Search

In this paper a new method for automatic word clustering is presented. We used this method for building n-gram language models for Persian continuous speech recognition (CSR) systems. In this method, each word is specified by a feature vector that represents the statistics of parts of speech (POS) of that word. The feature vectors are clustered by k-means algorithm. Using this method causes a reduction in time complexity which is a defect in other automatic clustering methods. Also, the problem of high perplexity in manual clustering methods is abated. The experimental results are based on "Persian Text Corpus" which contains about 9 million words. The extracted language models are evaluated by the perplexity criterion and the results show that a considerable reduction in perplexity has been achieved. Also reduction in word error rate of CSR system is about 16% compared with a manual clustering method. [ABSTRACT FROM AUTHOR]

We report on a series of experiments addressing the fact that German is less suited than English for word-based n-gram language models. Several systems were trained at different vocabulary sizes using various sets of lexical units. They were evaluated against a newly created corpus of German and Austrian broadcast news. [ABSTRACT FROM AUTHOR]

Bilingual documentation has become a common phenomenon in many official institutions and private companies. In this scenario, the categorization of bilingual text is a useful tool, that can be also applied in the machine translation field. To tackle this classification task, different approaches will be proposed. On the one hand, two finite-state transducer algorithms from the grammatical inference domain will be discussed. On the other hand, the well-known naive Bayes approximation will be presented along with a possible modelization based on n-gram language models. Experiments carried out on a bilingual corpus have demonstrated the adequacy of these methods and the relevance of a second information source in text classification, as supported by classification error rates. Relative reduction of 29% with respect to the best previous results on the monolingual version of the same task has been obtained. [ABSTRACT FROM AUTHOR]

Most of state-of-the-art large vocabulary continuous speech recognition systems use word-based n-gram language models. Such models are not optimal solution for inflectional or agglutinative languages. The Polish language is highly inflectional one and requires a very large corpora to create a sufficient language model with the small out-of-vocabulary ratio. We propose a syllable-based language model, which is better suited to highly inflectional language like Polish. In case of lack of resources (i.e. small corpora) syllable-based model outperforms word-based models in terms of number of out-of-vocabulary units (syllables in our model). Such model is an approximation of the morpheme-based model for Polish. In our paper, we show results of evaluation of syllable based model and its usefulness in speech recognition tasks. [ABSTRACT FROM AUTHOR]

In Natural Language Generation (NLG) systems, a general-purpose surface realisation module will usually require the underlying application to provide highly detailed input knowledge about the target sentence. As an attempt to reduce some of this complexity, in this paper we follow a traditional approach to NLG and present a number of experiments involving the use of n-gram language models as an aid to an otherwise rule-based text generation approach. By freeing the application from the burden of providing a linguistically- rich input specification, and also by taking some of the generation decisions away from the surface realisation module, we expect to make NLG techniques accessible to a wider range of potential applications. [ABSTRACT FROM AUTHOR]

In this article, we propose the use of suffix arrays to efficiently implement n-gram language models with practically unlimited size n. This approach, which is used with synchronous back-off, allows us to distinguish between alternative sequences using large contexts. We also show that we can build this kind of models with additional information for each symbol, such as part-of-speech tags and dependency information. The approach can also be viewed as a collection of virtual k-testable automata. Once built, we can directly access the results of any k-testable automaton generated from the input training data. Synchronous back-off automatically identifies the k-testable automaton with the largest feasible k. We have used this approach in several classification tasks. [ABSTRACT FROM AUTHOR]

This paper presents two different approaches to automatic captioning of geo-tagged images by summarizing multiple web-documents that contain information related to an image΄s location: a graph-based and a statistical-based approach. The graph-based method uses text cohesion techniques to identify information relevant to a location. The statistical-based technique relies on different word or noun phrases frequency counting for identifying pieces of information relevant to a location. Our results show that summaries generated using these two approaches lead indeed to higher ROUGE scores than n-gram language models reported in previous work. [ABSTRACT FROM AUTHOR]

Statistical Machine Translation (SMT) is an evolving field where many techniques in Syntactic Pattern Recognition (SPR) are needed and applied. A typical phrase-based SMT system for translating from a T (target) language to an S (source) language contains one or more n-gram language models (LMs) and one or more phrase translation models (TMs). These LMs and TMs have a large memory footprint (up to several gigabytes). This paper describes novel techniques for filtering these models that ensure only relevant patterns in the LMs and TMs are loaded during translation. In experiments on a large Chinese-English task, these techniques yielded significant reductions in the amount of information loaded during translation: up to 58% reduction for LMs, and up to 75% for TMs. [ABSTRACT FROM AUTHOR]

For n-gram language models the vital configuration parameters are given by the size of the lexicon used and the length of the context to be considered. When creating HMMs for a certain application, in contrast, it is not immediately clear, what model size should be chosen, which type of emission modeling should be used, and whether the number of possible paths through the model could eventually be restricted in a suitable way. [ABSTRACT FROM AUTHOR]

Statistical pattern recognition has proved to be an interesting framework for machine translation, and stochastic finite-state transducers are adequate models in many language processing areas such as speech translation, computer-assisted translations, etc. The well-known n-gram language models are widely used in this framework for machine translation. One of the application of these n-gram models is to infer stochastic finite-state transducers. However, only simple dependencies can be modelled, but many translations require to take into account strong context and style dependencies. Mixtures of parametric models allow to increase the description power of the statistical models by modelling subclasses of objects. In this work, we propose the use of n-gram mixtures in GIATI, a procedure to infer stochastic finite-state transducers. N-gram mixtures are expected to model topics or writing styles. We present experimental results showing that translation performance can be improved if enough training data is available. [ABSTRACT FROM AUTHOR]

One area of pattern recognition that is receiving a lot of attention recently is handwritten text recognition. Traditionally, handwritten text recognition systems have been modelled by means of HMM models and n-gram language models. The problem that n-grams present is that they are not able to capture long-term constraints of the sentences. Stochastic context-free grammars (SCFG) can be used to overcome this limitation by rescoring a n-best list generated with the HMM-based recognizer. Howerver, SCFG are known to have problems in the estimation of comlpex real tasks. In this work we propose the use of a combination of n-grams and category-based SCFG together with a word distribution into categories. The category-based approach is thought to simplify the SCFG inference process, while at the same time preserving the description power of the model. The results on the IAM-Database show that this combined scheme outperforms the classical scheme. [ABSTRACT FROM AUTHOR]

This paper proposes a method of combining two n-gram language models to construct a single language model. One of the corpora is constructed from a very small corpus of the right domain of interest, and the other is constructed from a large but less adequate corpus. This method is based on the observation that a small corpus from the right domain has high quality n-grams but suffers from sparseness problem, while a large corpus from another domain is inadequately biased, but easy to obtain bigger size. The basic idea behind dual-source backoff is basically the same with Katz's backoff. We ran experiments with 3-gram language models constructed from newspaper corpora of several millions to tens of millions words together with models from smaller broadcast news corpora. The target domain was broadcast news. We obtained significant improvement by incorporating a small corpus around one thirtieth size of the newspaper corpus. [ABSTRACT FROM AUTHOR]