Abstract

Designing embedded systems efficiently has always been of significant interest. This is primarily motivated by the fact that these embedded systems find application in a wide range of fields: industrial automation, commercial electronics, scientific and medical applications, and many more. With the inherent drive towards automation in today’s world, these applications are only ever-increasing. With the tremendous rise in the complexity of such systems and the need to satisfy multiple and mutually conflicting constraints, the design optimization problem has become extremely large and complex. Hence, there is an urgent necessity to develop new and efficient techniques for design of embedded systems. In this thesis, we apply and refine the design strategy of Hardware Software Partitioning (HW/SW Partitioning) to embedded systems. In the first part of the thesis, HW/SW Partitioning is introduced and then, a co-design methodology is utilized to enhance the widely used Fast Fourier Transform algorithm. A high performance and low-cost implementation is proposed which exploits the principles of Partitioning on a target system consisting of a Spartan 3E FPGA and an Atmega-16 micro-controller. A comparative analysis is then demonstrated, with respect to both, the hardware-only and software-only implementations, which depict the superior performance of the proposed implementation. It is then observed that as the problem size, i.e. complexity of target application/algorithm increased, making Partitioning decisions manually was not feasible. Hence, in order to utilize the true potential of HW/SW Partitioning (especially for large-scale applications), automation of the partitioning process was required. This formed the next focus of this thesis, which comprised of developing high speed and resource efficient algorithms for the problem of HW/SW Partitioning. The problem has been modeled as a multi-dimensional optimization problem with the aim of minimizing the area utilization, power dissipation, time of execution and system memory requirement of the implementation. The novelty of this work is in developing a fast paced, state-of-the-art algorithm which takes into consideration all the major design costs as mentioned above. The Greedy Metaheuristic Algorithm (GMA), with O(n) time complexity has been proposed, which significantly reduces the complexity of algorithms previously described in the literature. The proposed approach can also be efficiently implemented on resource constraint devices; such as low cost System On Chips (SoCs). The problem model being considered thus far was lacking one major factor, the communication costs incurred by the different hardware and software processing engines (PEs). With the contemporary age of Multi-Processor System on Chips (MPSoCs) and hybrid multi-core systems, this became a very important cost which could not be neglected. Thus, in this work, an extension to the previous model was introduced next. Subsequently, a two-phased algorithm, called Phased Greedy Metaheuristic Algorithm (PGMA) was proposed, which solved the problem of Partitioning, considering all the relevant communication costs. In order to evaluate the efficiency and performance of the proposed algorithms, a detailed empirical study was also performed. A comparison of the execution time and solution quality of the proposed algorithms with other major partitioning algorithms in the literature was depicted. The time of execution was found to be as low as 0.011 seconds, with deviation from optimal solution by less than 2% (solution error of under 2%), for partitioning an algorithm consisting of 2500 blocks. To the best of our knowledge, this is the fastest performance till date. Due to the heuristic nature of the algorithms, it was necessary to ascertain the solution quality (how optimal were the generated solutions) and compare them with other approaches. This analysis also confirmed the usefulness of the proposed algorithms. After an intensive theoretical design and performance evaluation, it became imminent to apply the proposed algorithms to real life embedded systems and determine their effectiveness. A comprehensive literature survey was conducted to find a suitable problem of relevance to HW/SW Partitioning. With the growing use and reach of digital photography, embedded systems have been increasingly used in the area of digital signal processing. Joint-Photographic Expert-Group (JPEG) Encoder was one such contemporary embedded system that is widely used in numerous applications. This has resulted in a demand for improved and more efficient constructions of the system. Any research carried out to improve the construction of JPEG encoder will have a significant impact in enhancing contemporary embedded system research. Thus, for this thesis, we chose to propose an efficient partitioning scheme for the JPEG encoder, using the proposed algorithm PGMA. A real-life case study was performed. Th

