Abstract

Data scientists very often find that a central step in their work, is to implement an appropriate transformation restructuring the originally given data into a new and more revealing form. Although specific domain knowledge can be used to help design representations, data-driven learning with generic priors can also be used, and the quest for AI is motivating the design of more powerful representation-learning algorithms implementing such priors. In order to make it possible to apply machine learning to different domains, it is very important to make learning algorithms less dependent on manual feature engineering, so that novel AI based applications could be developed faster. The work in this thesis focuses not only on the mapping from representation to output but also the representation itself by mining hidden patterns from available data. Keeping this in mind, we propose a learning model to alleviate difficulties involved in feature engineering through automation. We develop a learning model, based on regression between feature pairs, that discovers underlying patterns and its variations in the data, by the way features are related to each other and selects a very small number of new features to create a significant improvement in predictive performance. Because this model takes into account the inherent feature structures, we took our next motivating example from bioinformatics. The features in this domain have some natural spatial order (for instance, in the medical domain, genes are often associated with different types of clinical features) and thus incorporating such structure can help select more important features and achieve more accurate classification accuracy. It also provides better readability and interpretability to the models. The final component of the work focuses on the problem of fake news detection. The low cost, easy access and rapid circulation of information over the internet has encouraged more people than ever, to seek out and consume news from online sources or social media rather than traditional news organizations. While the traditional (count-based) feature engineering strategies for textual data are effective methods for extracting features from text, due to the inherent nature of the model being just a bag of unstructured words, we lose additional information like the semantics, structure, sequence and context around nearby words in each text document. This formed as a motivation for us to explore more sophisticated models which can capture this information and give us features which are vector representation of words. Keeping in mind that fake news is intentionally written to mislead readers to believe false information, it can be difficult to detect it, solely based on news content. Therefore, we need to include auxiliary information, such as user social engagements on social media, to help make adetermination. We therefore design, a novel attention based hybrid network which integrates the article information along with the meta data and claims of the news articles.

Abstract

Autonomous Driving is one of the hot areas of research in the current world. Deep Machine Learning and Computer Vision together have shown promising results in this field. A relatively new domain is Reinforcement Learning. RL is associated with behavior based learning. An RL agent learns from experiences, by exploring it action and state space. In this work, a Deep Reinforcement Learning algorithm, Deep Deterministic Policy Gradient(DDPG) is used to learn various driving behaviors. DDPG is an off-policy, continuous control framework. It is the off-policy nature of this algorithm which enables it to efficiently explore the environment. Curriculum learning and Intrinsic Motivation are used to extract advantage of the off-policy nature. Different behavior based agents are trained. The important ones being Overtaking in highways and Opportunistic agent in dense unstructured traffic scenes. For overtaking in highways, the agent learns in a Curriculum based approach. First, it learns to drive on an empty road and next it learns overtaking strategies. The reward function is handcrafted in a manner that the desired behavior is learned in least number of training episodes. Various experiments were conducted to handcraft the reward, a detailed analysis of which have been provided. The learned agent is able to overtake not only on straight but curved roads as well. Another behavior that was learned was blocking. Blocking is an hostile behavior and should not be practiced in real life. By slightly changing the reward function for Overtaking on highways and by positioning the learning agent ahead of the opponent, blocking behavior was learned. The approach for overtaking and blocking is compared with existing RL based approach for the same. For dense traffic scenes, we learn two behaviors: Opportunistic and Defensive. Opportunistic agent actively looks for free space ahead of it and navigates itself there, while defensive stays in its own lane and changes its speed each time any vehicle approaches it, to avoid collisions. There is no prior work which deals with end-to-end driving in dense unstructured traffic. Learning the behavior for defensive agent was not easy to achieve by simply using noise. We used intrinsic motivation based approach by explicitly showing it the rewarding actions. For all of the behaviors the learning is scalable and robust in terms of the speeds of opponent vehicles and the number of cars in the surroundings.

