Abstract

In this thesis, models are proposed for statistical machine translation and word-alignment which are well-suited for translation between English and Hindi (or any other Indian Language (IL)) apart from addressing the limitations of existing machine translation systems. In our models, we allow the flexibility of using linguistically motivated features but leave the learning to machine based on data. The purpose is better translation quality. Some of the issues in building a statistical translation system from English to Hindi are, 1. Large distance word-reordering. 2. Morphological richness of Hindi. 3. Lack of large training dataset. 4. Importance of function words in conveying grammatical roles in ILs. 5. Translation/Word-alignment of multi-word expressions. I address these issues in the statistical machine translation and word-alignment models proposed by us. The main contributions of my work are the two approaches that I propose for machine translation. They are, 1. Discriminative machine translation using global lexical selection (GLS approach): In this system, I propose a novel approach for machine translation using global lexical selection. Global lexical selection allows us to not rely on word-alignment information (which are often unreliable) between sentence pairs. Also, the emphasis in this approach is on lexical selection rather than lexical reordering. Indian languages are relatively free word order and hence, lexical selection (specially the function words) plays the most important role while translating to them. The results of this approach are presented in Chapters 4 and 5 [128, 129]. 2. Dependency based statistical machine translation system (Vaanee): I propose a dependency based statistical system that uses discriminative techniques to train its parameters. The use of syntax (dependency tree) allows us to address the large word-reorderings between English and Hindi. Discriminative training allows us to use rich feature sets, including linguistic features that are useful in the machine translation task. I present results of the experimental implementation of the system in this thesis (see Chapter 6) [134]. The model for discriminative machine translation and reranking using Global Lexical Selection (GLS) performed better than the state-of-art system (Moses). The main reasons for the success of this model are, 1. Not relying on word-alignment information. The dataset available for English-Hindi is small and noisy. This results in a low quality word-alignment between the sentences of this language pair. As the GLS based models do not depend on word-alignment, they performed very well. 2. The GLS based model relies on a global context in the source language to generate words in the target. This helps in an accurate generation of the target language words. In our approach, we used GLS to generate high quality local word groups, where the function words are associated with their corresponding content words. The high accuracy of predicting local word groups results in an accurate generation of function words which are the key grammatical role markers in Indian languages. This, in turn, results in target sentences that have high fluency. The high quality generation of lexical words in target also helps in reranking sentences generated by the state of art translation system. The results obtained are better than the state of art results. In the thesis, we present an experimental implementation of the discriminative machine translation model. This system provides the flexibility to incorporate rich linguistic information into a statistical machine translation system. It should act as an excellent platform for future experiments in syntax based statistical translation. It is available for download at the link: http://shakti.iiit.ac.in/sriram/vaanee.html. However, this model, which uses rich linguistic information doesn’t perform better than the state-of-art machine translation system mainly because of the availability of poor quality training data. Because of the nature of the dataset, the power of this kind of a model hasn’t been as effective as expected. In addition to the main contributions, I present two supporting contributions that are crucial building blocks of my work. The first work is on discriminative word-alignment. In this work, I show that the features representing compositionality of multi-word expressions, syntactic divergence and alignment structure help in the task of word-alignment and subsequently, in the task of machine translation (see chapter 7 [133, 131]). In the second supporting work, I propose various measures for computing the relative compositionality of multi-word expressions, of verb-object type in English and noun+verb type in Hindi (see Chapter 8 [130, 127, 85]). In the Introduction chapter, I build the background on the problem of machine translation and the various approaches