Abstract

In today’s era social media is our window into creating contextual systems of advanced technology. These systems span a breadth of industries and domains. Whether it is contextual phone assistants like Siri and Crotana or search based advertisements, all of them rely deeply on user data to teach themselves human-like responsiveness. A major chunk of the data available at our disposal is lexical data (primarily from social media platforms). Thus, psychology infered via lexical markers online is an indispensable part of the future of our technology. Lexical psychology and personality are paradigms that have been given immense attention for over a decade. Most state of the art techniques use pre-existing features from social media functions (for instance Number of followers , Likes on Facebook etc. to predict various personality models (such as Big Five , MBTI etc.) that have been around for over 70 years. Parts of our work attempt to tackle this approach of using pre-existing features (available for direct reference from social media). In doing so, we traced the conception of the Big Five model of personality to be concieved by means of extensive lexical analysis (both empirical and theoretical). Based on this knowledge, we illustrate the importance of additional features (informed by the fundamentals of psychology) in training robust personality pre- diction models. Training one such model ourselves, the results we achieved were a clear indicator of the effectiveness of lexical features versus stand-alone technological features to predict user personality. We then move on to tackle the drawbacks of the continued usage of the Big Five model to understand user personality since the last 70 years. At the time this model was proposed by Goldberg et al. the only mode for obtaining relevant user data was by means of psychometric tests. While the databases available to us have advanced significantly, the models used to capture lexical personality from these data reserves remain constant. They make available to us an overall persona of an individual. Contrary to these, abundant literature in the domain of psychology suggests that the persona of an individual is ever evolving. It is dynamic and changes temporally with the individual’s cognition, action and experiences. As opposed to this our technology today is still dependent on such overall personality models (static personality traits from Big Five and other’s of it’s likes). We have by means of this study attempted to discover, formulate and compute a dynamic model of representing the persona of an individual to overcome these limitations. In each of the domains described above an iterative model will elevate the contextual information by a manifold as opposed to the approximate static summarization of an individual’s persona. Thus, inspired by the Abhidhamma scholarship of Buddhism, we proposed PACMAN - the psycho computational model of an individual. The model theoretically represents a stochastic finite state ma- chine that encapsulates the moment by moment mental states of an individual. It is empowered by its genesis which lies in the foundations of Brentano’s western psychological theories, phenomenological thinking and strong parallels with the characteristic biomedical processes that define the psychology of an individual. Another feature which helps this model to carve it’s own niche is it’s close resemblance to a formal system. Much like a formal model is defined by a set of states and their inter state transitions, this model is also defined by it’s moments and transitional rules between these moments. It is adept in establishing a definite schema which can be ubiquitously populated by any given data reserve and helps in visualising the persona depicted. Thus, we first categorize the temporal (time to time) social media data into pre-defined mental states (concieved via Abhidhamma). These mental states are then popu- lated into the respective temporal moments of our automaton. Finally, they are used to draw various inferences about the lexical psychology and/or personalities of the respective social media users. We bring closure to our work by illustrating the functionality of our model in three different use cases. We first use our model to formalise the abstract idea of an individual in the concept of social machines. We also exemplify how our model tackles sparsity and non-uniformity of data in social media. The second use case is to model the psychological phenomenon of Anxiety, its conception, continuation etc. by means of lexical data available online. We finally introduce the lexico-psychological mental engagement factor to capture the maximum engagement of each individual with any given mental state. As opposed to the psychometric methods used to capture such a factor, our model uses lexical features to compute it. We discuss the importance of each

