Abstract

Internet and social media have fundamentally transformed the landscape of news consumption. Consumption of news from the social media is on increase and the consumption through the print media is on decline. The practice of digital journalism is on the rise, thanks to the low barrier of entry, almost zero distribution costs and availability of mobile computing. There has been a proliferation of news publishers and distributors online in the recent times. As more and more audience are consuming news from the internet, it is very important for the news publishers to engage the users on their websites to maximise the revenue, a majority of which is generated from the advertisements on the page. This also caused a raise in the concept of yellow journalism and tabloid journalism, that emphasizes sensational crime stories, gossip columns about celebrities and sports stars, extreme political views and opinions from one perspective, junk food news, and astrology. In fact, aggregating soft news has become a prospective and lucrative business compared to the business of aggregating only the hard news. Many of the publishers publish a mix of both soft news and hard news to maximise their revenue. Sensationalism is often the most common tactic used by the news editors to increase the engagement of the users is sensationalism. The topics and the news stories published are worded in a way to excite the readers and lure the user to click on the link of the story. This type of reporting by the editors can portray the news in a biased manner and may sometimes manipulate the truth of the story. A part of the sensationalism tactic is to report news on very insignificant matters and assigning undue importance to them. Reporting on controversial topics, deliberate omission of facts, using exaggerated words to seek attention, exaggerating trivial events out of proportion, reporting on the content which is insignificant and irrelevent to the audience are some of the tactics used to achieve higher page impressions in the online news media. News editors stoop to low journalistic reporting standards to maximize their engagement and the revenue, at the expense of the quality of the reporting. Such kind of reporting may lead to specific categories of news like Clickbait, Bizarre news, Fake news, Hoax news, Misleading news, Satire and Funny news and Hyperpartisan news. Mimics and clones of popular news outlets are also common. Some categories of the news like Bizarre News, Satire and Funny news are entertaining and harmless. However some categories like, fake news in specific, might cause problems ranging from financial loss, disturbances in society and may cause loss of human life in extreme cases. Clickbaits in the guise of providing entertainment may waste a lot of precious time of the readers and may ultimately cause disappointment to the readers. Fake health remedies disguised in the form of health tit-bits can harm the readers by compromising their health. There are both positive and the negative sides of the sensational softnews. Therefore, it is extremely important to automatically identify, evaluate and appraise the sensational content in the news media. In this thesis we consider the problem identifying sensational soft news content. In specific, we study two types of sensational content - bizarre news and clickbaits. We selected these two categories of the news as these news is usually sensationalized to increase the engagement of the readers. Identifying the sensational soft news by looking at the title is challenging because the titles are usually short text and they are often written in convoluted ways which requires requires high-order semantic understanding, usually with support of facts from some knowledge base. We do a thorough study of both of these types of content and build computational models to categorise the content as hard vs soft news category (like bizarre or clickbait) and also rank its intensity. We selected bizarre and clickbait as the broad categories as both of them have the characterstics of sensationalism in them. We begin the problem of bizarre news identification. We compile a dataset of bizarre news and normal news. We do a deep analysis of bizarre news and regular news and identify handcrafted features like semantic, lingustic and blah blah features for bizarre news detection. We also proceed to categorize the bizarre news into different categories. We then consider the problem of clickbaits. Clickbaits are ubuquitous. In this work, we explore multiple techniques to identify the clickbaits in online social media. We collect and annotate a huge dataset of clickbaits, which is many times larger than any of the existing datasets and use it in a semi supervised approach for doing clickbait classification. We explore various word and sentence embeddings for clickbait identification. We then consider the problem of clickbaits in Spanish. Specifically, we chose to build models to iden

