Abstract

The social and political shifts being perpetuated by the Internet and social media are dominating contemporary discourse around politics, society and technology. The difficulties in regulating online practices, in advancing Internet literacy and moderating machine learning algorithms that govern online behaviour, and moreover in preventing malicious users from infiltrating the Internet have been highlighted by scholars in this domain. As the Internet and social media have become a ubiquitous and inseparable part of our lives, so have the social shifts brought on by them. The influence of modern technology on human behaviour, both personal and social, poses a great challenge to one of the most fundamental democratic principles; human free will. This thesis highlights the anti-democratic phenomena emerging out of the incessant use of the Internet in daily life, while also looking at precursors for these phenomena through history. This thesis employs formal analysis using natural language processing as well as actor-network theory to comprehend some of these phenomena. It also connects various theories proposed by thinkers like Kierkegaard, Chomsky and Foucault, and ideas expressed by researchers, authors and journalists in recent years, to sufficiently understand these phenomena. Through the discussions and analyses presented in this thesis, the dynamics of the social ills posed by the Internet can be better understood, and this can inform further research as well as policy-making aimed at combating these issues

Abstract

Hypernasality is a prominent characteristic of speech production observed in subjects with Cleft Lip and Palate (CLP). In the clinical diagnosis of individuals with CLP, it is essential to identify and assess the hypernasality in speech. The assessments techniques are invasive, which affect the natural way of speech production. So automatic objective assessments have gained popularity as they are repeatable, easily accessible, and cheap. Nevertheless, several works in literature that try to address this problem and try to extract fine features which capture the effects of hypernasality on the speech signal. However, almost all the current works try to address the problem in individual vowel sound units. In contrast, this work tries to address the problem of detection and assessment of hypernasality from continuous speech which involves complex articulatory movement. It is similar to the assessments carried out by speech and language pathologists. The hypernasality is a result of an abnormal vocal tract, where the air leaks through the clefts into the nasal tract, resulting in nasalization. The nasalization alters the spectrum of the speech signal in different ways depending on the configuration of the vocal tract. So accounting all different changes for all possible sounds is impossible. Generally, the nasal tract is different from oral tract as it is longer and has no moving articulators. This thesis proposes Single Frequency Filter Bank Based Long Term Average Spectral (SFFB-LTAS) features that enhance the contrast between nasal tract an oral tract for detection and assessment of hypernasality from continuous speech. Further, the impact of nasalization on excitation source has been studied. The experimental results show that the information of hypernasality is manifested even in excitation source evidence and it is complementary to vocal tract features. The results show that both the excitation source and vocal tract system features are useful for the evaluations. We also present a new database, called New Mexico Cleft Palate Center- Cleft Lip and Palate (NMCPC-CLP) database. It consists of continuous speech utterances from CLP subjects and normal subjects collected at New Mexico Cleft Lip and Palate Center.

Abstract

The advancements in hyperspectral remote sensing escalated the research in forestry-related to species identification, discrimination, and health assessment. The technology aided in the development of a spectral library of various species that enable researchers to map species without the need foran expensive field inventory. The research focus of this thesis is to generate a temporal spectral database for ten selected species of Araku forest, the Eastern Ghats using field inventories (multiple seasons), hyperspectral satellite data, and canopy upscaling models. The properties of leaf/bunch of leaves sampled from the bottom of trees may vary, and a direct analogy with canopy properties will lead to substantial overestimation/underestimation of the spectral reflectance of the tree.Hitherto researchers applied upscaling techniques with wellknown PROSPECT, SAIL, FLIGHT models, but so far, no model (in forestry) developed for leaf upscaling, within the country using field-collected leaf reflectance spectra and their biochemical parameters. Towards this objective, the current study proposes an upscaling model “SEMULS” (Semi-Empirical Model for Upscaling Leaf Spectra), a unique and first of its kind of model to upscale leaf spectra to canopy. The spectra extracted for the selected species from satellite data are correlated with canopy spectra derived from leaf reflectance spectra obtained from the field using SEMULS. The outputs of the model are validated for its applicability by comparing with upscaled spectra derived from popular PROSPECT and DART 3D models. The SEMULS model is unique in retaining the minute dips and peaks of the spectra, which are important in discriminating (80% matching at 5deg or less threshold) species, unlike other models that smoothen spectra using different filters. The upscaled canopy spectra (SEMULS) is further used to classify selected species in the study area from CHRIS-PROBA satellite data, using the SAM method and validated based on GPS locations observed from the field. The study also proposed few experiments such as upscaling spectra for chlorophyll region without field inventory or upscaling model, mapping species - from the spectra extracted from literature and use of spectra of one location to map same species in other regions, discriminating species using DART model and an algorithm to decompose Sentinel data to hyperspectral data. The developed database enhances the reliability and comparability of the datasets allowing multi-temporal and multi-scale studies. The database is uploaded in the NISA portal (spectral library) which will serve present and future researchers, in identifying the species without field inventory,and/or for any spectra related research.

