Phrase-based statistical language generation using graphical models and active learning

Most previous work on trainable language generation has focused on two paradigms: (a) using a statistical model to rank a set of generated utterances, or (b) using statistics to inform the generation decision process. Both approaches rely on the existence of a handcrafted generator, which limits their scalability to new domains. This paper presents BAGEL, a statistical language generator which uses dynamic Bayesian networks to learn from semantically-aligned data produced by 42 untrained annotators. A human evaluation shows that BAGEL can generate natural and informative utterances from unseen inputs in the information presentation domain. Additionally, generation performance on sparse datasets is improved significantly by using certainty-based active learning, yielding ratings close to the human gold standard with a fraction of the data. © 2010 Association for Computational Linguistics.

The effect of language models on phonetic decoding for spoken term detection

Spoken term detection (STD) popularly involves performing word or sub-word level speech recognition and indexing the result. This work challenges the assumption that improved speech recognition accuracy implies better indexing for STD. Using an index derived from phone lattices, this paper examines the effect of language model selection on the relationship between phone recognition accuracy and STD accuracy. Results suggest that language models usually improve phone recognition accuracy but their inclusion does not always translate to improved STD accuracy. The findings suggest that using phone recognition accuracy to measure the quality of an STD index can be problematic, and highlight the need for an alternative that is more closely aligned with the goals of the specific detection task.

Query expansion using term relationships in language models for information retrieval

Language Modeling (LM) has been successfully applied to Information Retrieval (IR). However, most of the existing LM approaches only rely on term occurrences in documents, queries and document collections. In traditional unigram based models, terms (or words) are usually considered to be independent. In some recent studies, dependence models have been proposed to incorporate term relationships into LM, so that links can be created between words in the same sentence, and term relationships (e.g. synonymy) can be used to expand the document model. In this study, we further extend this family of dependence models in the following two ways: (1) Term relationships are used to expand query model instead of document model, so that query expansion process can be naturally implemented; (2) We exploit more sophisticated inferential relationships extracted with Information Flow (IF). Information flow relationships are not simply pairwise term relationships as those used in previous studies, but are between a set of terms and another term. They allow for context-dependent query expansion. Our experiments conducted on TREC collections show that we can obtain large and significant improvements with our approach. This study shows that LM is an appropriate framework to implement effective query expansion.

Sampling stored-product insect pests : a comparison of four statistical sampling models for probability of pest detection

Background: Developing sampling strategies to target biological pests such as insects in stored grain is inherently difficult owing to species biology and behavioural characteristics. The design of robust sampling programmes should be based on an underlying statistical distribution that is sufficiently flexible to capture variations in the spatial distribution of the target species. Results: Comparisons are made of the accuracy of four probability-of-detection sampling models - the negative binomial model,1 the Poisson model,1 the double logarithmic model2 and the compound model3 - for detection of insects over a broad range of insect densities. Although the double log and negative binomial models performed well under specific conditions, it is shown that, of the four models examined, the compound model performed the best over a broad range of insect spatial distributions and densities. In particular, this model predicted well the number of samples required when insect density was high and clumped within experimental storages. Conclusions: This paper reinforces the need for effective sampling programs designed to detect insects over a broad range of spatial distributions. The compound model is robust over a broad range of insect densities and leads to substantial improvement in detection probabilities within highly variable systems such as grain storage.

Language models for online handwritten Tamil word recognition

N-gram language models and lexicon-based word-recognition are popular methods in the literature to improve recognition accuracies of online and offline handwritten data. However, there are very few works that deal with application of these techniques on online Tamil handwritten data. In this paper, we explore methods of developing symbol-level language models and a lexicon from a large Tamil text corpus and their application to improving symbol and word recognition accuracies. On a test database of around 2000 words, we find that bigram language models improve symbol (3%) and word recognition (8%) accuracies and while lexicon methods offer much greater improvements (30%) in terms of word recognition, there is a large dependency on choosing the right lexicon. For comparison to lexicon and language model based methods, we have also explored re-evaluation techniques which involve the use of expert classifiers to improve symbol and word recognition accuracies.

Development and Evaluation of Statistical Downscaling Models for Monthly Precipitation

Several statistical downscaling models have been developed in the past couple of decades to assess the hydrologic impacts of climate change by projecting the station-scale hydrological variables from large-scale atmospheric variables simulated by general circulation models (GCMs). This paper presents and compares different statistical downscaling models that use multiple linear regression (MLR), positive coefficient regression (PCR), stepwise regression (SR), and support vector machine (SVM) techniques for estimating monthly rainfall amounts in the state of Florida. Mean sea level pressure, air temperature, geopotential height, specific humidity, U wind, and V wind are used as the explanatory variables/predictors in the downscaling models. Data for these variables are obtained from the National Centers for Environmental Prediction-National Center for Atmospheric Research (NCEP-NCAR) reanalysis dataset and the Canadian Centre for Climate Modelling and Analysis (CCCma) Coupled Global Climate Model, version 3 (CGCM3) GCM simulations. The principal component analysis (PCA) and fuzzy c-means clustering method (FCM) are used as part of downscaling model to reduce the dimensionality of the dataset and identify the clusters in the data, respectively. Evaluation of the performances of the models using different error and statistical measures indicates that the SVM-based model performed better than all the other models in reproducing most monthly rainfall statistics at 18 sites. Output from the third-generation CGCM3 GCM for the A1B scenario was used for future projections. For the projection period 2001-10, MLR was used to relate variables at the GCM and NCEP grid scales. Use of MLR in linking the predictor variables at the GCM and NCEP grid scales yielded better reproduction of monthly rainfall statistics at most of the stations (12 out of 18) compared to those by spatial interpolation technique used in earlier studies.

Dasher - A data entry interface using continuous gestures and language models

Existing devices for communicating information to computers are bulky, slow to use, or unreliable. Dasher is a new interface incorporating language modelling and driven by continuous two-dimensional gestures, e.g. a mouse, touchscreen, or eye-tracker. Tests have shown that this device can be used to enter text at a rate of up to 34 words per minute, compared with typical ten-finger keyboard typing of 40-60 words per minute. Although the interface is slower than a conventional keyboard, it is small and simple, and could be used on personal data assistants and by motion-impaired computer users.

Syllable language models for Mandarin speech recognition: exploiting character language models.

Mandarin Chinese is based on characters which are syllabic in nature and morphological in meaning. All spoken languages have syllabiotactic rules which govern the construction of syllables and their allowed sequences. These constraints are not as restrictive as those learned from word sequences, but they can provide additional useful linguistic information. Hence, it is possible to improve speech recognition performance by appropriately combining these two types of constraints. For the Chinese language considered in this paper, character level language models (LMs) can be used as a first level approximation to allowed syllable sequences. To test this idea, word and character level n-gram LMs were trained on 2.8 billion words (equivalent to 4.3 billion characters) of texts from a wide collection of text sources. Both hypothesis and model based combination techniques were investigated to combine word and character level LMs. Significant character error rate reductions up to 7.3% relative were obtained on a state-of-the-art Mandarin Chinese broadcast audio recognition task using an adapted history dependent multi-level LM that performs a log-linearly combination of character and word level LMs. This supports the hypothesis that character or syllable sequence models are useful for improving Mandarin speech recognition performance.