Abstract

Ray Tracing is one of the two major approaches to 3D rendering. Images produced through ray tracing show a higher degree of visual realism, and are more accurate compared to other 3D rendering methods like rasterisation. It is primarily used to generate photo-realistic imagery, as it is based on tracing the path of light going into the scene space. Ray tracing was initially an offline algorithm, and was seen as more suitable for applications which required high quality rendering, where the render time constraints were not as strict. Still images, animated films and television series thus used Ray tracing for rendering visual effects like reflections, shadows, refractions etc. accurately. Computer Generated Imagery (CGI) was first used in the science fiction movie Westworld in 1973. In 1995, ’Toy Story’ became the first film to be fully developed using computer animation. RenderMan, a Ray tracing engine developed by Pixar Studios, made this possible. At that time, rendering enough frames to make a 80-minute film took more than 800,000 machine hours. With massive increases in computational power, and the invention of Graphics Processing Units (GPUs), things have considerably improved, leading to renewed interest in ray tracing research. GPUs gave a huge boost to Ray tracing because of it’s inherently parallel nature. Though GPUs accelerate almost all parts of the Ray trac- ing pipeline, it was still considered to be impractical for interactive, real-time rendering applications, like games. With the annoucement of specalised hardware units for Ray tracing in the Turing GPU microarchitecture by NVIDIA, real-time ray tracing is believed to be possible now. Camera technology has advanced a lot over the years. With standard optical cameras, it was impor- tant to focus accurately on the subject, as slight errors would cause a substantial loss of quality in the desired portions of the image. Digital cameras today have the capability to adjust focus and aperture on their own to accurately focus on a subject. While simulating a camera-esque setup using computer graphics, we don’t have the flexibility of a refocusing after rendering is completed. The only way is to render the setup again with the correct setting. The light field representation offers a viable solution to this problem. Light field is a 4D function that captures all the radiance information of a scene. Traditionally, the creation of light fields was done by image-based rendering mechanisms. These reconstruct the 4D space using pre-captured imagery from various views and employ refocusing algorithms to generate output images. Plenoptic cameras capture the information of all light coming towards the camera, and use that to generate images at any focus setting. Handheld plenoptic cameras, like the Lytro camera series also capture light fields using a microlens array between the sensor and the main lens, but physical constraints have limited their spatial and angular resolutions. The first part of this thesis presents a GPU-based synthetic light field rendering framework that is robust and physically accurate. It can create light fields of very high resolutions, which are orders of magnitude higher than currently available in commercial plenoptic cameras. We show how light field rendering is possible and viable in a synthetic rendering setup, and high-quality images at any desired focus and aperture setting can be generated in reasonable times with a very low memory footprint. We explain the theory behind synthetic light field creation, different representations possible for a 4D light field, and show results based on our implementation, as compared with state-of-the-art ray tracing frameworks like POVRay. In the second part of the thesis, we attempt to solve the problem of light field storage. The 4D light field is inherently rich in information, but is very bulky to store. We address this issue by separating the light field creation and output image generation passes of a synthetic light field framework. Our thesis presents a compact representation of the 4D light slabs using a video compression codec and demon- strate different quality-size combinations using this representation. We demonstrate the equivalence of the standard light field camera representation, called the sub-aperture representation, with light slab representation for synthetic light fields. We use this to exhibit the capability of our framework, not only to trace light fields of very high resolutions, but also to store them in memory. The required images can therefore be generated independently from the ray tracing pass with a very small cost