Abstract

Age old classical problems of the theory of society have a contemporary new dimension with the advent of computer-mediated communication. New communications rapidly transformed the internet introducing the idea of the social web. The concept of social machines as socio-technical systems, in whose machinery both humans and technology play participatory roles has given a new definition of what is meant by the term ‘social’. In this contemporary context of social machines, we intend to develop the concept of a social automaton based on a formal approach to a theory of society. The research presented is a result of analytic inquiry into the origin of society and not historical or empirical inquiry. We ask an ontological question - what is the stuff of which society is made of? After all society is objectified as an entity, to which individuals belong and which somehow is constructed/- constituted by individuals. As a result of our inquiry, we present a symbolic automaton of an individual based on Buddhist Abhidhamma formal ontology of the individual. Further, by focusing on individual’s action we develop a social automaton which is capable of handling social objects like persona, roles, relations, manners, rituals etc. Such an automaton would contribute to evolving a wholesome idea of social machine which have human as well as technology components. With a bottom up approach, we first start by understanding the fabric of an individual human being who is the basic component of a society. We establish that for understanding social reality it is imperative to begin with the formal framework of individual which accommodates this reality. To understand the framework of individual, we take inspiration from Buddhist Abhidhamma theory which provides a formal model of man as a stream of consciousness. With this groundwork, we develop a computational model of an individual as a stochastic automaton. Based on our individual automaton model, we establish that psychological processes within an individual create scope for generation of social objects. As a computational use case, we populate the automaton model with social media data of the individual and provide a machine learning module to identify mental states from social media data. We then introduce the idea of dynamic evolving persona to personality mining, by analyzing the user’s personality in terms of evolving mental factors. Once we lay the foundation for an individual as psychological being, we then proceed to give an in-depth analysis of action in an individual. With action as the key synthetic signature of the individual vi vii as a social being, we extend our individual automaton to a social automaton. Based on action theory and the sequential psychological processes given by Abhidhamma, we introduce action into our individual automaton as a three tuple - intention, body action and consequence. Further, we provide causal closure of one’s action within the framework of individual automaton. This step leads to a social automaton which has within its framework processes to identify and generate social objects such as relations, roles and institutions. With this move, social reality is conceived around the framework of social automaton.

Abstract

In recent years there has been an increased use of wheeled mobile robots for outdoor and unstructured terrain navigation. While traversing uneven terrain wheeled mobile robots are subject to various kinds wheel slip which can lead to localization errors and decrease motor efficiency. The objective of this work is to develop robust modelling techniques for simulating the kinematics, quasi-statics , and dynamics of wheeled mobile robots subject to kinematic and dynamic no-slip constraints and also develop state-of- the-art suspension mechanisms for hybrid wheeled-legged robots in order to obtain reduced slip while traversing on fully 3D uneven terrain. In order to achieve slip free motion on uneven terrain the wheeled / wheeled-legged robots have to be equipped with essential degrees of freedom that would enhance its ability to negotiate undulations on the terrain. First we perform the analysis of a wheeled mobile robot with a passive variable camber (PVC). We develop a methodology for simulating the kinematics of this robot using the techniques of kinematics for dexterous manipulation of multi-fingured palm. We also develop a framework for simulating the quasi-static motion of the robot on uneven terrain .The framework at each instant estimates the contact forces and velocity of the vehicle platform for a given set of joint velocities of the robot. This ensures that the vehicle satisfies not only kinematic no slip constraints but as well as no slip constraints that arise due to relations between traction and contact forces. In general a complete simulation of a WMR on a fully 3D terrain has been a difficult problem to solve. The best efforts so far have provided a simulation that incorporates the wheel ground contact constraints into a set of differential algebraic equations (DAEs) to estimate the full 6dof pose of the vehicle. This work integrates the quasi static contraints within the DAE framework to provide a complete 6dof evolution of vehicle on 3D terrain that respects both kinematic and quasi static constraints. We then present a motion planning algorithm connecting a starting and ending goal positions of a wheeled mobile robot (WMR) with a passive variable camber on a fully 3D uneven terrain without slipping. The overall planning framework is along the lines of the RRT (Rapidly Exploring Random Tree). The curve connecting the adjacent nodes of the RRT is a quasi-static path which is generated using the forward motion problem based on the Peshkin’s minimum energy principle which combines the force and kinematic relationships of the WMR into a nonlinear optimization problem. The output of this optimization routine is a set of ordinary differential equations (ODEs) representing the non-holonomic constraints and wheel ground contact conditions of the robot along with a set of differential algebraic equations (DAEs) representing the geometric/holonomic constraints of the robot. We extend this kinematic framework to develop the complete dynamic state space realization of the WMR on uneven terrain. The framework performs the affine form state-space realization for the WMR on uneven terrain. Second we present two models of Linear Force Actuator based hybrid wheeled-legged robots and propose a strategy to control the wheel ground contact forces to improve the no-slip margin while traversing a fully 3D uneven terrain. The mechanisms we propose aims at controlling the contact forces at the wheels in a way to improve traction as well as to ensure that the vehicle retains a desired posture. Torus wheels are used to simulate a single point of contact with the terrain which enables us to have lateral tilt on lateral surfaces which is essential while operating on fully 3D terrain. We present the quasi-static analysis of each of the mechanisms to depict the ability of the systems to control the wheel ground contact forces while negotiating uneven terrain. The first model of the Linear Force actuator based Vehicle has a one degree of freedom (1 dof ) leg and the second vehicle has a 2 dof leg. We depict the force controllability of these vehicles using feasibility plots and also perform simulations on terrains having varying geometrical parameters as well as on terrains having discontinuities twice as much as the wheel radius.

