Abstract

RNA is a versatile biopolymer. Different RNAs build up the protein synthesis machinery in living organisms. Non-protein coding RNAs often participate in gene regulation, catalysis of biochemical reactions, and many other cellular processes. The efficiency of functional RNAs depends on their accurate folding into compact three-dimensional structures. RNAs are composed of only four types of nucleotides having similar chemical properties. So, the sources of structural complexity observed in these folded RNAs remain ambiguous. Instead, hydrogen-bonded planar RNA base pairs have the potential to be the fundamental building block of RNA. They display remarkable diversities in their geometries, stabilities, and physicochemical properties. Properties of RNA base pairs are strongly influenced by nucleobase modification, such as post-transcriptional changes, tautomerization, ionization, etc. One such important modification is nucleobase protonation. Protonation at different polar sites of nucleobases are associated with positive free energy change. However, protonationinduced modifications to nucleobases result in subsequent stabilizing interactions, which compensate for the thermodynamic cost of protonation. For example, protonation converts a hydrogen bond acceptor to a hydrogen bond donor and therefore opens up new base-pairing opportunities through the protonated edge of the nucleobase (Class I). In other cases, the protonation-induced changes may influence the stability of the base pairs formed through the unprotonated edges of the base (Class II). Instances of Class II protonation in RNA bases is nontrivial to detect for both experiments and in-silico techniques, as the loaded proton does not form any inter-base hydrogen bonds. Hence, the contribution of Class II protonation in determining RNA’s structure and function remains majorly unexplored. Recently a strategy has been proposed to study the consequences of Class II protonation on the geometry and stability of RNA base pairs (1). It involves comparing the electronic structure properties of the different base pairs optimized with and without the Class II protonation. In the past, this method revealed that Class II protonation at the N7 and N3 positions of guanine stabilizes different RNA base pairs, which are intrinsically unstable. In this work, the same strategy is applied to explore the consequences of Class II protonation at the N3 position of guanine and N3 & N7 positions of adenine on the geometry and stability of different RNA base pairs. Those are the sites where the thermodynamic cost of protonation is higher than other sites. Analysis of RNA crystal structure datasets and electronic structure calculations carried out at DFT, and DFT-D levels reveal a number of interesting insights. In general, it is observed that charge redistribution caused by Class IIprotonation alters the hydrogen bonding potential at the polar sites located on the unprotonated edges of nucleobases. Such changes collectively determine the geometry and stability of the base pairs formed through the unprotonated edges. A set of RNA base pairs are reported which are sensitive to Class II type nucleobase protonation and hence have putative biotechnological application in pH sensing.

Abstract

Summarization involves reducing the size of the document without the loss of important content. The reduced content is called the summary of a document. Summarization can be divided into two broad categories: Extractive and Abstractive. In Extractive Summarization, the summary of a document is constructed by extracting phrases and sentences from the document itself. In this work, we are mainly concerned with single document extractive text summarization and employing deep-neural-networkbased solutions for the same. While on the surface, extractive summarization seems like a trivial problem since summarizing here involves copying content from the source document, in reality, it is far from trivial. One main reason for the complexity of the task is the absence of gold labels for what sentences/phrases to extract. Typically, extractive summarization systems are evaluated using human-written summaries. However, this summary cannot be directly used to train the neural networks. Indirect methods of labeling or training are commonly used in many of these cases. Extractive Summarization can be divided into two steps: Document Understanding and Document pruning. A document is not simply a bunch of sentences put together. There is an underlying structure between the way sentences are arranged, which makes a document coherent and meaningful. Moreover, summarisation can be said to utilize this exact structure (among other things) to determine which sentences are more important and relevant. Most works assume that the structure of the document is linear and use recursive networks to encode them. But it was shown by Rhetorical Structure Theory (RST) that documents can be represented by tree-like structures. One of the first things we explore is to see if we can use non-linear tree-based networks to encode the document. Moreover, since learning to summarize involves learning to understand documents implicitly, we see if we can try to capture such structures implicitly when learning to summarize. We also explore if such structures can aid in summarisation. Another reason for the complexity of the task is that extractive summarization is different from sentence ranking. Sentence ranking can be defined as ranking sentences based on how good a sentence is for representing the document. However, in summarisation, the selection of each sentence also depends on the other sentences selected. This makes the task very complex due to the exponential number of possibilities in selecting a subset of sentences to represent the document. Each of these possible subsets of sentences can be called a candidate summary of the document. A document of length 25 has over two thousand such candidate summaries, only when considering those of length 2 and 3 sentences. This creates a combinatorial explosion problem that we should overcome by usnig indirect means. Many works deal with this combinatorial problem in their own way. Here we also explore different novel and unique methods for doing the same. We introduce a new pipelined architecture for extractive summarization, which helps in dealing with precisely this problem. We also see how we can adopt different complex reinforcement learning (RL) formulations to extractive summarisation. Many works use one of the two formulations in training their models: either policy-gradient-based algorithms or deep-Q-networks. In this work, we see how we can improve both of them. We also experiment with neural-network-based loss functions for summarisation. At last we see the problems that can arise in training extractive summarisation models when we try to circumvent the combinatorial explosion problem. Many works try to simplify their training objective by approximating summarisation as sentence classification, or use very limited set of samples to train their models. This creates a lot of issues in what the model actually learns. We show this by experiments and analysis, and also provide a glimpse at a solution for solving the same.

