Narrow Results

Neural Network Language Models (NNLMs) are a successful approach to Natural Language Processing tasks, such as Machine Translation. We introduce in this work a Statistical Machine Translation (SMT) system which fully integrates NNLMs in the decoding stage, breaking the traditional approach based on $n$-best list rescoring. The neural net models (both language models (LMs) and translation models) are fully coupled in the decoding stage, allowing to more strongly influence the translation quality. Computational issues were solved by using a novel idea based on memorization and smoothing of the softmax constants to avoid their computation, which introduces a trade-off between LM quality and computational cost. These ideas were studied in a machine translation task with different combinations of neural networks used both as translation models and as target LMs, comparing phrase-based and $N$-gram-based systems, showing that the integrated approach seems more promising for $N$-gram-based systems, even with nonfull-quality NNLMs.

We address the problem of smoothing the probability distribution defined by a finite state automaton. Our approach extends the ideas employed for smoothing n-gram models. This extension is obtained by interpreting n-gram models as finite state models. The experiments show that our smoothing improves perplexity over smoothed n-grams and Error Correcting Parsing techniques.

The RNA sentences present structured regions caused by pairwise correlations, and nonstructured regions where any global relation can be found. In this paper, we present a combination of stochastic context-free grammars (SCFG) and bigram models. The SCFGs are used to represent the long-term relations of the structured part of RNA sequences, while the bigram models are used to capture the local relations of the nonstructured part. A stochastic version of Sakakibara's algorithm is used to study the SCFGs. Finally, experiments to evaluate the behavior of this proposal were carried out.

Current multiple classifier systems for unconstrained handwritten text recognition do not provide a straightforward way to utilize language model information. In this paper, we describe a generic method to integrate a statistical n-gram language model into the combination of multiple offline handwritten text line recognizers. The proposed method first builds a word transition network and then rescores this network with an n-gram language model. Experimental evaluation conducted on a large dataset of offline handwritten text lines shows that the proposed approach improves the recognition accuracy over a reference system as well as over the original combination method that does not include a language model.

Enterprise corpora contain evidence of what employees work on and therefore can be used to automatically find experts on a given topic. We present a general approach for representing the knowledge of a potential expert as a mixture of language models from associated documents. First we retrieve documents given the expert's name using a generative probabilistic technique and weight the retrieved documents according to expert-specific posterior distribution. Then we model the expert indirectly through the set of associated documents, which allows us to exploit their underlying structure and complex language features. Experiments show that our method has excellent performance on the expert search task of the TREC Enterprise track and that it effectively collects and combines evidence for expertise in a heterogeneous collection.

Deep parsing of Chinese sentences is a very challenging task due to their complexity such as ambiguous word boundaries and meanings. An alternative mode of Chinese language processing is to perform shallow parsing of Chinese sentences in which chunk segmentation plays an important role. In this paper, we present a chunk segmentation algorithm using a combined statistical and rule-based approach (CSRA). The decision rules for refining chunk segmentation are generated from incorrectly segmented chunks from a statistical model which is built on a training corpus. Experimental results show that the CSRA works well and produces satisfactory chunk segmentation results for subsequent processes such as chunk tagging and chunk collocation extraction.

Probabilistic language models have gained popularity in Natural Language Processing due to their ability to successfully capture language structures and constraints with computational efficiency. Probabilistic language models are flexible and easily adapted to language changes over time as well as to some new languages. Probabilistic language models can be trained and their accuracy strongly related to the availability of large text corpora.

In this paper, we investigate the usability of grapheme probabilistic models, specifically grapheme n-grams models in spellchecking as well as augmentative typing systems. Grapheme n-gram models require substantially smaller training corpora and that is one of the main drivers for this thesis in which we build grapheme n-gram language models for the Albanian language. There are presently no available Albanian language corpora to be used for probabilistic language modeling.

Our technique attempts to augment spellchecking and typing systems by utilizing grapheme n-gram language models in improving suggestion accuracy in spellchecking and augmentative typing systems. Our technique can be implemented in a standalone tool or incorporated in another tool to offer additional selection/scoring criteria.

In this paper, character-based Japanese dictation method is proposed. This method is based on the kana and kanji string language model proposed by Ito et al. First, sentences in the training corpus are split into character-based units (CBUs). Then strings of CBUs (CBUSes) are chosen from the CBU corpus based on a statistical criterion. We examined three criteria for the CBUS selection. They are the frequency-based selection, the mutual-information based selection and their combination. From the experimental results, it was found that the combined method gave the best result (7.19% and 8.75% CBU error rates for the 20k and the 60k word vocabulary conditions, respectively) which was better than the ordinary word-based method (7.61% and 9.15% CBU error rates for the 20k and the 60k word vocabulary conditions, respectively).

In addition, we carried out a recognition experiment for the Corpus of Spontaneous Japanese to confirm that the proposed method is effective for not only the read speech but also for spontaneous speech. As a result, we obtained the best result (29.82%) using the frequency-based method, which is better than the word-based recognition result (32.80%).

Inspired by machine translation task, recent text simplification approaches regard a task as a monolingual text-to-text generation, and neural machine translation models have significantly improved the performance of simplification tasks. Although such models require a large-scale parallel corpus, such corpora for text simplification are very few in number and smaller in size compared to machine translation task. Therefore, we have attempted to facilitate the training of simplification rewritings using pre-training from a large-scale monolingual corpus such as Wikipedia articles. In addition, we propose a translation language model to seamlessly conduct a fine-tuning of text simplification from the pre-training of the language model. The experimental results show that the translation language model substantially outperforms a state-of-the-art model under a low-resource setting. In addition, a pre-trained translation language model with only 3000 supervised examples can achieve a performance comparable to that of the state-of-the-art model using 30,000 supervised examples.

The established approach to unsupervised protein contact prediction estimates coevolving positions using undirected graphical models. This approach trains a Potts model on a Multiple Sequence Alignment. Increasingly large Transformers are being pretrained on unlabeled, unaligned protein sequence databases and showing competitive performance on protein contact prediction. We argue that attention is a principled model of protein interactions, grounded in real properties of protein family data. We introduce an energy-based attention layer, factored attention, which, in a certain limit, recovers a Potts model, and use it to contrast Potts and Transformers. We show that the Transformer leverages hierarchical signal in protein family databases not captured by single-layer models. This raises the exciting possibility for the development of powerful structured models of protein family databases.