We show that Viterbi (or “hard”) EM is well-suited to unsupervised grammar induction. It is more accurate than standard inside-outside re-estimation (classic EM), significantly faster, and simpler. Our experiments with Klein and Manning’s Dependency Model with Valence (DMV) attain state-of-the-art performance — 44.8% accuracy on Section 23 (all sentences) of the Wall Street Journal corpus — without clever initialization; with a good initializer, Viterbi training improves to 47.9%. This generalizes to the Brown corpus, our held-out set, where accuracy reaches 50.8 % — a 7.5 % gain over previous best results. We find that classic EM learns better from short sentences but cannot cope with longer ones, where Viterbi thrives. However, we explain that both algorithms optimize the wrong objectives and prove that there are fundamental disconnects between the likelihoods of sentences, best parses, and true parses, beyond the wellestablished discrepancies between likelihood, accuracy and extrinsic performance. 1

We show how punctuation can be used to improve unsupervised dependency parsing. Our linguistic analysis confirms the strong connection between English punctuation and phrase boundaries in the Penn Treebank. However, approaches that naively include punctuation marks in the grammar (as if they were words) do not perform well with Klein and Manning’s Dependency Model with Valence (DMV). Instead, we split a sentence at punctuation and impose parsing restrictions over its fragments. Our grammar inducer is trained on the Wall Street Journal (WSJ) and achieves 59.5 % accuracy out-of-domain (Brown sentences with 100 or fewer words), more than 6 % higher than the previous best results. Further evaluation, using the 2006/7 CoNLL sets, reveals that punctuation aids grammar induction in 17 of 18 languages, for an overall average net gain of 1.3%. Some of this improvement is from training, but more than half is from parsing with induced constraints, in inference. Punctuation-aware decoding works with existing (even already-trained) parsing models and always increased accuracy in our experiments. 1

Probabilistic grammars offer great flexibility in modeling discrete sequential data like natural language text. Their symbolic component is amenable to inspection by humans, while their probabilistic component helps resolve ambiguity. They also permit the use of well-understood, generalpurpose learning algorithms. There has been an increased interest in using probabilistic grammars in the Bayesian setting. To date, most of the literature has focused on using a Dirichlet prior. The Dirichlet prior has several limitations, including that it cannot directly model covariance between the probabilistic grammar’s parameters. Yet, various grammar parameters are expected to be correlated because the elements in language they represent share linguistic properties. In this paper, we suggest an alternative to the Dirichlet prior, a family of logistic normal distributions. We derive an inference algorithm for this family of distributions and experiment with the task of dependency grammar induction, demonstrating performance improvements with our priors on a set of six treebanks in different natural languages. Our covariance framework permits soft parameter tying within grammars and across grammars for text in different languages, and we show empirical gains in a novel learning setting using bilingual, non-parallel data.

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

We show that orthographic cues can be helpful for unsupervised parsing. In the Penn Treebank, transitions between upper- and lowercase tokens tend to align with the boundaries of base (English) noun phrases. Such signals can be used as partial bracketing constraints to train a grammar inducer: in our experiments, directed dependency accuracy increased by 2.2 % (average over 14 languages having case information). Combining capitalization with punctuation-induced constraints in inference further improved parsing performance, attaining state-of-the-art levels for many languages. 1

The unsupervised Data Oriented Parsing (uDOP) approach has been repeatedly reported to achieve state of the art performance in experiments on parsing of different corpora. At the same time the approach is demanding both in computation time and memory. This paper describes an approach which decreases these demands. First the problem is translated into the generation of probabilistic bottom up tree automata (pBTA). Then it is explained how solving two standard problems for these automata results in a reduction in the size of the grammar. The reduction of the grammar size by using efficient algorithms for pBTAs is the main contribution of this paper. Experiments suggest that this leads to a reduction in grammar size by a factor of 2. This paper also suggests some extensions of the original uDOP algorithm that are made possible or aided by the use of tree automata. 1

The work presented in this thesis is largely based on a work presented in

This paper investigates improving supervised word segmentation accuracy with unlabeled data. Both large-scale in-domain data and small-scale document text are considered. We present a unified solution to include features derived from unlabeled data to a discriminative learning model. For the large-scale data, we derive string statistics from Gigaword to assist a character-based segmenter. In addition, we introduce the idea about transductive, document-level segmentation, which is designed to improve the system recall for out-ofvocabulary (OOV) words which appear more than once inside a document. Novel features1 result in relative error reductions of 13.8 % and 15.4 % in terms of F-score and the recall of OOV words respectively. 1