Abstract

This thesis revolves around two questions. How have the ideological positions of political parties in India evolved in the 21st century and what changes are these bringing to the participation of people in democratic processes? Literature on ideology and participation are discussed in the initial part of this thesis. It uses election manifestos published by political parties for the elections from 2004 to 2019 to identify the ideology of political parties. A manifesto coding scheme adapted to Indian conditions was used to categorize manifesto points to different heads based on ideological leaning. The evaluation suggested a shift towards capital and private entrepreneurship, increasing focus on urban areas, changing attitude towards decentralisation depending on the power relations. Political parties showed a left leaning view on social affairs while positions varied from central right to left on economic and political affairs. The political parties are becoming more inclusive of different sections of the country though the trend is not consistent. To identify participation, the focus was given to Muslim participation. Detailed election results published by the Election Commission of India from 1977 to 2019 was used to identify participation in elections. Muslim names were identified using a code. geographical distribution was mapped using QGIS. Muslim participation was increasing consistently from 1991 to 2014 before dropping in 2019. The participation is largely through recognized political parties than as minor groups or independents. Muslim identity based political parties and Islamic parties are trying to spread the influence geography wise and vote wise. But they remain minor players in relation to Muslim participation where recognized political parties which are not based on Muslim identity fields more number of candidates.

Abstract

Research in Natural Language Processing is expanding in multiple languages and is seeping into all domains of life with time. The variety of text which can be processed is growing with every advancement. Online social media platforms have become a rich source of textual data for several NLP tasks. One such social media platform is Twitter. In the present day, tweets on Twitter have become vital for understanding humans. They should be analyzed to obtain more relevant information, such as sentiments and emotions. On the other hand, the massive growth in communications on these sites by users across the world has its ill effects. Many individuals are taking undue advantage of these platforms to post aggressive and hateful content on other individuals and groups. The amount of content that resembles this form has increased beyond the scope of manual filtering for any research group/organization. Hence, the availability of automatic identification systems and annotated datasets in this domain is of utmost importance. From many years, a phenomenon called “code-mixing” has attracted a lot of research and interest from sociolinguists. According to Wikipedia, Telugu is the fifteenth most-spoken language in the world. Telugu is observed to be mostly written in Roman script on social media platforms. Code-mixed Telugu-English content is growing digitally by people residing across the globe. This thesis focuses on detecting two elements, i.e., aggression, and hate speech from code-mixed Telugu-English online posts, which are two different text classification problems. Currently, there are no publicly available datasets in Telugu-English for these tasks. In our attempt to create data resources for these tasks, this work presents a corpus of 3677 and 3361 tweets annotated with aggression and hate speech, respectively. These datasets are of great value as all annotations are done manually by two annotators, which makes them very rich for the research community. In this thesis, we describe the creation and annotation process, data, and the possible uses of the datasets. We have also built a baseline classification system to identify hate speech and aggression using the corpus created. The annotations have been used for training and testing the classification models. This work gave us an opportunity to work on code-mixed Telugu-English data, which is a relatively less explored field.

