Abstract

Abstract Machine Translation (MT) is a task in Natural Language Processing (NLP), where the automatic systems are used to translate the text from one language to another while preserving the meaning of source language. In this work, we provide our efforts in developing a rule-based translation system on the Analyze-Transfer-Generate paradigm which employs morphological and syntactic analysis of source language. We utilized shallow parser for Hindi language along with dependency parse labels for syntactic analysis of Hindi language, developed modules for transfer of Hindi to English and generation of English language. Due to wide difference in word order of the two languages (Hindi following SOV and English SVO word order), a lot of re-ordering rules need to be crafted to capture the irregularity of the language pair. As a result of drawbacks of the aforementioned approach, we shifted to statistical methods for developing a system. A wide variety of machine translation approaches have been developed in past years. As each model has its pros and cons, we propose an approach where we try to capture the advantages of each system, thereby developing a better MT system. We then incorporate semantic information in phrase-based machine translation using monolingual corpus where the system learns semantically meaningful representations. Recent studies in machine translation support the fact that multi-model systems perform better than the individual models. In this thesis, we describe a Hindi to English statistical machine translation system and improve over the baselines using multiple translation models. We work on MOSES which is a free statistical machine translation framework, which allows automatically training translation model using parallel corpus. MOSES provides support for multiple algorithms for training a model. We propose an approach for computing the quality score for each translation by using automatic evaluation metric as our quality score. The computed quality score is used for guided selection among the translations from multiple models, thereby providing a better system. We have used support vector regression to train a model using syntactic,textual and linguistic features extracted from the source and target translation with evaluation metric as our regression output. Quality Estimation of Machine Translation is a task where the system tries to predict the quality of output on the basis of features extracted from source and target languages. The system dynamically(run time) computes a quality score corresponding to each translation, which is the measurement for the correctness of the output. Different from MT evaluation, quality estimation systems do not rely on reference translations and are generally addressed using machine learning techniques. The approach offers a great advantage to the readers of target language as well as for professional translators.

Abstract

Histopathology, the study of tissues, has a major contribution in determining prognosis of diseases at the tissue level. Quantitative analysis of histopathological images is used to determine the future course of treatment. However, the analysis is a very laborious process given the complexity of structures in tissues. Pathologists require a lot of expertise and skill to perform the analysis. Hence, development of Computer Aided Diagnostic (CAD) algorithms which rely on automated image analysis, is of great interest. CAD development in histopathology is a relatively new field of research. Some of the major hurdles in this field are non - availability of public datasets, difficulty in acquiring data and ground - truth marking. This thesis aims at developing CAD tools for assisting pathologists. In our work, we have identified two problems in this field and provided solution for the same. The first is identification of stromal regions in tissue and the second is detecting mitotic candidates. Both these are ne cessary for assessing and staging cancer. For stroma identification we propose a novel block - based approach of suppressing foreground texture to learn discriminative textural features in a block. The results of the bloc k - based stage are refined to obtain t he stroma segments. We have teste d our method on 3 datasets (two public and one locally acquired) varying in size and type (breast and prostate tissue) and have obtained around 80% accuracy and 90% sensitivity. Both the qualitative as well as the quantitat ive results show an improvement over the existing methods. The existing algorithms for mitosis detection employ a traditional approach of classification post feature - extraction. A major road - block in this problem is the presence of visual clutter ( false positives , similar in appearance to mitotic candidates ). We use biological cues to design rejection stages to get rid of the false positives for mitosis detection followed by classification. The pipeline design comprising of rejection stages adapt to variability in the magnification levels of datasets. This was confirmed by determining percentage of rejection after each stage on a locally acquired dataset at different magnification settings. We further improve the performance by incorporating an eff icient classifier regenerative random forests . The algorithm was validated on MITOS dataset against the ground truth and a recall of ,0.81,precision of 0.84 and f - measure of 0.83 was obtained.