Abstract

Humour is an intricate part of human behaviour and a valued quality when developing interpersonal relationships. It is an interesting and challenging domain to work in owing to its subjectivity. Humour is one of the most sophisticated forms of human intelligence and creativity, intimately related with factors such as world knowledge, reasoning and perception. With our AI systems becoming increasingly intelligent, there is now a demand for systems which are more human-like than ever before. No AI system could be complete without the ability to to process and use an integral human phenomenon like humour. Humour is often expressed through language in various forms such as funny stories/anecdotes, jokes, witty sayings, puns etc. This verbally expressed humour is of interest to NLP researchers because it deals with spoken and textual language. The subfield of Computational Linguistics and Artificial Intelligence which aims to develop systems which can use and understand humour like humans do is called Computational Humour. A complete computational humour system has to contain two main modules - a humour interpreter, and a humour generator. Both these modules could have their independent uses in a variety of applications, but when it comes to interactive and comprehensive AI systems, both are equally important. In this thesis, we present a first ever work for computational humour for Hindi. We consider both sub-fields in computational humour - Humour Understanding and Humour Generation, and work on specific applications in both. We study Hindi-English code mixed puns by focusing on automatic recovery of the target words/phrases in them. We define code-mixed puns, analyse the type of structures present in them, and classify them into two main categories - intra-sentential and intra-word code-mixed puns. We then propose a four-step methodology to recover intra-sentential code-mixed puns. On evaluating our model on a test dataset of such puns, we observe that we are able to successfully recover 67%. We recognise several limitations of our system, like the inability to deal with highly unusual and creative language use. For the problem of computational generation of humour, we consider two tasks. First, we build a prototype system which automatically generates a simple form of Dur-se-Dekha jokes in Hindi using a small lexicon and handcrafted rules. Second, we design an algorithm which can induce humour in a short non-humorous text by performing single word subsitutions on the basis of phonetic similarity and sentiment opposition. We perform human evaluation on the results of both and observe that we are successfully able to generate funny instances in both these cases.

Abstract

The major contribution of this thesis is creation and enrichment of resources (a lexicon) for a resource poor language viz. Telugu. This thesis also describes the enrichment of OntoSenseNet-a verb-centric lexical resource. Our work aims to preserve and enhance the usage of an authentic Telugu dictionary by developing a computational version of the same. This enables the native speakers of the language to actively involve in the research as most of the computer science experts or algorithms work on top of the annotated language data. With the aid of developed Telugu dictionary, native speakers can perform better annotations as both the word and its meaning are in a language they are familiar with. Hence, efforts are made to develop the aforementioned Telugu lexical resource and numerous annotations are done manually by language experts. Primarily, we attempted two types of annotations 1) Ontological classification of verbs, adverbs and adjectives; 2) Annotation of unigrams and bigrams for sentiment polarity. Based on the proposed ontological classification, the manually annotated gold standard corpus consists of 8483 unique verbs and 253 unique adverbs. Annotations are done by native language speakers according to the set of provided annotation guidelines. We discuss annotation procedure in detail and present the validation of the developed resource through inter-annotator agreement as a measure. Additional words extracted from Telugu WordNet are combined with our resource and annotations are done. Furthermore, we discuss the enrichment of this manually developed resource of Telugu lexicon, OntoSenseNet. OntoSenseNet is a ontological-sense annotated lexicon that classifies each verb into 7 sense-types and adverbs into 4 sense-classes. The developed OntoSenseNet for Telugu has primary and secondary sense-types identified for the verbs and primary sense-class tag for adverbs. The area of research is relatively recent but has a large scope of development. We provide an introductory work to enrich the OntoSenseNet to promote further research in Telugu. Classifiers are adopted to learn the sense-type of the words in the resource and thus, we automate the tagging of sense-types for verbs. We perform a comparative analysis of different classifiers applied on OntoSenseNet. The results of the experiment prove that automated enrichment of the resource is effective using SVM classifiers and Adaboost ensemble. However, the accuracy is low compared to the task of manual annotations. To perform manual annotations more extensively, we have developed a tool for crowd-sourcing the task of annotating. Access to contribute to the resource is given only to certified individuals after preliminary assessment.This tool consists of guidelines for annotations and list of words that are to be annotated and the annotator is given the freedom to choose “uncertain” option in case of an unclear judgment. Mechanisms adopted to minimize the disagreements and measures are taken while adding these annotations to our resource are discussed in this thesis in detail. Additionally, we discuss the potential applications of this ontological resource. Moreover, efforts have been put to enhance the sentiment analysis task through phrase-level annotations. We developed a systematically annotated corpus that can support the enhancement of sentiment analysis tasks in Telugu by annotating bigrams in Telugu. These are the second kind of annotations that are mentioned before. The developed polarity-annotated corpus is called ‘BCSAT’. From the developed Telugu dictionary, we extracted 11,000 adjectives, 253 adverbs, 8483 verbs. We extracted words from SentiWordNet and bigrams from target corpus. Sentiment based polarity annotations for these extracted lexemes are done by language experts. We discuss the methodology followed for the polarity annotations and provide validation for the developed resource. The fundamental aim is to validate and study the possibility of utilizing phrase-level sentiment annotations in the task of automated sentiment identification. This work aims at developing a benchmark corpus, as an extension to SentiWordNet, and baseline accuracy for a model where phrase-level sentiment annotations are applied for sentiment predictions. The method we present outperforms all known methods when tested on the recognized and standard benchmarks for sentiment analysis task in Telugu.