Abstract

In recent years, we have witnessed a tremendous growth in the volume of text documents available on the Internet, digital libraries, news sources, company-wide intranets and so on. This has led to an increased interest in developing methods that can help users to effectively navigate, summarize, and organize this information. The ultimate goal is to help users find what they are looking for. There are mainly two approaches to enhance the task of document organization: supervised approach, where pre-defined category labels are assigned to documents based on the likelihood suggested by a training set of labeled documents; and unsupervised approach, where there is no need for human intervention or labeled documents at any point in the whole process. Fast and high-quality document clustering algorithms play an important role towards the goal of document organization, as they have been shown to provide both an intuitive navigation/browsing mechanism by organizing large amounts of information into a small number of meaningful clusters. It can be noted that the performance of information retrieval method depends on feature selection and weight assignment methods which are being employed to extract the features from the documents. Normally, features from a document are selected based on some criteria (e.g. Frequency). These features are weighed by TF-IDF (Term frequency and inverse document frequency) method. There are research efforts to improve the performance of feature extraction methods by extending the concepts from the areas of ontology, open web directories and so on. The vector space model is one of the most widely used models to represent text document for computing similarity. Cosine similarity is a popular method to calculate similarity between two vectors. The similarity computation by cosine similarity method is influenced by common features (and their weight) between two document vectors. It can be observed that the cosine similarity is unable to find similarity based on the meaning or semantics. For example, consider two documents, one with a word “BMW” and the other having a word “Jaguar”. The similarity computation with cosine similarity between these two documents equals to zero. Even though these two documents have different vocabulary, they are semantically same as both refer car models. In the thesis, we made an effort to capture the context of the document to incorporate it into similarity computation. We exploit the generalization ability of hierarchical knowledge repositories such as “Open Web Directory”, in which, a given term relates to a context, and the context, in turn, relates to a collection of terms, then we can extract related terms for each term in the document. In a simple generalization hierarchy of web directory, a term at a higher level is a generalized concept for all the terms under this node e.g. sport is a generalized concept for football, cricket, baseball etc. We can add these generalize terms along with their weight to the document vector. By enriching tthe document with generalized contextual terms there is a scope to increase the number of common terms, thus the similarity between two documents. An improved approach is also proposed in which these generalized terms are later removed from the document vector leaving behind only the document terms. So, as output, we get a document vector with boosted weights and the dimension space of the document vector remains same. In addition to feature extraction and feature weighting of a document vector, we propose a method to improve clustering quality by assigning weight to the feature vector of a cluster. The contribution of this thesis is threefold: 1. We propose a framework that performs feature generation (using open web directory) and enriches the feature vector with new, more informative and discriminative features. We use topic paths of a term to obtain related contextual and generalized terms. 2. We propose an improved term weighting method called BoostWeight by considering the semantic association between the terms. BoostWeight method increases the weight of terms if they have common generalized term. 3. We propose a methodology to refine a given set of clusters by incrementally moving documents between clusters. More weight is given to the representative and discriminative features of a cluster. To compare the performance of proposed approaches with existing weighting approaches, we have conducted clustering experiments on two data-sets: WebData and Reuters21578 news corpus dataset. Experimental results show that the proposed approach improves clustering performance over other term extraction and weighting approaches.

Abstract

Multicarrier Code Division Multiple Access (MC-CDMA) is an attractive technique for high speed wireless data transmission in view of its advantages over orthogonal frequency division multiplexing. In this thesis, we analyze the performance of fully loaded downlink MC-CDMA systems with minimum mean square error (MMSE) equalizer in the presence of residual frequency offset (RFO) in multipath Rayleigh fading channels. We first show that as the SNR is increased beyond a value, referred as threshold SNR, the performance degrades. We then analyze the cause for this behavior and propose a remedy to prevent the degradation by regularizing the co- efficient(s) of the equalizer, and use the regularized equalizer for SNRs beyond the threshold value. The threshold SNR depends on the RFO and the profile of multipath channel. We suggest two methods for estimating this SNR, one gives close to the true value but requires the knowledge of RFO and the channel state information (CSI), while the other gives an approximate value but requires only CSI. We first show that if the actual value of RFO is less than the assumed, the threshold SNR estimate based on the assumed RFO will still be appropriate. Next, we show that the regularization based on the approximate value also prevents the degradation, but the performance at higher SNRs is slightly poorer compared to that with the better estimate. Numerical and simulation results are provided to support the analysis.