Abstract

Abstract Speech communication is a major medium of communication among human beings. All human beings have flexibility in changing parameters of speech like loudness, duration, pitch and intonation within their voice limits. Voice imitation is a fine art in which the professional imitator develops his ability to mimic other speakers. Professional imitators have the ability to convince the listeners that they are listening to someone else. It is the flexibility of speech production mechanism that allows imitators to perform voice imitation. In this work, analysis and synthesis of mimicry speech has been carried out. Previous studies on voice imitation have focussed on features at segmental and suprasegmental levels. The present analysis of voice imitation is carried out at suprasegmental, segmental and subsegmental levels. The supraseg- mental features studied in this work are the instantaneous fundamental frequency (f0) contour and du- ration. The segmental feature used is linear prediction cepstral coefficients (LPCCs). The strength of excitation at the instants of significant excitation and a loudness measure reflecting the sharpness of the impulse-like excitation around epochs are the subsegmental features. The study focusses on how close the imitation is to the target speech and how much deviation happens from his natural speech. The ob- servations are correlated with perceptual studies. The suprasegmental and subsegmental features show a tendency to get closer to the target features. The segmental features which represent the vocal tract shape and size are difficult to change for the imitator. The importance of source and system parameters is studied by synthesis experiments. The natural utterance of the professional imitator is transformed into imitated utterance by variations in excitation source and system parameters. Subjective studies on the synthesised speech shows that suprasegmental features play a significant role in imitation. This analysis is extended into an application where the natural and imitated speech are distinguished using neural network models. The models are built using both excitation source and system features. The models built using excitation source features have better performance than the models built using system features.

Abstract

Inductors are one of the major constraints in the design of Radio Frequency Integrated Circuits (RFICs). This is because they consume large area and the quality factor is very low. Low quality factor is attributed to the parasitic losses at the higher frequencies. Substrate eddy current losses is one among the contributing parasitic losses. Research in this area is towards fabrication of 3D-MEMS inductors and synthesis of new magnetic materials. 3D MEMS inductors are difficult to implement in monolithic RFICs due to packaging and stability issues. Patterned electrical ground or amorphous magnetic ground planes beneath the inductor are employed to reduce the eddy current losses. To improve the inductance density and quality factor of on-chip spiral inductors in RFICs employment of magnetic materials is one of the popular approaches. In cases where bulk magnetic material is used around inductors, eddy current losses become significant at higher frequencies restricting their usage to lower frequencies. Patterning the magnetic material has shown to reduce these drawbacks. Permalloy, which is known for its soft magnetic properties and compatibility with CMOS process, has been used widely for enhancement of inductance of spiral inductors but the frequency is restricted to <1GHz. When patterned as magnetic bars and square dots, improvement in the performance of inductors has been observed at frequencies >1 GHz and in such cases Ferromagnetic Resonance(FMR) of the patterned material has set the limitations to employment at further higher frequencies. For magnetic materials, formation of magnetic domains and their dynamics may play a major role in defining their properties at higher frequencies. In this thesis, effect of size and shape of Permalloy patterns in the formation of magnetic domains, their dynamics, and the influence of these domain controlled patterns on properties of spiral inductor are investigated. It is shown experimentally that the shape and size induced domain patterns in Permalloy play a varied role in the improvement of inductance. Micromagnetic simulations with Object Oriented Micro-Magnetic Framework (OOMMF) are used to determine the domain patterns. Secondly, 60 nm thick domain-patterned-Permalloy have given more improvement in inductance and quality factor compared to 96 nm thick patterns. OOMMF simulations show the same trend that 60 nm patterns which contain cross tie walls perform better than 96 nm patterns which contain asymmetric Bloch wall. Permalloy domain-controlled patterns have also been incorporated as magnetic shield and an improvement in the performance of the inductors is observed. All the experimental observations confirm the necessity of incorporation of proper magnetic patterns whose shape and size could be controlled to form appropriate domain patterns for high frequency applications like inductors in RFICs.