Abstract

Autonomous navigation of generic monocular quadcopter in the natural and urban environment requires sophisticated mechanism for perception, planning and control. This thesis is primarily focused on describing a framework which performs perception using monocular camera and generates minimum time collision free trajectory and control for any commercial quadcopter flying through cluttered unknown environment. The proposed framework first utilizes supervised learning approach to estimate the dense depth map for video stream obtained from frontal monocular camera. This depth map is initially transformed into Ego Dynamic Space and subsequently, is used for computing locally traversable way-points utilizing binary integer programming methodology. Finally, trajectory planning and control module employs a convex programming technique to generate collision-free trajectory which follows these way-points and produces appropriate control inputs for the quadcopter. These control inputs are computed from the generated trajectory in each update. Hence, they are applicable to achieve closed loop control similar to model predictive controller. A majority of the work in this thesis is attributed to developing a Autonomous Navigation Framework for generic quad copter which incorporates various theoretical concepts into practical implementation with ROS (Robot Operating System) as middle-ware which is running on a conventional desktop computer with Ubuntu 12.04 LTS. The functionality of the proposed approach could be described as a threefold navigation framework: perception, planning and control. The thesis is concluded by demonstrating the applicability of developed navigation framework in controlled indoors and in unstructured natural outdoors environment on a low cost commercial quadcopter ParrotTM Ardrone 2.0.

Abstract

Syntactic parsing, a major component of natural language processing, involves understanding the structure of a sentence as per given rules. Despite of enormous research in this field, syntactic parsing for Indian languages is still not at par with other languages like English. The reason being, most of the Indian languages are morphologically rich, free word-order languages (Mor- FWO) and parsing Mor-FWO languages is a challenging task. Apart from this, parsing requires sizable annotated resource called a treebank for machine learning algorithms. Treebank is a text corpus with manually annotated grammatical analysis of each sentence. Unavailability of resources for these languages is a major reason why we do not have good parsers for these languages. In this work, we have proposed methods which aid in semi-automatic development of good quality treebanks for Indian languages by proposing a method to automatically detect potential errors in treebank. With the help of this approach, manual validators can validate treebanks in less time and effort. Since the availability of Penn Treebank [44], treebanks have played a crucial role in building automatic natural language processing tools for various languages. In particular, treebanks have helped in building robust and efficient syntactic parsers. The availability of syntactic parsers is critical for the further processing of a sentence, e.g. semantic analysis [31]. We, therefore, explore methods to develop treebanks for resource poor languages because we might have huge unannotated data available with us but labeled data is very less for many languages. Anno- tation being an expensive task, if we obtain a tool which helps in building quality treebanks, development of statistical parsers would be affected a lot. Statistical parsers are generally trained on data from a singe domain and evaluation is also generally done on the same domain. While when put into natural language applications, we expect input data from any random domain. In such cases, parsers behave clumsily and their performance degrades drastically from what was evaluated. This might be because, while training parsers on data from one domain, parser might learn domain specific features which might not be valid for other domains. So, while working on these two aspects of grammatical analysis of sentences in Indian languages, I also ventured into exploring whether existing parsers can be made robust for new domains.