Abstract

Monitoring or surveillance by machines has become very popular nowadays because of the advantages that it provides over monitoring by human beings. Remote sensing has a huge potential to monitor the events on a large scale that millions of cameras put together on the surface of earth cannot monitor. Remote sensing data analysis involves examining the data using various viewing and interpretation devices to analyze the remotely sensed digital data. Reference data about the resources being studied are used when and where available to assist in the data analysis. Results of data analysis are then compiled generally in the form of hardcopy maps and tables or as computer files that can be merged with other ‘layers’ of information in a geographic information system (GIS). Finally, the information is presented to users who apply it to their decision making process. Image classification is one of the most widely used techniques on remote sensing data. The overall objective of image classification procedures is to automatically categorize all pixels in an image into land cover classes or themes. Land cover can be defined as the physical material at the surface of the earth. The range of projects, programs and organizations that use land cover maps to meet their planning, management, development and assessment objectives is expanding every day and the remote sensing community has been challenged to produce regional- to global-scale data sets on a repetitive basis that characterize ‘current’ Land Use Land Cover (LULC) patterns and also document all the LULC changes, however large or small they be. Spectral pattern recognition refers to the family of classification procedures that utilize the pixel-by-pixel spectral information as the basis for automated land cover classification. That is, different feature types manifest different combinations of digital numbers based on their inherent spectral reflectance and emittance properties. Spatial pattern recognition involves the categorization of image pixels on the basis of their spatial relationship with pixels surrounding them. Temporal pattern recognition uses time as an aid in feature identification. The multitemporal data that is now available because of better revisit times of the satellites allows us to characterize objects based on their dynamic processes rather than static properties like color, shape, etc and this is being successfully demonstrated in this research for a few classes. In case of multi-temporal studies, atmospheric influences and sensor malfunctioning leads to data gaps and data anomalies that have to be corrected for and is usually a preprocessing stage in all time series studies. This research has adopted a technique called local maximum fitting to smooth the noisy time series. Local maximum fitting combines temporal window operation and harmonic series fitting to provide a faithful representation of the ground process. We have viewed the classification of the pixel-time trajectories as a curve matching problem comparing the unknown time series of a pixel to the temporal signatures available in the library derived from the training samples obtained either through our proposed automatic method or any other method. An important part of curve matching problem is the definition of similarity measure. We have used various similarity measures and concluded that constrained dynamic time warping is best suited for classifying enhanced vegetation index (EVI) time series. Conventional classification methods inherently use Euclidean distance as a similarity measure even for time series classification. Through our empirical experiments, we have concluded that a Sakoe-chiba radius of 4 (equivalent to 2 months) for constraining DTW (Dynamic Time Warping) provides the best classification accuracy which is reasonable as there can be a maximum shift of this period in growing practices. The classification accuracy for Dharwad district was 86.57% with Euclidean distance as compared to 90.89% with constrained dynamic time warping distance measure for the moderate resolution imaging spectroradiometer (MODIS) dataset. The results obtained though the proposed classification procedure were also being compared with the conventional kNN and SVM classifiers. kNN classifer with k = 5 provided a classification accuracy of 79.75% and SVM classifier, an accuracy of 83.89%. The low classification accuracy of these classifiers for time series satellite image classification can be attributed to the use of Euclidean distance in the feature space. Whereas the actual supervised classification of satellite image is a highly automated process, assembling the training data needed for such classification is anything but automatic. In many ways, the training data needed for land cover classification is both an art and a science. It requires close interaction between the image analyst and the image data thus making the clas