Abstract

This thesis revolves around a rather simple question - Is Sport Art or Not? We discuss some of the scholarly works that debate around this question before we charter our own course of answering this question. We engage with this discourse by keeping Sport at the centre of our analysis. We discuss the Idea of Sport and the concepts that are central to it before analysing the structure of Sport.To analyse the Idea of Sport, we study the historical transformation of Play and how it degraded to be something else. We also analyse social structures that replicate the structure of a Game to further understand the position of Sport relative to Play and Game. For analysing the structure, we selected the following sports - Track, Hurdles, Shot Put, Long Jump, Bowling, Weightlifting, Pacman, Badminton. These sports were selected as they display a serial logic in how their events are arranged. We construct the graphical representation of these sports by listing out the events and the rules that render the events their serial causality.We propose a method of construction grammars for serializable sports and provide a prototype of a classification of sports based on the grammar generated. Based on our observations of the above mentioned exercise, we arrive at a conclusion that Play leads to Art and that Sport is not Art.

Abstract

There is an explosion of information in the digital world. In a crowded digital world, recommender systems are a prerequisite for any digital media platform that serves content to users of varied interests. Platforms like Netflix, Amazon, Youtube have leveraged and contributed to the extensive work in recommender systems and are continuously improving the status quo. In the last decade, another kind of content-based platform has emerged – the EdTech kind. Though having the same need, to serve relevant and engaging content to its users, this domain however has very distinct nuances that make the traditional recommendation systems and the theory, metrics and impact factors around it misfit for its use case. For example, it makes sense for youtube to recommend a jingle that a user keeps listening to, it makes sense for netflix to recommend an all time favourite episode from a sitcom. Numerous experiments have shown that this infact helps to boost the users engagement on the platform. This approach does not work for EdTech where the intention is to have the user consume optimal content for knowledge gain. Repeat recommendations will in fact fatigue the user and drive him/ her off the platform. Another key nuance that we make a central theme in the thesis is the concept of ‘knowledge gain’ and posit that recommender systems for eLearning have to optimise this as a parameter. This is something that is unique to eLearning only and needs a careful thought. Overspecialization in the recommender systems is a big problem for edtech as it inhibits the users learning curve and ultimately reflects poorly on the edtech platforms catalogue and user journey. Thus, a healthy amount of knowledge gain injected by appropriate recommendations goes a long way in the eLearning domain. The limitations of the current ecosystem propelled us into an inquiry on the exact nature of these gaps. Building on, we propose a content based recommendation algorithm - ‘KNN-RkNN’. Inspired by the theory of active learning, our ‘KNN-RkNN’ (K Nearest Neighbour - Reverse K nearest neighbour) algorithm is optimised for information gain that’s learnt from the users history. The data was sourced from a popular EdTech platform, Unacademy. The effectiveness of this algorithm was measured against relevance, diversity and NDCG (Normalised discounted cumulative gain). This algorithm was tuned in perspective of optimising for knowledge gain. We show that taking informativeness as a feature, independent of the diversity and relevance, can lead to an unbiased exploration of good content in a large content repository. This thinking is a new line of thought in edtech and paves way for further research on development of recommendation algorithms primed on informativeness as a feature.We also research on building an enhanced sequential collaborative filtering technique that learns from the users watch history and recommends a learning path/playlist. This recommendation engine consists of a seq-2-seq model with multiple LSTM layers with attention mechanism. The results of this experiment are encouraging. An accuracy above 90% is achieved when the number of classes to recommend are reduced. Different experiments with similar architectures yielded different results which have been documented in this study. The chief contribution of this research is to qualitatively establish how knowledge gain when considered as a objective for eLearning is a nuance that was missing in eLearning thus far. We also quantitatively establish this via multiple case studies. Byproduct of these studies is to demonstrate how knowledge gain can be incorporated in both content based and collaborative filtering techniques. A more subtle contribution of our work is its lab to market nature. The sequential collaborative learning algorithm explained in the thesis is industry ready both for performance and scale. This research hopes to serve as a base for future studies to tune knowledge gain not just in content but also in other eLearning aspects like polls, tests, discussions etc.

Abstract

In drug discovery applications, high throughput virtual screening exercises are routinely performed to determine an initial set of candidate molecules referred to as “hits”. In such an experiment, each molecule from large small-molecule drug library are evaluated for physical properties such as the LogP, Topological Polar Surface Area (TPSA), and docking score against a target receptor. In real-life drug discovery experiments, the drug libraries are extremely large but still a minor representation of the essentially infinite chemical space , and evaluation of physical property for each molecule in the library is not computationally feasible. In the current study, a novel Machine learning framework for Enhanced MolEcular Screening (“MEMES”) based on Bayesian optimization is proposed for efficient sampling of chemical space. The proposed framework is demonstrated to identify 90% of top-1000 molecules from a molecular library of size about 100 million, while calculating the docking score only for about 6% of the complete library. We believe that such a framework would tremendously help to reduce the computational effort in not only drug-discovery but also areas that require such high-throughput experiments.