Abstract

In this thesis we present the development of a low cost point of care testing(POCT) device for mon- itoring cerebral oxygenation using near infrared spectroscopy (NIRS). The changes in optical measurements of tissue oxy-(HbO2) and deoxy-(Hb) hemoglobin concentration can be monitored using NIRS and these changes can be coupled with anodal transcranial direct current stimulation (tDCS) to find impairments in the cerebral micro-vessels functionality, where impairments in cerebral microvessels functionality may lead to impairments in the cerebrovascular reactivity (CVR). In fact severely reduced CVR predicts the chances of transient ischemic attack and ipsilateral stroke. The feasibility studies on healthy subjects and stroke survivors showed detectable changes in the NIRS responses to anodal tDCS of current density 0.526 A/m 2 . A clinical studyhas been conducted on 14 stroke survivors using our NIRS data acquisition (DAQ) device and our anodal tDCS device. This study revealed that the lesioned hemisphere with impaired circulation showed significantly (p <0.01) less change in HbO2 concentration than the non-lesioned side in response to anodal tDCS. The NIRS sensors used in the study are not disposable as opposed to the present state of the art technology. Our low cost NIRS-DAQ device has been further integrated with the commercially used CE-approved SomaSensor (Covidien, USA). This was done under a ”clinical study in human subjects of a CE-marked device (being used within indication) to evaluate additional parameters.” Here, Medical Devices Directive covers a diverse range of devices that are classified into one of four risk-based Classifications. Our SomaSensor based low-cost NIRS data acquisition device to monitor hemodynamics during non-invasive brain stimulation (NIBS) should be classified as Class IIa - Medium risk, often short-termnon-invasive device. The reusability of the SomaSensor was also tested as the commercially used SomaSensors are labelled as disposable. A large chunk of the complete cost of the NIRS setup comes from the procurement of the CE-approved SomaSensors. By proving the reusability of the SomaSensors the complete cost of the NIRS setup was also reduced. The performance of the SomaSensor interfaced with our low cost NIRS-DAQ device was also tested with the help of a dual layer dynamic phantom filled human blood and intralipid. NIRS and anodal tDCS have been further combined with electroencephalography (EEG) to develop a NIRS-EEG/tDCS joint-imaging method. This was done to investigate local post-tDCS alterations from baseline at the site of anodal tDCS using NIRS-EEG/tDCS joint-imaging. In a separate study the changes in local post-tDCS alterations in motor evoked potentials (MEP)-measure of corticospinal excitability were investigated. We found that post-tDCS changes in the mean regional cerebral oxygen saturation (rSO 2 ) from baseline mostly correlated with the corresponding post-tDCS change in log-transformed mean-power of EEG within 0.5Hz - 11.25Hz frequency band. Moreover, a decrease in log-transformed mean power of EEG within 0.5Hz - 11.25Hz frequency band corresponded with an increase in the MEP-measure of corticospinal excitability found in the second study. Therefore, we propose to combine NIRS-EEG/tDCS joint-imaging with corticospinal excitability investigation in a single study to confirm these findings.