Abstract

Software Defined Radio (SDR) is an emerging technology that is predicted to revolutionize the world of communications. An ideal SDR can be configured on the fly to operate over different wireless communication standards without any change in the hardware. SDR based systems require changes in various operating parameters like modulation techniques, wide-band or narrow-band operations, frequency of operation, output power level, security functions and several other baseband waveform requirements of different current and evolving standards. Thus, they suffer from some disadvantages, mainly due to redundant logic which is required to accommodate reconfigurability. Therefore, an optimal reconfigurable design is very complex. Even more difficult is to design or find an algorithm which in itself can support reconfigurability. This thesis deals with high performance algorithms and architectures of FFT for SDR. FFT is one of the most commonly used operations in any communication or multimedia system. It takes a large part of the baseband processor for operations varying from simple transform calculations to convolution, filtering, channelizer, OFDM and spectrum evaluation etc. These tasks represent some of the most computationally intensive operations of any communication system. In this thesis, an attempt has been made to find, analyze and design efficient FFT computation units to reduce area, computational delay and power consumption associated with reconfigurable and converged systems. A detailed analysis and comparison between different FFT algorithms and architectures is presented fitting to the requirement of the SDR based platforms. It is found that Bruun’s FFT is a better FFT algorithm, for most cases of practical importance in SDR. Also, a distributed precision based approach is suggested to reduce the arithmetic complexity and Noise-to-Signal ratio (NSR) contribution of Bruun’s FFT. Theoretical results have been validated by carrying out ASIC and FPGA implementations for the proposed designs. As predicted by the mathematical analysis, a significant reduction has been achieved in all the critical parameters like area, computational delay, reconfiguration overhead and power consumption for FFT designs as compared to popular FFT architectures currently used in various wireless communication standards. It is suggested that the proposed algorithm and designs can be used as a better alternative in SDR-based systems. As an application of the proposed theory, Bruun’s FFT has been applied in evaluating the comb spectrum for SDR based systems. Comb spectrum is an important and integral part of all multi-channel based communication systems like Wi-Max and WLAN. It is required in different standards varying from those based on OFDM and OFDMA (e.g. WLAN and WiMax) to Cyclic Prefix Code Division Multiple Access (CP-CDMA). It is shown that Bruun’s FFT is very useful in such a widely applicable, computationally intensive operation whose configuration requirement changes not only with the standard but also within a particular standard.

Abstract

User interface (UI) forms play a key role in information systems. The design of UI form is an important activity of an information system development. The improper UI increases navigational burden, increase data-entry errors, reduces the e±ciency of the user, and increases the cost of both data-entry and data maintenance. Normally, user expects an e®ective, easy-to- learn and easy-to-operate UI. Currently, context-based framework is well-known in designing UIs. A context covers a particular functionality or task for entering data into one or more relational tables of an information system. Developing an e±cient UI using context-based approach is a challenge, as the behavior of a system is known based on the contextual factors (such as the background knowledge of a user, the speci¯c goals and objectives of a user, the work context etc).