Abstract

This work investigates some interesting questions arising out of mechanistical studies of purine riboswitches. As in any other riboswitch, the purine riboswitch has an aptamer domain linked to the expression platform of mRNAs. Discriminative ligand binding by the aptamer domain is known to regulate the mRNA expression. The aptamer domain of the purine (adenine, guanine and their variants) riboswitches con- sists of three helices P1, P2 and P3 which form a ‘Y’-shaped three-way junction (3WJ). The junction loops are J1-2, J2-3 and J3-1 which connect the corresponding helices. Loops L2 and L3 cap the helical stems P2 and P3, respectively. The helices P1 and P3 are collinear - forming a coaxial stack. Each of the stems P2 and P3 usually consist of 7 base-pairs. Helix P2 is adjacent to P3 and the two are held in parallel arrangement by interactions between the terminal loops L2 and L3. Complementary base pairs between nucleotides of the two loops have been found to form a kissing loop interaction. Earlier work carried out in our group on the study of the add adenine (A) riboswitch ap- tamer. It has provided insightful answers to the question how the recognition of the adenine by the ligand free APO form of the riboswitch is communicated to the downstream region via a conformational change in the downstream expression platform which is responsible for gene regulation [1], [2]. These ideas, as understood from computational studies as well as from biochemical and biophysical experiments conducted in vivo or in vitro have been discussed in the first chapter. Crucial to this mechanism were (i) involvement of kissing loop interactions in facilitating the stabilization of the binding pocket, and (ii) ligand binding mediated stabi- lization of a helix which is involved in the switching event. In the second chapter we have reviewed single-molecule experiments carried out with sim- ple hairpin, hairpins involving kissing loops and aptamer domain of the adenine riboswitch, both in the presence as well as the absence of the ligand. It has been shown that a detailed understanding of the mechanism of helix stabilization due to tertiary interactions may be ob- tained using computational models of these motifs. The third chapter explores the standard worm-like chain (WLC) model for semi-flexible poly- mers. Extensive simulations of the WLC models have been carried out to observe the variation of the end-to-end distances of the polymer at different stiffness of the polymer using a bead- spring and a bending potential. The fourth chapter applies the WLC model to the problem of understanding the behaviour of RNA hairpins by suitably assigning the bond stiffness, base-pair strength, base-pair breaking threshold and strand stiffness under external pulling force in an attempt to simulate some of the observations in Chapter 2. In addition to this, extensive work was done in implementing a querieable database which relates the occurrence frequencies of different base pairs in RNA with their crystal and ground state optimized geometries (calculated at different levels of QM theory) and corresponding intrinsic interaction energies, RNABP COGEST, which was published. Since it is work that is distinct from the rest of the study, the work is reported as an Appendix titled ‘RNA Base Pair Count Geometry and Stability (RNABP COGEST) - Database Development and Imple- mentation of Useful Features’.

Abstract

Abstract Natural language can be easily understood by everyone irrespective of their differences in age or region or qualification. Based on various theories proposed in psycholinguistics, Roger Schank came up with the proposition that “the basis of natural language is conceptual and the conceptual base is inter-lingual”. According to Indian grammatical tradition, “all content words in a language are either verbal roots or derived from verbal roots”. Inspired by these two ideas we are proposing a new set of meaning primitives based on overlapping senses of verbs. Using these meaning primitives we are trying to develop a new ontological classification of verbs. We propose that meaning of a verb is composed of seven fundamental senses implicating different attribute sets. In this proposal maximally overlapping sense of verbs is taken as “happening” (Sanskrit bh¯ava), as in the Indian linguistic traditions. We have rendered it as a recursive form with a structure “change of state in context”. Each fundamental sense is a “kind of happening” giving different contiguity of states with different attribute set as context. We derive the meaning of verbs from a universal verb happening such that every verb is uniquely specified across languages. The method has been applied to develop ontologically informed etymon in English and Indian language, Sanskrit. The primitives along with ontological attributes have been used to explicate the meaning of verbs. This form of meaning specification of verbs is used to develop language independent ontological classification of verbs which in turn can be used in Natural Language Processing applications like semantic search and machine translation.