Abstract

Text articles propagating false news have lately become annoying for the Internet users. These articles circulate quickly in the social media platforms and make it difficult for users to bifurcate actuality from fiction. Various social media platforms and instant messaging applications have become the source of these kinds of text articles. Since a large audience is active on these platforms, it becomes easy to propagate media content among the active users. These text articles range from news content belonging to international news to hyperlocal journalism which is propagated for financial or political gains for some organization or community. Users on the Internet easily believe any kind of shared news on such social media platforms. Very few try to verify and then share further these news articles. But many tend to blatantly accept and proceed further to share on various social media and instant messaging platforms because most of the time false news tend to be more interesting with misleading facts. Lately, the scientific community has shown a keen interest in finding better computational methods to segregate such misleading facts from a lot of news articles. Past work on computational credibility analysis has concentrated on fact examination and linguistic features. The primary focus of the credibility analysis task is to learn features which differentiate the false claims from the true claims by finding the conflicting viewpoints. This task of news verification and scoring have been attempted in various ways ranging from traditional NLP methods to machine learning approaches. This work describes a novel architecture and methodology of Credibility Outcome (CREDO) which processes a given text article and gives the trustworthiness score of a textual claim in an open domain setting. CREDO constitutes of various modules which capture different characteristics responsible for the credibility of the textual claim. These characteristics include reliability of the article’s source and writer, semantic comparability between the article and related dependable articles recovered from an information base and the opinion expressed in the textual claim. All these characteristics are learned from various modules and the scores are fed into a neural network architecture which learns the features from these modules pertaining to the overall credibility of the textual claim. The noisy text is also attempted to normalize by using a novel normalizing approach using distributed semantics and ranking approaches to find normalized word during the preprocessing of text, in case the system receive noisy text article as input. Various experiments are conducted on a popular dataset which exhibits that CREDO outruns the state-of-the-art approaches based on linguistic features

Abstract

Micro-blogging sites such as Twitter, Facebook have gained tremendous popularity in the last decade, and have allowed the users to freely write content on the Internet for the purpose of sharing, providing and using the information. They encourage users to express their daily thoughts in real time, which often results in millions of emotional statements being posted online, everyday. This huge amount of user-generated data has resulted in the emergence of a research community that aims to mine social media content and evaluate various linguistic properties associated with the text, such as emotion prediction, irony detection, hate speech detection, gender prediction, sarcasm detection, fake news detection, troll detection etc. However, the writing style of users on microblogging sites tends to be quite colloquial and non-standard, different from the style found in more traditional, edited genre. Authors from multilingual societies tend to write code-mixed posts frequently on social media which results in emotions expressed in text due to the mixture of different natural languages. Code-Mixing (CM) is a natural phenomenon of embedding linguistic units such as phrases, words or morphemes of one language into an utterance of another. The presence of huge amount of code-mixed data on social media has attracted researchers to understand and study the code-mixed texts. While some work has been done on code-mixed social media text and in emotion prediction and irony detection separately, our work is the first attempt which aims at identifying the emotion and detecting the irony associated with Hindi-English code-mixed social media text. In this thesis, we analyze the problem of emotion identification and irony detection in code-mixed content. We present a corpus of Hindi-English code-mixed tweets for Emotion Prediction and Irony Detection. Corpus for Emotion Prediction is annotated with the associated emotion and also the causal language of the expressed emotion. We annotated the corpus for Irony Detection with the labels ironic or non-ironic. For every tweet in the datasets, we annotate the source language of all the words present. Finally, we propose a supervised classification system which uses various machine learning techniques for prediction of emotion and detecting the irony associated with the text using a variety of character level, word level, and lexicon based features. We evaluate our systems on the presented datasets and carry out 10-fold cross-validation.

