsing tasks, for a variety of languages, showed that lateen strategies significantly speed up training of both EM algorithms, and improve accuracy for hard EM. 1