Abstract

Proteins are important building blocks of biological cells. They are not static entities, but they exhibit a wide spectrum of dynamics ranging from atomic vibrations, side chain motions to large scale domain motions. Their dynamics help them perform many biological functions in cells. Recent experimental studies on the binding of ligands to proteins have highlighted the importance of low-frequency vibrational dynamics in the kinetics and thermodynamics of ligand binding and in bringing about binding-induced structural changes in proteins. It is necessary to investigate the vibrational dynamics to comprehend on the functional aspects of proteins. In the present thesis, we have characterized the vibrational dynamics of proteins and how it is aected by perturbations such as ligand binding and surface water using normal mode analysis (NMA) and molecular dynamics (MD) simulation. Calmodulin, a calcium binding protein found in most eukaryotic cells, is used as a model system to study the changes in the vibrational dynamics after it binds to various peptides. We have also developed a web application with a GUI interface to perform NMA and to analyze vibrational properties of proteins. Chapter 1 presents an overview of the introduction to proteins and the importance of vibrational dynamics. A detailed overview of various computational methods used to characterize protein dynamics is presented in Chapter 2; specically, the fundamental theory behind NMA, MD and the methods to calculate the binding anity of protein-ligand complexes are provided in Chapter 2. The computational details of the development of the web application to perform NMA of proteins are presented in Chapter 3. The specics of GUI interface to our web application that enables calculation of vibrational properties such as the vibrational density of states (VDOS), vibrational entropy and participation ratios of proteins and that displays vibrational modes of interest are also reported in Chapter 3. Chapter 4 explores on a study of the vibrational dynamics of CaM in the presence and absence of ligands to understand ligandinduced changes in the vibrational motions of CaM. The correlation between binding-induced changes in vibrational dynamics of CaM and the binding anity of CaM-peptide complexes is also explored. The eect of solvent on the vibrational properties of CaM is also investigated by varying its hydration level and the corresponding results are also reported in Chapter 4. Chapter 5 concludes this thesis by presenting the concluding remarks of the work performed herein.

Abstract

Sports data is probably the most widely looked at and analysed data source in the world. Data captured during a sports match is of much interest to fans who use data to validate their hypotheses about player/team performances and to conduct an analysis of the happenings during the game. Data collection has largely been a manual process and thus the data was discrete in nature and outcome based. For example in the case of cricket, for a ball bowled the outcome was recorded as to how many runs were scored but the position the ball was pitched and where it passed the batsman can not be recorded with a reasonable level of consistency, if done manually. With the introduction of various technology based tracking solutions, such kinds of spatial data that is consistent, can now be complied. When you think of continuous sports such as football, hockey this kind of data is continuous and holds much more context than that the discrete, outcome based data used to present. The 2 methods used currently are either tracking through hardware tracking devices that need to be attached to each object of interest or by using computer vision techniques to analyze videos of the game that must be recorded by multiple specially installed single purpose cameras. Hence, as the data made available is becoming richer we see that sports players and teams are increasingly making use of data to devise strategies and to gain a competitive advantage. This is making such data very powerful as sports teams can use it for prescribing tactics, betting companies for better predictions, and sport broadcasters and fans for analysis. Spatial data is especially interesting in football because the dynamic positions of the players on the playing field are highly important to decide the outcome of the game. Football players are often advised by their managers precisely as to how they must position themselves during different passages of play. Managers are often thought to be playing chess with each other as they plan their moves and organize their troops in various tactical formations aimed at giving them a leg-up over their opponent. The importance of spatial data in football cannot be overstated. We have taken a different approach of using computer vision to gather data pertaining to player and ball location by working purely off a broadcast video for the sport of football. The benefits of this approach are that it is economically cheaper and can also be used to analyze any televised match from the past. The 2 other techniques discussed are highly intrusive and expensive to setup. Another drawback with these techniques is that they need specially installed hardware prior to the start of the game to gather data. Using our technique we can analyze data of any televised football game easily and also use it to analyze any televised game for the past. As data hungry machine learning models are being increasingly used to prescribe and predict, opening up the data from all the games in the past could be hugely significant and we hope our research proves a significant step in that direction