Abstract

Microblogging services like Twitter enable communication at a massive scale. It has been recently reported1 by Twitter that every month, 248 million users access the microblogging platform, and create around 500 million posts everyday in more than 35 languages. This tremendous growth gives an opportunity to mine useful information that is being shared across social media. However, there exists a 140 character limitation on the posts (also known as “tweets”) that users can create on Twitter. This results in heavy use of emoticons, abbreviations, misspellings and has lead to various linguistic “innovations” that render the traditional text analysis techniques less effective. Another interesting aspect of such tweets is the usage of semantico-syntactical constructs called “hashtags”. Hashtags are “#”-prefixed keywords used by people in order to organise the meaning of their tweets. Also, hashtags enable classification of tweets, since the posts using same or similar hashtags are expected to be semantically related to each other. The challenge posed by hashtags is the fact that most hashtags are not simply “#”-prefixed keywords, but “#” symbol prefixed with concatenation of various words or phrases which are not space delimited. For example, consider the hashtag - “#NSAvsSnowden”. We observe that this hashtag is essentially “NSA vs Snowden”, which is not a single keyword, but a concatenation of various words. In this thesis, we discuss and compare various approaches in order to “segment” the hashtag into meaningful words. Also, our task extends beyond just the segmentation of hashtag - we present a unified framework to also perform “entity-linking” on various constituent entities in a hashtag. Entity Linking is an established IR task, where the goal is to extract latent semantics from plain text by linking the text to a knowledge base (KB) such as Wikipedia. Consider, for example, the following text - “Snowden reveals classified information from NSA”, we first need to identify various entities in this piece of text, followed by “disambiguating” them, and establishing a link between those entities and some knowledge base (KB) so that we have additional contextual information available about the concerned entity. This approach has been found to be instrumental in order to teach the meaning of text to machines, which is otherwise meant for human consumption. Hence, after performing entity linking on the segmented hashtag - “Snowden vs NSA”, we would have enriched the text with additional semantic information by establishing links between “Snowden” and corresponding Wikipedia page - “Edward Snowden”, between “NSA” and “National Security Agency”. “NSA”, in principle, could also refer to “National Sports Academy” or “National Security Act”, and this is exactly where “disambiguation” becomes important for mention-resolving, which employs contextual information to perform this task. Since hashtags are human curated labels associated to tweets, our premise is that segmenting and linking the entities present within the hashtags could therefore help in better understanding and extraction of information shared across the social media. Traditionally, most of the IR tasks have treated hashtags as either a single word, or have ignored them for all practical purposes. We demonstrate how extraction of semantics from tweets improved, when additional semantic information was made available by our system by segmenting and entity-linking hashtags. We demonstrate this by performing various experiments on NEEL Challenge Dataset, and a human annotated subset of Stanford Sentiment Analysis Dataset, which has also been made public to ease future research in this area. We have achieved the P@1 score of 0.914 on NEEL Dataset and 0.873 on the manually annotated Stanford Sentiment Analysis Dataset for hashtag segmentation and linking. We also showcase how our approach leads to improvements in the task of “Semantic Microblog retrieval” and “Semantic Hashtag retrieval”. Microblog retrieval refers to retrieval of a ranked list of microposts given a query Q. Hashtag retrieval, on the other hand is a relatively newer IR task. It basically refers to retrieving a ranked list of the top-k hashtags relevant to a user’s query Q. To retrieve information related to a user’s interest, for instance, “Rock concerts”, it’d be very helpful to the user if they can be suggested a list of hashtags which are commonly used in relation to “Rock concerts”. By tracking these hashtags, a user can gain information about rock concerts via the posted tweets. However, it’s not possible for the user to manually figure out all the hashtags that are used across Twitter, relevant to their interest. In this thesis, we also address this problem. In order to solve these two retrieval problems, we propose and discuss a virtual docume

Abstract

One of the DNA lesions formed due to free radical damage of thymine is urea, occurrence of which in DNA blocks DNA polymerases, and has been shown to be lethal. Recently, it has been shown that urea is capable of forming hydrogen bonding and stacking interactions with nucleobases, which are responsible for the unfolding of RNA in aqueous urea. Base pairing and stacking are inherent properties of nucleobases; since urea is able to form both, this study attempts to investigate if urea can mimic nucleobases in the context of the nucleic acid structures by examining the effect of introducing urea lesions complementary to the four different nucleobases on the overall helical integrity of B-DNA duplexes and their thermodynamic stabilities using molecular dynamics (MD) simulations. The MD simulations resulted in stable duplexes without significant changes in the global B-DNA conformation. In agreement with experimental results, the urea lesions occupy intrahelical positions by forming hydrogen bonds with nitrogenous nucleobases. Further, these urea lesions form hydrogen bonding and stacking interactions with other nucleobases of the same and partner strands, which are sometimes similar to the nucleobases in typical B-DNA duplexes. Direct hydrogen bond interactions are observed for the urea-purine pairs within DNA duplexes, whereas two different modes of interactions, namely direct hydrogen bonds and water-mediated hydrogen bonds, are observed for the urea-pyrimidine pairs. The latter explains the complexities involved in interpreting the experimental NMR data reported previously. Binding free energy calculations were further performed to understand the thermodynamic stability of the urea incorporated DNA duplexes with respect to pure duplexes. This study suggests that urea potentially mimics nucleobases by pairing opposite to all the four nucleobases and maintains the overall structure of the B-DNA duplexes.