Abstract

The ever increasing demand for communication bandwidth and inefficient usage of the existing spectrum has led to spectrum scarcity. In this light, spectrum should be managed as a scare resource. For radio communication systems, efficient utilization of spectrum is the key requirement. The inefficient usage of the existing spectrum can be improved through opportunistic access to the licensed bands without interfering with the primary users. This introduces the concept of dynamic spectrum access and cognitive radio. Applying these techniques provides better throughput even in congested spectrum exploiting better propagation characteristics. Cognitive radio introduces cognition and adaptation at the physical layer. Cognition should also be included in upper layers of the network to effectively utilize the benefits of cognitive radio and dynamic spectrum allocation. Thus, cognitive network is an intelligent network that can adapt the transmission, reception and other network parameters of either a network or a wireless node to achieve efficient communication without interfering with primary users. Routing and spectrum access in cognitive environment is a challenging task as the channel availability is constrained by the presence of primary users. The problem of routing in cognitive networks targets the creation and maintenance of wireless multi-hop paths among cognitive nodes by deciding both the spectrum to be used and the relay nodes of the path. In the recent times various protocols were designed for cognitive networks. Cross layer protocol design is suitable for cognitive networks as dynamic adaptation of parameters is required. This thesis proposes a cognitive cross-layer routing protocol which jointly does power control and spectrum selection for multiple hops in the routing path. In the proposed protocol, routing is defined as a multi-objective optimization problem to derive the cost function which is used as the routing metric. The work shows that the proposed routing metric is convex and gives Pareto optimal solution. The thesis also proposes a medium access protocol for cognitive networks. The proposed protocol provides solutions for the multi-channel hidden node problem, channel synchronization, channel scheduling and other challenges that arise in the design of cognitive networks. For performance evaluation, the proposed protocols are implemented in NS2 and a comparative analysis is also done against existing protocols from the literature. The impact of cognitive user density, cognitive user speed and primary user activity model is analyzed in the thesis. The thesis also develops a Markov model for the performance analysis of the proposed cognitive MAC protocol. The developed model is generalized and can be used for performance analysis of any IEEE 802.11 power saving mode based cognitive MAC protocol for ad hoc networks. This analysis is used to decide the design parameters of the MAC protocol. Protocols specific to cognitive wireless sensor networks are also proposed in this thesis. Cognitive wireless sensor networks are a specific type of cognitive networks which are used to monitor some entity. Traditional wireless sensor networks work in ISM band which is shared by some other technologies like WiFi and Bluetooth. Sensor nodes with cognitive capability are still limited with battery constrains, and efforts are required to increase the lifetime and other performance measures of the network. This thesis also proposes an network architecture for cognitive wireless sensor networks which consists of hierarchical distributed minimum spanning trees. The work also shows that this multilevel network provides fault tolerance, admits simple routing, and offers easy extensibility with power efficiency. The hidden primary user problem arises in cognitive users with passive listening primary users. For avoiding interference to hidden primary users, disable region should be calculated considering the communication range of both primary and cognitive users. Disable region is a region in which cognitive user communication should be suspended to avoid interference to primary users. Cognitive users must be aware of primary user’s location within the region of interest for calculating disable region. This thesis applies a leveling and sectoring based approach for primary user localization. This will avoid the interference to hidden primary users. The cognitive users within the disabled region are sent to sleep mode until a free channel is detected in that area by spectrum sensing module. This approach also saves energy of cognitive users.

Abstract

This thesis deals with design and development of an Oral Personal History system and the issues associated with the spontaneous speech. The primary goal of the thesis is to provide a comprehensive procedure with a step-by-step approach towards collecting and analysing the data. The important contributions of this thesis are : • An important issue is how are we designing the questionnaire and the precautions taken in preparing the questions. These narratives are passed down verbally from generation to generation beyond the lifetime of any one individual. • We explore the issues faced during the interview and how are we transcribing the raw material into written form. We also address the advantage of usability of the spontenous speech which can be used for further research purposes and the morphing of the data where we show privacy concern when the data is exposed to public. • The experiments were conducted by two methods, one was by conducting a survey and another one was by subjective analysis. Surverys are conducted to determine the effectiveness of a problem. They are conducted to build the similarities among the different categories of their views based on some segment, and it is easy to generalise the results. • In this thesis, the issue of privacy of the narrator using survey-based approach was implemented to obtain the responses to privacy-based concerns of the users who were taken from top categories of the society namely, Academicians and IT Professionals. • Based on inferences from the survey, we propose a speaker anonymization technique using MSASB to address the privacy concern.

