Contextual advertising on web pages has become very popular recently and it poses its own set of unique text mining challenges. Often advertisers wish to either target (or avoid) some specific content on web pages which may appear only in a small part of the page. Learning for these targeting tasks is difficult since most training pages are multi-topic and need expensive human labeling at the sub-document level for accurate training. In this paper we investigate ways to learn for sub-document classification when only page level labels are available - these labels only indicate if the relevant content exists in the given page or not. We propose the application of multiple-instance learning to this task to improve the effectiveness of traditional methods. We apply sub-document classification to two different problems in contextual advertising. One is "sensitive content detection" where the advertiser wants to avoid content relating to war, violence, pornography, etc. even if they occur only in a small part of a page. The second problem involves opinion mining from review sites - the advertiser wants to detect and avoid negative opinion about their product when positive, negative and neutral sentiments co-exist on a page. In both these scenarios we present experimental results to show that our proposed system is able to get good block level labeling for free and improve the performance of traditional learning methods.

Speaker verification can be treated as a statistical hypothesis testing problem. The most commonly used approach is the likelihood ratio test (LRT), which can be shown to be optimal using the Neymann-Pearson lemma. However, in most practical situations the Neymann-Pearson lemma does not apply. In this paper, we present a more robust approach that makes use of a hybrid generative-discriminative framework for text-dependent speaker verification. Our algorithm makes use of a generative models to learn the characteristics of a speaker and then discriminative models to discriminate between a speaker and an impostor. One of the advantages of the proposed algorithm is that it does not require us to retrain the generative model. The proposed model, on an average, yields 36.41% relative improvement in EER over a LRT.

Information workers are often involved in multiple tasks and activities that they must perform in parallel or in rapid succession. In consequence, task management itself becomes yet another task that information workers need to perform in order to get the rest of their work done. Recognition of this problem has led to research on task management systems, which can help by allowing fast task switching, fast task resumption, and automatic task identification. In this paper we focus on the latter: we tackle the problem of automatically detecting the tasks that the user is involved in, by identifying which of the windows on the user's desktop are related to each other. The underlying assumption is that windows that belong to the same task share some common properties with one another that we can detect from data. We will refer to this problem as the task assignment problem.To address this problem, we have built a prototype named Swish that: (1) constantly monitors users' desktop activities using a stream of windows events; (2) logs and processes this raw event stream, and (3) implements two criteria of window "relatedness", namely the semantic similarity of their titles, and the temporal closeness in their access patterns.In addition to describing the Swish prototype in detail, we validate it with 4 hours of user data, obtaining task classification accuracies of about 70%. We also discuss our plans on including Swish in a number of intelligent user interfaces and future lines of research.

We introduce an elegant and novel design for a speech detector which estimates the probability of the presence of speech in each time-frequency bin, as well as in each frame. The proposed system uses discriminative estimators based on logistic regression, and incorporates spectral and temporal correlations in the same framework. The detector is flexible enough to be configured in a single level or a "stacked" bilevel architecture depending on the needs of the application. An important part of the proposed design is the use of a new set of features: the normalized logarithm of the estimated posterior signal-to-noise ratio. These can be easily and automatically generated by tracking the noise spectrum online. We present results on the AURORA database to demonstrate that the overall design is simple, flexible and effective.

In this paper, we introduce a new framework for speech detection using convolutional networks. We propose a network architecture that can incorporate long and short-term temporal and spectral correlations of speech in the detection process. The proposed design is able to address many shortcomings of existing speech detectors in a unified new framework: First, it improves the robustness of the system to environmental variability while still being fast to evaluate. Second, it allows for a framework that is extendable to work under different time-scales for different applications. Finally, it is discriminative and produces reliable estimates of the probability of presence of speech in each frame for a wide variety of noise conditions. We propose that the inputs to the system be features that are measures of the true signal-to-noise ratio of a set of frequency bands of the signal. These can be easily and automatically generated by tracking the noise spectrum online. We present preliminary results on the AURORA database to demonstrate the effectiveness of the detector over conventional Gaussian detectors. 1.