Abstract

Today’s society cannot be visualized without the dependency of humanity on technology. Every walk of life in human survival today is linked to computers and electronic systems, which undoubtedly improve the quality of life. This comfort comes with a restriction because of the vulnerability of the electronic system to anti-electronic threats. Not orienting towards offensive, it is indeed the need of the hour to protect our systems by characterizing their susceptibility levels. Many researchers formulate the threat scenarios, and international committees constituted for this purpose. High power electromagnetic (HPEM) systems that could generate and radiate few gigawatts were developed worldwide. Likewise, an antenna also plays a crucial role in the HPEM systems in any transmitting system. However, there are more challenges in designing and developing antennas for HPEM systems. Among the categories of antennas developed for this purpose, wideband/Ultra-wideband (WB/UWB) antennas that are capable of transmitting transient impulse-like high voltage signals are potential contenders. A well-designed WB/UWB antenna should have high fidelity in transmitting the transient signal without altering its characteristic parameters. This thesis focuses on the design, analysis, development, and measurements of a wideband antenna, spark gap integrated dipole (SGID) antenna. The main focus is placed on transient analysis, its construction, and performance evaluation. Spark gap integrated dipole (SGID) are simple in construction, compact, and reliable transient wideband antennas for high power applications. They offer the promise of utilizing them to very high powers without any complications and can easily be fabricated with low-cost materials. This makes them a viable alternative to complex horn-based antennas that are generally used in high-power scenarios. Despite the aforementioned advantages, the bandwidth offered by these antennas is limited to mesoband, whereas Transverse Electromagnetic (TEM) based horn antennas can scale up to hyper bands. As a result, their use is limited to Directed Energy Weapon (DEW) applications that demand damped sinusoidal radiation and, to some extent, in UWB radars. While horn-based antennas have matured technology, both experimentally and theoretically, only a few researchers have reported the analysis of spark gap integrated dipole antennas. In this thesis, we mathematically modelled the SGID antenna based on Maxwell’s equations, and a theoretical transient response is generated. The performance of SGID is thoroughly analysed using computational electromagnetics for its transient response. The physical dimensions of the dipole are optimized for maximum radiation efficiency. The spark gap efficiency was also studied by evaluating the performance for various electrode profiles. Modifications are proposed to the existing designs in order to enhance the performance and increase the figure of merit. This includes optimal dimensions visà-vis input signal parameters which help not only for the maximum radiation but also its spectral occupancy. In addition to this, the selection of Bruce’s profile has ensured uniform electric field distribution and has allowed using the dipole for higher voltages. To further enhance the reliability of the antenna, hydrogen gas is selected as the spark gap medium due to its high recovery rate property. These proposed methods are mathematically formulated, simulated, and analyzed, and a considerable improvement in the overall SGID performance is reported. The theoretical analysis is validated by measurements by subjecting the antenna to a 100kV pulsed input voltage. A logical measurement procedure is proposed that removes the ambiguity in measurements as the radiation from the spark gap antenna is highly probabilistic. The proposed design and development methodology of the SGID has resulted in an efficient transient radiator with a figure of merit (FoM) significant compared to those reported in the literature.