Abstract

The amount of information available on the web is increasing rapidly. The number of internet users is also increasing every day. A significant section of internet users is monolingual. They want to express themselves in their native language and also seeking information in the same. Hence, multilingual content over the internet is also increasing at a rapid pace. There is a need of systems which empowers such internet users to let them express themselves in the language of their choice and can provide them answers to information requests irrespective of the language barrier. Cross-lingual term associations are very important for many interlingual applications. Machine Translation system is a sub-field of computational linguistics which aims at translating text from one language to other. All MT systems at the core depend on bilingual dictionaries. The bilingual dictionaries are important resources for such NLP applications as statistical machine translation and cross-language information extraction systems. The bilingual dictionary is one such important resource where entries are word translations. They also can serve to enhance existing dictionaries, for second language teaching and learning. Manually created resources are usually more accurate and do not contain noisy information, in contrast, to automatically learned dictionaries. Scientific community seeks for the methods to achieve similar accuracy level and broader terminology scope by automatic means. In this thesis, we try to address the problem of generating bilingual dictionaries automatically. We propose two different approaches to generating bilingual dictionaries for English-Hindi pair. Both the approaches proposed in this thesis are language independent and hence can be used to build dictionaries for other language pairs which are phonetically rich. Hindi is very under-represented on the web because of many technical and socio-cultural reasons. Many languages, like Hindi, suffer from the limited availability of the language resources/ tools. It is crucial to develop a language-independent approach for such languages. In this thesis, though we have shown our experiments for English and Hindi language pair, our approach can be easily extendable to other language pairs as well. We didn’t use any language specific resource in the approaches which we proposed here. We have chosen two of the most reliable document resource for English-Hindi pairs viz. Wikipedia and news streams. We use Wikipedia as comparable corpora and ever growing news corpus as Quasi Comparable corpora. In the first approach proposed, we exploit the structural properties of documents to build a bilingual English-Hindi dictionary. The main intuition behind this approach is that documents in different languages discussing the same topic are likely to have similar structural elements. We proved by applying our approach on Comparable corpora containing documents having structural properties. Documents are inherently structurally divided into many sections and subsections. We link these sections to corresponding in the other language using a statistical approach. These mini sections are further divided into comparable sentences. After using some basic NLP techniques, we chose the most co-occurring pairs, which are a candidate for being a dictionary pair. One of the major contributions of this approach is that the dictionary contains translation and transliteration of words which include Named Entities to a large extent. The approach proposed here is language independent. In the other automatic method, we proposed exploits continuous quasi-comparable corpora to derive term level associations for the enrichment of dictionaries generated by the first approach. The approaches uses easily accessible (i.e, comparable and quasi-comparable) data sources to derive dictionary entries. The News is one such ever generating quasi-comparable corpora. Using this approach on a news dataset new vocabulary can be added to the dictionary. According to the same taxonomy, we try to link news stories across languages with the same focal event which might serve as comparable corpora. Afterwards, we apply dictionary generation algorithm to extract bilingual dictionary. Our method is completely automatic and does not rely on any labelled information. We compare the dictionary generated by our method to the dictionary generated via the first approach. In this thesis, we show this approach is able to derive interesting term level associations across languages. The dictionary generated from the Wikipedia is of high quality because of the highly structured information like title, section, sub-section etc. and interlingual links. We demonstrate that our method is able to find some associations which are not present in the Wikipedia using this approach.