Abstract

The English word ‘daedalian’, which means ‘ingenious, intricate, and confusing’, is a word that originates from the Greek mythological character Daedalus, who is believed to have built a labyrinth so complex, that he himself couldn’t escape its confines- but you would be forgiven for not knowing that. Daedalian is not a commonly used word, and its synonyms like ‘complex’ and ‘confusing’ are favored in day-to-day speech. But like the title of this thesis shows, rare words are still found in written text and this creates a need to simplify this text to make it accessible to the layperson. In this thesis, we study the problem of text simplification, the challenges it presents, and the different ways these challenges can be overcome. This thesis touches upon various topics along the vast expanse of text simplification research, starting from generic improvements to encoder-decoder models that are used for controlled Natural Language Generation (NLG) tasks (including text simplification), to domain-specific text simplification challenges and the specialized solutions that are required to solve them. Recent advances in deep learning and NLG research have enabled sophisticated text simplification models. The central idea behind most approaches to text simplification, and indeed most other controlled text generation tasks, is the encoder-decoder model. Thus, we believe that controlled text generation tasks like text simplification stand to gain by using improved encoder-decoder models. Encoder-decoder models have been vastly improved by the introduction of the attention mechanism, and architectures like the transformer, but there is scope to further improve these models by explicitly improving the encoder outputs. We do this by using a co-attention based loss in addition to the standard cross-entropy loss to train the model. We evaluate this proposal on multiple tasks and show that this not only benefits text simplification but also improves performance in adjacent tasks like style transfer and sentence compression. Another problem with current text simplification research is the lack of suitable datasets that are representative of real-world text simplification needs. We enumerate the problems with existing text simplification datasets and propose a new dataset that comes from diverse sources, has a richer vocabulary and is more representative of real-world text simplification tasks. We present various approaches to simplify text from our proposed dataset and recommend methods that can be used to make the generated text more readable and semantically similar to the source. We also study domain-specific text simplification, with a specific focus on the medical domain. We study the unique challenges that come with simplifying medical text and propose an unsupervised approach to solve the problem of medical text simplification, which goes a long way towards overcoming the challenge of the lack of parallel data in the medical domain.

Abstract

Positive counterions, such as metal cations, are necessary for the electrostatic neutralization of the polyanionic ribonucleic acids (RNA) as the negatively charged RNA backbone folds onto itself to adopt a 3D structure. Magnesium ions are the most frequently observed metal ion in experimentally determined RNA structures, and have been noted for their influence on the stability, dynamics and function of RNA molecules. Identifying motifs at magnesium-RNA sites is necessary to better characterize the role played by the ions in RNA molecules. Previously available tools to detect magnesium-RNA motifs are either based on manual annotation or rigidly defined ion-RNA binding characteristics that cannot be used to describe many recently identified motifs. In this work, a set of methods have been developed to help detect, characterize and discover magnesium-RNA motifs that overcome the limitations of previously available tools. A graph-based querying system “MagMotiRNA” has been developed to query for welldefined motifs in literature as well as user-defined patterns at magnesium-RNA sites. In keeping with the evidence regarding variability in magnesium-RNA binding patterns and the role played by magnesium in stabilizing certain inter-nucleotide contacts, the tool allows for defining queries with fuzzy ion-RNA binding descriptions and queries involving the local RNA structure at the site. The results from MagMotiRNA were benchmarked against MgRNA tool, the prevalent magnesium motif detection tool, and the discrepancies in the results could be largely attributed to the higher sensitivity in the hydrogen bond detection method used in MagMotiRNA. The occurrences of five distinct base-coordinated magnesium-RNA motifs in a redundant dataset were analyzed and the distinct structural contexts in which they occur were studied. The case study revealed that the ion-RNA binding patterns at these contexts are often variable, and hence it may not be appropriate to define magnesium-RNA motifs solely based on ion-RNA binding details. Instead, a bottom-up approach is proposed for defining motifs based on the invariant features in the local RNA structure at these sites that are indicative of the putative role played by magnesium ions. A set of heuristics has been developed as part of this work whereby conserved magnesiumRNA sites can be detected in an RNA molecule by analyzing multiple structures of the molecule. Building upon these heuristics, a metal-RNA binding profile was created for 19 RNA molecules by transposing all metal-RNA sites associated with these molecules in a redundant dataset onto a representative structure. The rather complex problem of studying magnesium-RNA sites given the sparsity of reliable experimental data would benefit from these profiles of frequently conserved metal-RNA sites provided by the MagMotiRNA-Atlas database. The Atlas equips users to perform top-down exploration of magnesium-RNA motifs, and can act as a complementary tool to the bottom-up analysis of magnesium-RNA sites developed in the other MagMotiRNA tools.

