No abstract

We discuss data-driven morphological segmentation, in which word forms are segmented into morphs, that is the surface forms of morphemes. We extend a recent segmentation approach based on conditional random fields from purely supervised to semi-supervised learning by exploiting available unsupervised segmentation techniques. We integrate the unsupervised techniques into the conditional random field model via feature set augmentation.Experiments on three diverse languages show that this straightforward semi-supervised extension greatly improves the segmentation accuracy of the purely supervised CRFs in a computationally efficient manner.

gram models are the most widely used language models in large vocabulary continuous speech recognition. Since the size of the model grows rapidly with respect to the model order and available training data, many methods have been proposed for pruning the least relevant -grams from the model. However, correct smoothing of the -gram probability distributions is important and performance may degrade significantly if pruning conflicts with smoothing. In this paper, we show that some of the commonly used pruning methods do not take into account how removing an -gram should modify the backoff distributions in the state-of-the-art Kneser-Ney smoothing. To solve this problem, we present two new algorithms: one for pruning Kneser-Ney smoothed models, and one for growing them incrementally. Experiments on Finnish and English text corpora show that the proposed pruning algorithm provides considerable improvements over previous pruning algorithms on Kneser-Ney-smoothed models and is also better than the baseline entropy pruned Good-Turing smoothed models. The models created by the growing algorithm provide a good starting point for our pruning algorithm, leading to further improvements. The improvements in the Finnish speech recognition over the other Kneser-Ney smoothed models are statistically significant, as well.

Morfessor is a family of probabilistic machine learning methods for finding the morphological segmentation from raw text data. Recent developments include the development of semi-supervised methods for utilizing annotated data. Morfessor 2.0 is a rewrite of the original, widely-used Morfessor 1.0 software, with well documented command-line tools and library interface. It includes new features such as semi-supervised learning, online training, and integrated evaluation code.

Determining optimal units of representing morphologically complex words in the mental lexicon is a central question in psycholinguistics. Here, we utilize advances in computational sciences to study human morphological processing using statistical models of morphology, particularly the unsupervised Morfessor model that works on the principle of optimization. The aim was to see what kind of model structure corresponds best to human word recognition costs for multimorphemic Finnish nouns: a model incorporating units resembling linguistically defined morphemes, a whole-word model, or a model that seeks for an optimal balance between these two extremes. Our results showed that human word recognition was predicted best by a combination of two models: a model that decomposes words at some morpheme boundaries while keeping others unsegmented and a whole-word model. The results support dual-route models that assume that both decomposed and full-form representations are utilized to optimally process complex words within the mental lexicon.