Abstract

Quantum superposition and entanglement (a special case of superposition) play a pivotal role in achieving information processing tasks which could otherwise be not possible through any classical resource. It is further interesting to note that these very properties constrict us from performing some tasks which are otherwise achievable classically. As quantum entanglement has proven to be a key resource for various information-theoretic tasks, it’s numerous applications makes it imperative to study it’s behaviour. Considerable studies have been done on how to generate entanglement, the chief of which is by using global unitary operations. The thesis further examines the limits imposed on quantum information theory due to the probability structure of mixed states. In mixed states the probability occurs at two levels, first it starts as a description of the system as a statistical mixture, second as the intrinsic superposition that is contained by each entity of the ensemble. We witness how the global unitaries, that can make any pure separable state entangled, fail to generate entanglement for a class of mixed states. The discussion begins with the definition of separable states and non-local states. It is then followed by a concise study on the absoluteness of the above mentioned class of states. Following which is an extensive study on absoluteness of non-negative conditional von Neumann entropy, which happens to be the decisive part of the second chapter. The thesis concludes by discussing another limit that superposition imposes on quantum theory. The impossibility to clone coherence of an arbitrary state. It also addresses the logical implication of no-cloning of coherence which is the famous no-cloning theorem.

Abstract

Developmental dyslexia, considered genetic in origin, is associated with unexpected difficulties in reading, writing, and comprehension, despite normal intelligence and socioeconomic status. Its nature and prevalence rate differ across languages. These differences gave rise to many debates about its universality and specificity. Additionally, prevalence estimate of dyslexia in different countries seems to be related to the shallowness of the orthography. Telugu is an alpha-syllabic/ Akshara-based language with a one-to-one mapping between the grapheme and its constituent phoneme. Moreover, on the granularity-transparency scale, Telugu orthography falls on the coarse-transparent level. In Telugu, phoneme is not the smallest grain size nor the reading instructions phoneme based as in English. In that case one can ask how a phoneme deficit could underlie reading difficulties in Telugu. Thus we hypothesized that reading difficulties of Telugu native speakers would be perceptual in nature rather than phonological. This thesis aims to address the reading difficulties of Telugu native speakers, specifically perceptual and phonological deficits. Further, appropriate intervention methods are devised in this thesis to address specific deficits. We conducted a set of perceptual and phonological tests on age-matched groups (dyslexic and non-dyslexic) and observed significant group differences on perceptual tests but not on the phonological tests. The results are in line with those obtained on other transparent languages. We propose a bilingual reading model for Telugu and other Akshara languages. The qualitative analysis of the reading errors indicates that there is impairment of the visual/letter detector components that lead to peripheral dyslexia and at the sub-lexical route causing the vowel-letter dyslexia among Telugu native speakers. These errors are also related to visual and auditory attention deficits. Consequently, we designed interventions that could assist in improving visual and auditory skills. The contributions of the thesis could provide the foundation for devising more accurate diagnostic protocols and offer effective interventions for children with reading difficulties.