Abstract

This project aims at investigating spatial cognition and melodic contours in Hindustani classical music. The descriptions of ragas in musical literature up until recently has been more in terms of static qualities such as discrete notes, scales, and prominent pitches which are not enough to describe the dynamic character of Hindustani music. The allusions to motion and flow in Hindustani music are evident in the sound and the physical movements of the singers, as well as in the vocabulary used to describe the music. Performers of this music are known to bring out the motional thinking in the improvisation of this music through elaborate hand gestures and body movements. Through this thesis, we explore the role of gestures and motion metaphors in concept forming for various parameters of music, and how gestures express these concepts. We carry out this analysis by an experiment consisting of four ragas performed by different perform-ers, and analyse their gestures through annotations and automatic feature extraction. We find evidence linking rhythmic and melodic gestures strongly with stress and prosody, and flow and motion respec-tively. We study the metaphorical character of rhythmic gestures by studying its role in representing stress points in spoken language. We also analyse the dependency of gestures related to melodic flow by their relationship with the poetry of the compositions and prosodic features. We investigate these mental models further through the method of sound tracings. We study these through taking listener’s sketch-based responses to musical phrases. We analyse this data by building a pattern classification system to classify different phrase types. We then map the analysis to generate new musical material as a computational creativity tool. This tool uses a constraint satisfaction based model which is possible to adapt to newer schemes of music generation and visu-alization. We therefore capture more of these rules through an ontology of knowledge in Hindustani music. We apply spatial layouting algorithms to analyse the features of such an ontology of categories in Hindustani music, and analyse it as a tonal space based model. We provide a way for it to represent proximity conceptualization in Hindustani music, and a way to generate different affective combinations of tonal clusters. We compare it with the combinatorial space of colour combinations and demonstrate a way to computationally pick diverse or similar tonal clusters varying in terms of different parameters. Thus we present a way to understand different ways in which Hindustani music is conceptualized spatially, both during improvisation, and as categories.

Abstract

Abstract With the advent of web 2.0, the world has witnessed an immense growth in social interactions over the web. Social media allows online users to share their opinions and express their thoughts with others. Users can post their views in the form of blog posts, tweets, product reviews, etc. With the exponential growth in the use of social media, there was a need to automatically interpret and analyze this large volume of data. Researchers worked on analyzing user sentiments and summarizing users’ views and opinions. Online debate sites are one of those social media platforms where users can take a stance and argue in support or opposition of the debate topics. This platform provides a rich collection of differing opinions on various topics. Online users’ views mainly consists of opinions and facts. Facts include objective world known information which strengthens users’ stance. Objective information is very reliable to persuade others. On the other hand, opinions include users’ sentiments towards a debate topic which means subjective remarks are made towards either stance. Users’ posts can have a mixed set of emotions in which some sentences might support one topic and other might oppose that same topic. It can contain positive or negative remarks about both the debate topics. Users structure their arguments in such a way that overall debate post helps them achieve the task of supporting their chosen stance. Considering the rich collection of mixed bag of sentiments provided by online users, we carried out the task of identifying stance users are supporting in these online debates. We identify sentence level user intentions by assigning sentiment carrying words in the sentence to appropriate debate topic. For this subtask, we have devised domain independent intention tagging schema. Argument structure of the debate post is then built using these individual sentences’ intentions and thus debate posts’ stance are classified accordingly. The online debate is a multi-party conversation where users can rebut other users’ opinions. We generate inter-debate user interactions graph and use gradient ascent strategy to improve stance classification accuracies further. With the huge amount of opinions available about the debate topics, it was required to summarize these debates to facilitate users’ for easier understanding. Considering the current state of Natural Language Generation systems, we have opted for Extractive Summarization approach where debate sentences are ranked with features like sentiment richness, topic directed sentiment strength, topic relevance, etc. Our system significantly outperforms several baseline systems and show 5.2% (ROUGE-1), 7.3% (ROUGE-2) and 5.5% (ROUGE-L) improvement over the state-of-the-art opinion summarization system. The results verify that topic directed sentiment features are most important to generate effective vii viii debate summaries. Major contributions made in this thesis include: • Novel Domain independent intention tagging schema to mark sentence level user intentions. • Argument structure of the debate post to aggregate contributing sentence level intentions. • Gradient Ascent approach to improve stance classification based on inter-debate user interactions structure. • Novel debate summarization system to summarize the online debates. Keywords: Topic Directed Sentiment Analysis, Argument Structure, Gradient Ascent approach, Online debate summarization features