This paper describes an attempt to design a knowledge-based large vocabulary speech recognition system. Our motivation is to replace features based on the short-term spectra, such as Mel-frequency cepstral coefficients (MFCC), by features that explicitly represent some of the distinctive features of the speech signal. However, rather than attempting to compute acoustic correlates of these distinctive features, we have engineered an approach where neural networks are trained to map short-term spectral features to the posterior probability of some distinctive features. These probabilities are then used as features in a large vocabulary tied-state HMM-based recognizer. Experimental results on the Wall Street Journal Task show that such a system, while not outperforming a MFCC-based system, generates very different error patterns. After combining the results of a base-line MFCC system with the results of several systems based on the proposed approach, we were able to obtain reductions in word error rates of 19% and 10 % on the 5K and 20K tasks respectively over our best MFCC-based systems.

In most adaptation schemes, when a limited amount of data is available, the estimation errors affect the feature coefficients which have a small dynamic range i.e. delta and delta-delta coefficients. On the other hand, adapting only the cepstral coefficients compromises the performance when large amounts of data are available. In this paper we present a novel approach to solving this problem based on hierarchical Bayesian adaptation. First the delta parameters are cast as transformations of the cepstra. Then the cepstra are adapted using Bayesian techniques. The parameters of posterior distribution of the cepstra are transformed to get the hyperparameters of the delta coefficients which are then adapted. In this paper, the solution is presented in the framework of Bayesian predictive adaptation, but the approach is general, hence it can be applied in any other adaptation framework. Preliminary results are presented that demonstrate the effectiveness of the new method.

Modelling non-native speakers with different mother tongues is a difficult task for automatic speech recognition due to the large variation among speakers. One possibility for jointly modelling all speakers is to use the same speaker independent acoustic mod-els and a joint lexicon to capture the variation. We have modified the reference lexicon using pronunciation rules that are derived in a totally data-driven manner from a set of adaptation data using the reference recognizer and the reference lexicon. Deriving common rules for such diverse sources simul-taneously is difficult. The challenge is to combine these rules to a common set without increasing the confusability. In this paper we compare several methods of combining the individual rules to form a common lexicon for all speakers. Us-ing a new log likelihood rule pruning measure presented in this paper, we achieved improved performance compared with more traditional rule pruning methods based on rule probability, and with much fewer rules. With a confusability reduction scheme we reduced the number of rules even further. 1.

Most state-of-the-art speaker verification systems need a user model built from samples of the customer speech, and a speaker independent (SI) background model with high acoustic resolution. These systems rely heavily on the availability of speaker independent databases along with a priori knowledge about acoustic rules of the utterance, and depend on the consistency of acoustic conditions under which the SI models were trained. These constraints may be a burden in practical and portable devices such as palm-top computers or wireless handsets which place a premium on computation and memory, and where the user is free to choose any password utterance in any language, under any acoustic condition. In this paper, we present a novel and reliable approach to background model design when only the enrollment data is available. Preliminary results are provided to demonstrate the effectiveness of such systems.

The performance of an automatic speech recognizer degrades when there exists an acoustic mismatch between the training and the testing conditions in the data. Though it is certain that the mismatch is nonlinear, its exact form is unknown. Tackling the problem of nonlinear mismatches is a difficult task that has not been adequately addressed before. We develop an approach that uses nonlinear transformations in the stochastic matching framework to compensate for acoustic mismatches. The functional form of the nonlinear transformation is modeled by neural networks. We develop a new technique to train neural networks using the generalized EM algorithm. This technique eliminates the need for stereo databases, which are difficult to obtain in practical applications. The new technique is data-driven and hence can be used under a wide variety of conditions without a priori knowledge of the environment. Using this technique, we show that we can provide improvement under various types of acoustic mismatch; in some cases a 72% reduction in word error rate is achieved.