Abstract

Human brain is undoubtedly the most magnificent yet delicate arrangement of tissues. This serves as the seat of such a wide span of cognitive functions and behaviors. The neuronal activities within every neuron, collectively observed over network(s) of these interconnected neurons, manifest themselves into patterns at multiple scales of observations. Many brain imaging techniques such as fMRI, EEG, MEG etc. measure these patterns as electro-magnetic responses. These patterns supposedly play the role of unique neuronal signatures of the vast repertoire of cognitive functions. Experimentally, it is observed that different neuronal populations participate coherently to generate a signature for a cognitive function. These signatures could be investigated at the micro-scale corresponding to responses of individual neurons to external-current stimuli, at the meso-scale related to populations of neurons that show similar metabolic activities and in turn these populations, also known as regions of interest (ROIs), communicate via complex arrangement of anatomical fiber pathways leading to signatures at the macro-scale. The holy grail of neuroscience is thus to computationally decipher the interplay of this complex anatomical network and the complex functional patterns corresponding to the cognitive behaviors at various scales/levels. Each scale of observation, depending on the instruments of measurement, has its own rich spatiotemporal dynamics that interacts with higher and lower levels in complex ways. Large-scale anatomical fiber pathways are represented in a matrix that accounts for inter-population fiber strength known as structural connectivity (SC) matrix. One of the popular modalities to capture large-scale functional dynamics is resting-state fMRI, and statistical dependence between these inter-population BOLD signals is captured in functional connectivity (FC) matrix. There are many models that provide computational accounts for the relationship between these two matrices as deciphering this relationship will provide the mechanism by which cognitive functions arise over the structure. On one hand, there are many non-linear dynamical models that describe the biological phenomenon well but are expensive and intractable. On the other hand there are linear models that compromise on the biological richness but are analytically feasible. This thesis is concerned with the analysis of the temporal dynamics of observed resting-state fMRI signals over the large-scale human cortex. We provide a model that has a bio-physical explanation as well as an analytical expression for FC given SC. Reaction-diffusion systems provide a computational framework for the emergence of excitatory-inhibitory activities at the populations as reactions and their interactions as diffusion over space and time. The spatio-temporal dynamics of the BOLD signal governed by this framework is constrained with respect to the anatomical connections thereby separating the spatial and temporal dynamics. Covariancematrix of this signal is estimated thus getting an estimate of the functional connectivity matrix. The covariance matrix or the BOLD signal in general is expressed in terms of the graph-diffusion-kernels thus forming an analytically elegant expression. Most importantly, the model for FC abstracts out biological details and works in the realm of spectral graph theoretic constructs providing the necessary ease for computational analysis. As this model learns the combination parameters of multiple diffusion kernels and kernels themselves, it is called Multiple Kernel Learning (MKL) model. Apart from superior quantitative performance, the model parameters may act as biomarkers for various cognitive studies. Albeit, the model parameters are learned for a cohort, the model preserves subject-specificity. These parameters can be used as a measure for inter-group differences and dissimilarity identification as has been employed for age-group identification as an example in this thesis. Essentially MKL model partitions FC into two constituents: influence of the underlying anatomical structure into diffusion kernels and the cognitive theme of temporal structure into the model parameters, thus predicting FCs specific to subjects within the cognitive conditions of the cohort. Even though MKL is a cohort based model, it maintains sensitivity towards anatomy. Performance of the model drastically drops down with alterations in SC and model parameters, but does not overfit to the cohort. Resting state fMRI BOLD signals have been observed to show non-stationary dynamics. Such multiple spatio-temporal patterns, represented as dynamic FC matrices, are observed to be cyclically repeating in time motivating use of a generic clustering scheme to identify latent states of dynamics. We propose a novel solution that learns parameters specific to the dynamic states using a graph-theoretic model (temporal-Multiple Kernel Learning, tMKL) and finally p

Abstract

Synthetic or artificial speech has been developed steadily during the last decades. Especially, the intelligibility has reached an adequate level for most applications, especially for communication impaired people. This thesis is concerned with the development of deployable text to speech (TTS) systems for Indian languages. The research described here was conducted with two principle aims: To build stable, scalable and unrestricted synthesis systems for Indian languages; and to design such systems by (a) making observations about the structure of the data and (b) keeping the human in loop so that we end up with systems both acceptable and efficient. In order to do that, I first present a baseline synthesis system for Indian languages built employing unit selection and concatenation as the algorithm and syllables as the basic unit with reduced vowel epenthesis based backoff to handle the missing units. During synthesis, the required utterance specified as a string of syllables help in the selection of appropriate units from the stored database. The system uses word to phone mapping to handle the words from a different language (English) and the selected units are concatenated using a variant of Waveform Similarity Overlap Addition (WSOLA) based method. Two extensions to the system are discussed. • Inspired by the smoothness of the speech obtained after the backend concatenation algorithm, a method is presented to exploit the signal correlation as a join cost in the viterbi algorithm itself with an aim to further improve the naturalness of the system. Two variants of signal similarity are examined - cross correlation based and average magnitude difference function based. • Modeling the prosodic characteristics is an important trait for a text to speech system. Deep generative models which are the standard today benefit from a low dimensional continuous input rather than an array of 1-hot k encoding for the phonemes as is used today. A method has been discussed to obtain continuous representation for the phonemes using a matrix factorization approach and show that the representation can be used to model prosody, by predicting the phoneme durations. The system can be retrained in 6 hours on the same languages and should take approximately the same amount of time for any new language.