Abstract

This thesis is a culmination of the work done on event detection, annotation and analysis. We present here the development of the detection of events on a large scale for low resource languages in two ways from a computational and a linguistic perspective. From the linguistic perspective, we discuss the creation of a language specific event annotation and representation task for Kannada, a morphologically rich resource poor Dravidian language and Hindi, a popular Indo-Aryan language. From a computational perspective, we look into leveraging information from resource rich languages and use transfer learning in order to detect events in a resource poor environment as well. We analyze events in Hindi briefly and Kannada, in depth. As most information retrieval and extraction tasks are resource intensive, very little work has been done on Kannada NLP, with almost no efforts in discourse analysis and dataset creation for representing events or other semantic annotations in the text. In this thesis, we linguistically analyze what constitutes an event in this language, the challenges faced with discourse level annotation and representation due to the rich derivational morphology of the language that allows free word order, numerous multi-word expressions, adverbial participle constructions and constraints on subject-verb relations. Therefore, this is one of the first attempts at a large scale discourse level annotation for Kannada, which can be used for semantic annotation and corpus development for other tasks in the language. On the other hand, from a processing viewpoint, detection of TimeML events in text have traditionally been done on corpora such as TimeBanks. Traditional architectures revolve around highly feature engineered, language specific statistical models. In this thesis, we also present a Language Invariant Neural Event Detection (ALINED) architecture. ALINED uses an aggregation of both sub-word level features as well as lexical and structural information. This is achieved by combining convolution over character embeddings, with recurrent layers over contextual word embeddings. We find that our model extracts relevant features for event span identification without relying on language specific features. We compare the performance of our language invariant model to the current state-of-the-art in English, Spanish, Italian and French. We outperform the F1-score of the state of the art in English by 1.65 points. We achieve F1-scores of 84.96, 80.87 and 74.81 on Spanish, Italian and French respectively which is comparable to the current states of the art for these languages. We also introduce the automatic annotation of events in Hindi and Kannada with an F1-Score of 77.13 and 67.30 respectively.

Abstract

This work presents the idea of using a space-filling fractal curve as the heating element in micro heaters. Since the heating element of a micro-heater is based on metal coils, that will be the costliest component of a micro-heater. However, if we try to increase the spacing between the metal coil, that leads to hotspots and non-uniform distribution. Hence, a cost-effective heater should have an excellent uniform temperature distribution while having minimum metal consumption. Further, non-uniform temperature distribution will give rise to the non-uniform temperature gradient between the environment and body. The geometry of the Heater plays a crucial role in its average temperature (Tµ) and uniformity. This work talks about multiple types and orders of space filling fractal curves as the geometry for heaters, modelled and simulated all of them as heaters and showed that they are on par with the typical double-spiral heat-pad in terms of several criteria like average temperature and temperature uniformity while being cost-effective due to optimal distribution of metal. This work presents modelling of modified Hilbert, Moore and Peano fractal curves as micro-heaters. In order to make a fair comparison among the heaters, a constant power of 1 W was supplied to all the heaters, also the heaters were confined to a fixed area of 30 mm x 30 mm. Finite Element Analysis (FEA) results showed that heating elements based on the geometries of Peano order 3 curve, Hilbert & Moore order 4 curves outperformed double-spiral heating element in terms of temperature uniformity with less metal consumption (cost effective), also all the geometries are modelled with same material, thickness and width. Heaters based on Peano order 3, Hilbert & Moore order 4 curves, showed a non-uniform temperature distribution (𝜎/𝜇) of 2.2%, 4.4% and 4.3% respectively whereas double spiral had a non-uniformity of 5.8%. However, the average temperature (Tµ) was lower by 10 K for Hilbert and Moore order 4 and 50 K for Peano order 3. Further, heaters one fractal order lower than these, i.e., Peano order 2, Hilbert & Moore order 3, have given the average temperature (Tµ) close to double spiral heater while consuming only half the metal that a double spiral takes, thus promising significant reduction in metal consumption. Peano order 2, Hilbert and Moore order 3, showed a a non-uniform temperature distribution of 7.24%, 11.7% and 11.6% respectively, which makes them slightly more non-uniform than the double-spiral Heater. With many fractal patterns to choose from, space-filling fractal curves can provide very good performance while being cost effective as the heating elements in a micro heater.