Abstract

There are several types of spatio-temporal dynamic processes, both naturally occurring and manmade like climate change, floods or droughts, crime-spread etc; and some which are influenced by multiple factors like agricultural crop performance, occurrence of diseases, etc. It is important to understand the behaviour of such phenomena as they have a significant effect on the places where they occur, having both direct and indirect impacts on the environment as well as the population in that region. These phenomenon have an intrinsic spatial as well as temporal context associated with them. In order to gain insights into these processes and their relationships, there is a need for a methodology that can capture the context in both the dimensions- spatial and temporal, simultaneously. At a particular location and time, the observed output value of a dynamic phenomenon can be affected by many causative factors that feed as an input to the process and are at play simultaneously. While many studies look at the correlations of input to output to understand the phenomena, it’s also important to understand the outcome exhibited with its intrinsic spatio-temporal context, deciphering which may help us better appreciate the factors at play, both known and unknown dynamics of the factors themselves. Hence there is a need to highlight these patterns as seen in the observed data and provide insights into the process that govern these phenomena. Different spatio-temporal statistical modelling methods have been developed, that try to approximate the process and produce a mathematically-formalized model that would generate the spatio-temporal data. However, such modelling techniques are more focused on accurate estimation of input parameters and finding a suitable model that fits the data. They often end up ignoring the spatially correlated parameters, or parameters with a weaker influence on the phenomena, thus affecting the overall predictive performance of the model. Alternatively, data mining approaches are focused on uncovering and describing unknown patterns and trends in the data. However, in the case of spatio-temporal datasets, the direct application of classical data mining algorithms is not suitable, since they are intended to work on transactional-like data which has independent attributes. The attributes in spatial data are not independent, and there is an implicit interaction between them, which cannot be captured directly by the classical data mining techniques. Therefore, the conventional data mining methods fail to effectively capture the implicit spatial or temporal characteristics of the geographic phenomenon and are unable to exploit the rich spatial temporal relationships/patterns embedded in the datasets. In this thesis, we make use of a generic Spatio-Temporal data mining algorithm called MiSTIC, as part of a generalized framework called MiSTIC++. Using MiSTIC, the goal of the framework is to segregates the entire area of analysis into different zones, where each zone is comprised of contiguous spatial samples. The method tried to create the zones in a way such that within each zone, the spatial interactions among its comprising elements remains preserved over time, thereby capturing the consistent spatio-temporal context, which cannot be observed directly from the data. These spatial interaction exist between two neighboring regions, the directionality of which remains the same across time, despite the changes in magnitude. There might be several input factors that are responsible for bringing a net change in the observed value. However, in spite of the various factors that are at play, the interactions between these individual spatial units remains very similar across time. The MiSTIC algorithm tries to capture this property of preserved spatial interactions, based on only the observed output data and the neighborhood information. The knowledge of such stable zones and their intra-zonal analysis helps understand the dynamic phenomena better, enabling us to respond better to the situations and to take better decisions with respect to the process being analyzed. The utility and generality of the proposed method has been demonstrated with the help a case study of crop yields in India. The study uses the district level yield dataset of Paddy and Maize in Kharif season across India for the period 2000 to 2010. The method helps cluster a set of districts into zones, with each zone further divided into nearly concentric sub-zones of similar performers based on the empirical rule. The results for paddy in the Kharif season show that there are 18 production zones across India and 23 for Maize, except the districts of North East and Jammu and Kashmir, with zonal averages ranging from 861.9 kg/ha to 3076.3 kg/ha for Paddy. MiSTIC is further extended, in this research, to indicate regions and sub-regions that are consistently high and low performers across the years, irrespective of the input simila

Abstract

Language is the most important medium of communication and education. In a multi-lingual country like India where there are 22 scheduled languages, it is a matter of national policy to endorse and promote the development of all these languages and reduce the language barrier. Giving universal access to information to people in their language is the primary concern and need of society. In this technological revolution, access to all information and communication technology (ICT) and Human language technology (HLT) will make it convenient for people to gain all the benefits of this. Though there are several NLP Tools and approaches available, due to the scarcity of good quality benchmark data, the development of practical usable systems is hampered. Machine Translation (MT) systems play a major role in bridging the language gap. However, there are several pit-falls in the MT due to various linguistic, cultural, and social reasons. Hence, there is a need for an Accessor system where there is a guarantee of faithful translation. In this thesis, we have proposed and implemented an architecture to develop a usable system called ’Anusaaraka.’ The use of an expert system to develop this Accessor + MT system is unique and helpful for the common man to contribute to further improvement of the system. As discussed without proper quality linguistic resources, no usable NLP applications can be developed. To address this issue, we have proposed a unique concept of ’mutual-bootstrapping’ where various linguistic/NLP tools interact with each other to further enhance/boot-strap the development of highquality linguistic resources. We also demonstrated the proposed model with the task of ’word/phrase alignment’ of a parallel corpus by building an ’Alignment Tool’. Methods to build high-quality linguistic resources like bilingual dictionaries, multiword expression dictionaries, treebank, domain dictionaries, technical dictionaries, etc. through mutual interactions by the Alignment Tool are discussed. Neural network based Machine Translation (NMT) systems are now producing state-of-the-art results. But still, there are errors and idiosyncrasies in the output. We have applied the ’mutual-bootstrapping’ approach to improve the accuracy of the NMT. Through direct pre-processing and post-processing techniques using the ’Alignment tool,’ the boot-strapping of NMT to produce high-quality translations will be discussed in detail. Throughout the thesis, we have emphasized the concept of ’mutual-bootstrapping’ in developing various NLP Tools/resources. The results of the tools are compared with other state-of-the-art results and the error analysis is done