Abstract

User interface (UI) forms play a key role in information systems. The design of UI form is an important activity of an information system development. The improper UI increases navigational burden, increase data-entry errors, reduces the efficiency of the user, and increases the cost of both data-entry and data maintenance. Normally, user expects an effective, easy-to-learn and easy-to-operate UI. Currently, context-based framework is well-known in designing UIs. A context covers a particular functionality or task for entering data into one or more relational tables of an information system. Developing an efficient UI using context-based approach is a challenge, as the behavior of a system is known based on the contextual factors (such as the background knowledge of a user, the specific goals and objectives of a user, the work context etc). Research efforts are going on to investigate the design of efficient UIs under context-based framework. In context-based framework, the input attributes are divided into several contexts based on the functionality or task. The UI form is designed for each context. Normally, it is assumed that the system receives data values on all days for all the attributes in the context. It can be observed that, in some information systems, some attributes receive values only during certain time duration. Therefore, the attribute may be not required in the UI for the remaining duration. As a result, there is an opportunity to reduce UI size. We call such property of the attribute as “temporality” of attribute. So, for several information systems, there is a scope to exploit the temporality of attributes for the efficient UI design. As a part of this work, we have proposed improved UI design approaches both for desktop and small screens. * To develop an efficient UI design for desktop environment, the notion of attribute temporality is used to divide the large UI into several small UIs in desktop or PC-based systems. The temporality values can be fixed by analyzing the system functionality or task. By taking attributes and temporality values of a given context, we have proposed an approach to cluster the attributes into several active-contexts where each active-context can be considered as an independent UI for certain time period. We have designed UIs for a real system eSagu® (eSagu® is an IT-based personalized agro-advisory system) with this approach. The implementation results show that it is possible to improve the navigational burden significantly with the proposed approach as compared to the existing approaches. * We proposed an approach for small screens by exploiting attribute temporality to interface tailoring approach. In small screen environments, such as mobile phones and personal digital assistants, interface tailoring approaches are employed to develop UI forms. In these approaches, the PC-based UI is divided into multiple screens (or high-level components) in such a way that the UI can be accommodated in one or few mobile screens. In the proposed approach, the notion of attribute “temporality” is extended to divide a given context into several small active-contexts. However, we observed that certain active-contexts have more attributes which cannot fit in small screen. In such situations, we explored interface tailoring approaches in building suitable interfaces for mobile devices. We have carried out analysis by considering real world dataset of eSagu® system and found out that the proposed approach has a potential to improve the performance.

Abstract

Illiteracy is an impediment to India’s development as approximately only 66% of the population is literate, with about 150 million adult illiterates. Tata Consultancy Services (TCS), an Indian Software House has initiated Adult Literacy Programme (ALP) as a solution to adult illiteracy problem in India, developed ALP products under this initiative and has made over 1,20,000 people literate across India. An ALP product is an eLearning system that employs Computer Based Functional Literacy (CBFL) as the underlying technology to teach adult illiterates of a particular language. However, the goal of ALP initiative is to develop and maintain ALP products for 22 Indian languages and their major dialects. The scope also includes development of ALP products to help migrants from one linguistic region to another. It is labor intensive to develop and maintain a family of ALP products because of the enormous number and huge variety of products. The ALP Problem is concerned with the automating the development of a family of ALP products. In this thesis, we present ALP Factory as a solution to the ALP Problem, which has resulted significant improvements in productivity levels of up to 89%. We have also developed Prerak Training Factory that pertains to automating the development of training software for the ALP family of products. We found that development of a family of eLearning Systems is a common problem in developing computer-based systems for computer literacy, technical literacy, vocational training, school education and thus presenting us with the problem of automating software development. Despite many decades of intensive research towards software industrialization and software product lines to improve software productivity, the automation problem is a daunting task. In this thesis, we propose TALES as an approach towards automating the development of a family of eLearning Systems. This approach was borne out of our experience of providing a solution to the ALP problem and inspired from traditional manufacturing. This approach progresses by first gaining a deeper understanding of the problem domain, designing a solution based on this understanding to handle the scale and variety of software products. Then various kinds of standardization focusing on product structure, production processes, and raw material are developed to provide a standard platform that enables assembly and interchange of parts. The next step involves the development of Virtual Assembly Lines that facilitate the automation of production processes by weaving processes with tools and hence producing a family of similar but distinct products. This approach is supported by repositories of components, tools, and processes. The uniqueness of TALES approach is its ability to facilitate non-technical users during software development.