Abstract

The Cooperative Target Observation (CTO) problem is an optimization problem in which a set of observer agents attempt to maximize the collective time during which each target from among a set of targets, is observed by at least one observer agent within an area of interest. The CTO problem has been well pursued in the multi-agents and robotics literature due to the problem being at the core of a number of applications such as surveillance, inspection, search-and-rescue among others. However, prior works are limited to handling homogeneous target types and the sole criterion of the set of observers is to maximize the observation of targets. In this work, we expand the CTO problem to include heterogeneous target types and constrain observers to incorporate other criteria during decision making while maximizing target observation. To this purpose, we pose Wildlife Monitoring domain as a CTO problem and propose a Multi Criteria Decision Analysis (MCDA) based algorithm named MCDA-CTO, to maximize the observation of different animal species and to effectively handle multiple target types (animal species) and the multiple criteria that arise due to the targets and environmental factors. To demonstrate the working of our algorithm, we present the performance results on two settings: (a) UAV setting i.e., Unmanned Air Vehicles (UAV) are modeled as observers and (b) UGV setting i.e., Unmanned Ground Vehicles (UGV) are modeled as observers. UAVs have uncertainty in observation of targets which makes it challenging to develop a high-quality monitoring strategy while UGVs are affected by the terrain they operate in i.e., UGVs have a different maximum speed in different terrains, thus needing to make suitable trade-offs. We therefore develop monitoring techniques that explicitly take actions to improve belief about the true type of targets being observed for UAV or optimize the ability to observe targets based on terrains for UGV. Furthermore, it is often reasonable to assume that the observers may themselves be a subject of observation by unknown adversaries (poachers). Randomizing the observer’s actions can therefore help to make the target observation strategy less predictable. We provide experimental validation which shows that the techniques we develop provide a higher performance along with better randomization than state of the art approaches.

Abstract

Code-mixing (CM) refers to the mixing of various linguistic units (morphemes, words, modifiers, phrases, clauses and sentences) primarily from two or more participating grammatical systems within a sentence [10]. The development of CM NLP systems has significantly gained importance in recent times due to an upsurge in the usage of CM data by multilingual speakers. However, this proves to be a challenging task due to the complexities created by the presence of multiple languages together. The complexities get further compounded by the inconsistencies present in the raw data on social media and other platforms. In this thesis, we explore methods to efficiently parse the immensely popular Bengali-English CM which is widely spoken in India and Bangladesh. We present a neural stack-based dependency parser for CM data of Bengali-English by utilizing pre-existing resources for closely related Hindi-English CM as well as monolingual treebanks for Bengali, Hindi, and English. To address the issue of scarcity of annotated resources for the Bengali-English CM pair, we present a rule-based system to computationally generate synthetic CM dataset from parallel treebanks of Bengali and English. The generated synthetic CM data does not require an overhead of preprocessing for language identification and normalization as we get orthographic norms from standard corpora. Next, we project automatic annotations from Bengali and Hindi-English treebanks to Bengali-English using simple heuristics creating a synthetic CM treebank for Bengali-English (Syn-BE). Incorporating Syn-BE into our neural stacking parser further improves its performance. Our final model achieves an accuracy of 89.63% for POS tagging and 76.24% UAS and 61.41% LAS points for dependency parsing. This model improves parsing results by 1.37% LAS points when compared with a stacking model utilizing pre-existing Hindi-English and other monolingual resources. We also present a dataset of 500 Bengali-English tweets annotated under Universal Dependencies scheme which can be utilized for evaluation of the system as well as providing seed training data.

Abstract

Neural Machine Translation (NMT), together with multilingual formulations have arisen as the de-facto standard in translating a sentence from a source language to a target language. However, unlike many western languages, the available resources like training data of parallel sentences ortrained models which can be used to build and demonstrate applications in other domains are limited for the languagesin the Indian subcontinent. This work takes a major step towards closing this gap.In this work, we describe the development of state-of-the art translation solutions for 10 Indian lan-guages and English. We do this in four parts described below:1.Considering the Hindi-English language pair we successfully develop an NMT solution for anarrow-domain, demonstrating its application in translating cricket commentary.2.Through heavy data augmentation, we extend the above to the general domain and build a state-of-the art MT system for Hindi-English language pair. Further, We extend to five more languagesby taking advantage of multiway formulations.3.WedemonstratetheapplicationoftheNMTincontributingmoreresourcestothealreadyresource-scarce field, expanding to 11 langauges and its application in a multimodal task of translating a talking face to a target language with lip synchronization.4.Next, we improve both data-situation and performance for machine translation in 11 Indian Lan-guages iteratively to place our models in a standardized, comparable set of metrics setting up forfuture advances in the space to comprehensively evaluate and compare against.