Abstract

An accurate and up-to-date road network database is essential for GIS (Geographic Information System) based applications such as urban and rural planning, transportation management, vehicle navigation, emergency response, etc. Often, road database is generated through field surveys with the help of GPS (Global Positioning System) enabled instruments. This approach of road extraction, however, is time consuming and labour intensive. With increase in availability of satellite imagery both in high and low resolutions, automatic road network extraction from satellite imagery has received considerable attention and has been studied extensively since 1970s. Since road region appear as linear segment in low resolution images, earlier research on road extraction focused on extracting road center-line from low resolution images. On the other hand, high resolution satellite images provide an opportunity to extract entire road area, which are particularly useful for vehicle navigation, along with road network. However, factors like occlusion due to trees along the road, building shadows and vehicles, variation in road surface characteristics (example, bitumen road versus concrete road) and changing road geometry, which become more apparent in high resolution satellite images, make road extraction a challenging problem. Though several road extraction techniques have been proposed in the literature, because of various assumptions about the road region characteristics and imaging modality, such techniques often exhibit limited success in the real scenarios. Of these techniques, the region growing based road extraction approach, in which road extraction starts with road seed points (either provided manually or generated automatically) and extracts entire road region based on predefined matching criteria, holds considerable promise. In this thesis we present a novel automatic road extraction algorithm based on adaptive texture matching (ATM-R), which is a variant of region growing approach. The algorithm developed in this thesis is robust to variations in radiometric resolution of input images, road surface characteristics (texture) and road geometry and is scalable to inputs from different sensors. In the proposed algorithm, difference in a pair of closely dated multi-temporal images of same geographic area has been studied to generate road templates (seeds) and such templates are further utilized to extract remaining road region within a road template matching framework which adapts to local road texture. To assess the performance of the proposed algorithm, a set of images encompassing wide range of radiometric resolutions and different spatial resolutions were prepared. The image set consisted of grayscale equivalent of pan-sharpened IKONOS images, panchromatic CARTOSAT-2 images and grayscale equivalent of images captured from Google Earth. Images were chosen to include road regionswith varying width, wide range of road surface reflectance values and different textures. Experiments were conducted on images with disconnected road segments, multi-lane roads, road regions occluded with trees and images with road over-bridges. The measures used to evaluate the algorithm performance are road network completeness, road area completeness and correctness of road area. The algorithm is able to consistently extract 70% to 90% of the road network and has a high performance against all the three measures. Along with generating road network the proposed algorithm is able to extract road area which can be further used to assess road width, number of road lanes and other auxiliary road network information. The algorithm has considerably reduced manual intervention in road extraction process and avoided the need for GCPs (ground control points).

Abstract

Nanoparticles are the particles having dimensions in the range of 1- 100 nanometres. In recent years the synthesis and characterization of nanoparticles have gained a lot of attention due to their potential application in drug delivery system and as biosensors along with their unique optical, chemical, electronic, and mechanical properties which is missing from their bulk counterparts. The nanoparticles can be surface engineered to achieve specific goal. The ability to exploit and modify the properties of nanoparticles, makes it a potential agent for the enhancement of the performances of the enzyme-based biosensors. Iron nanoparticles are magnetic nanoparticles which make it easy to be detected and potential transducer. Biosensors are analytical devices that perform qualitative and/or quantitative estimation of analytes. An essential requirement is a biomaterial that selectively recognizes the molecule of interest. Applications of nanomaterials to biosensors have recently aroused much interest. This is because these interesting materials exhibit large surface-to-volume ratio, high surface reaction activity, high catalytic efficiency, and strong adsorption ability that are helpful for the immobilization of desired biosensing molecules. Moreover, nanoparticles have a unique ability to promote fast electron transfer between electrode and the active site of the enzyme. In this work, magnetite (Fe3O4) nanoparticles have been modified using glutathione and then proteins (BSA and HRP) have been coupled to study the characteristics of the nanoparticles for possible biosensor application. It has been found that the immobilized HRP displays excellent catalytic property to H2O2, and Fe3O4 nanoparticles in the biosensing interface offer not only friendly environment to immobilize HRP but also improve the electron transfer between analyte and glutathione modified Fe3O4 electrode surface. HRP/GT+Fe3O4 bioelectrode shows high sensitivity, low detection limit and fast response for analytes. In addition, the HRP/GT+Fe3O4 based biosensor exhibits both good reproducibility, long-term stability and cost effectiveness. It should be considered as an approach to develop biosensors for immunological and pathological analytes.