Abstract

Replacing overexposed or dull skies in outdoor photographs is a desirable photo manipulation. It is often necessary to color correct the foreground after replacement to make it consistent with the new sky. Methods have been proposed to automate the process of sky replacement and color correction. However, many times a color correction is unwanted by the artist or may produce unrealistic results. Style similarity is an important measure for many applications such as style transfer, fashion search, art exploration, etc. However, computational modeling of style is a difficult task owing to its vague and subjective nature. Most methods for style based retrieval use supervised training with pre-defined categorization of images according to style. While this paradigm is suitable for applications where style categories are well-defined and curating large datasets according to such a categorization is feasible, in several other cases such a categorization is either ill-defined or does not exist. In this thesis, we primarily study various image representations and their applications in understanding visual style and automatic background replacement. First, we propose a data-driven approach to sky-replacement that avoids color correction by finding a diverse set of skies that are consistent in color and natural illumination with the query image foreground. Our database consists of ∼1200 natural images spanning many outdoor categories. Given a query image, we retrieve the most consistent images from the database according to L2 similarity in feature space and produce candidate composites. The candidates are re-ranked based on realism and diversity. We used pre-trained CNN features and a rich set of hand-crafted features that encode color statistics, structural layout, and natural illumination statistics, but observed color statistics to be the most effective for this task. We share our findings on feature selection and show qualitative results and a user-study based evaluation to show the effectiveness of the proposed method. Next, we propose an unsupervised protocol for learning a neural embedding of visual style of images. Our protocol for learning style based representations does not leverage categorical labels but a proxy measure for forming triplets of anchor, similar, and dissimilar images. Using these triplets, we learn a compact style embedding that is useful for style-based search and retrieval. The learned embeddings outperform other unsupervised representations for style-based image retrieval task on six datasets that capture different meanings of style. We also show that by fine-tuning the learned features with datasetspecific style labels, we obtain best results for image style recognition task on five of the six datasets. To the best of our knowledge, ours is the first work that provides a comprehensive review and evaluation of style representations in an unsupervised setting. Our findings along with the curated outdoor scene database would be useful to the community for future research in the direction of sky-search and sky-replacement

Abstract

Climate change calls for a unanimous global initiative, and all nations have been requested to reduce their growing greenhouse gas emissions to tackle this issue. In the United Nations’ 21st Conference on Climate Change, India has pledged to reduce its emission intensity by 33-35% of its 2005 level by 2030. Several studies predicted that coal-based electricity generation in India could double by 2040. At present, the residential building sector alone accounts for 24% of total electricity consumption in India, representing second-highest compounded annual growth rate of 7.58% after the Industrial sector. The Alliance for an Energy Efficient Economy (AEEE) in India predicted that the electricity consumption growth rate in residences would further increase in the future, with increasing residential built-stock and growth in ownership of cooling appliances. However, the growth rate in consumption may not necessarily be linear with increasing residential built-stock and cooling appliances ownership per unit household. Climate Change impact studies conducted on residential buildings in cold and tropical climates mostly infer that cooling energy consumption would rise with the predicted change in the climate. This study aims to assess the potential impact of climate change on Indian residential building energy consumption in five climatic zones, using a transient building energy simulation approach. Future typical weather files are generated for the year 2050 for these climatic zones, using Climate Change World Weather File Generator developed by University of Southampton. The present conventional (PC) and present energy-efficient (PE) building cases are modelled to identify their energy performance for present and future weather conditions. Further, the study includes cases with future energy conservation measures for PC building – Present Conventional Buildings with Future Energy-efficiency retrofit Measures (PCFM) and PE building – Present Energy-efficient Buildings with Future Energy-efficient retrofit Measures (PEFM) in the scope. The PCFM and PEFM cases are simulated only for future climate to estimate energy savings through future energy conservation measures. A future energy-efficient (FE) building case is also modelled and simulated with features involving high insulated envelope and future energy-efficient appliances. Various global studies on residential energy and the impact of climate change have focused on measures for new buildings, and on revision of insulation code. These studies focus on insulation levels for building envelope and retrofit measures for cooling system towards energy conservation. This work contributes by incorporating future retrofit technologies such as electrochromic windows with solar PV glass, and simple home automation strategies for coupling ceiling fans with air conditioning systems for future energy conservation. The variable loads like ceilings fans and geysers are modelled and simulated through Energy Management System scripting in EnergyPlus. The Building Integrated Photovoltaic (BIPV) system is also modelled and simulated as part of future energy conservation retrofit measures Based on energy simulations, it is observed that the present-day built residential buildings energy performance would be influenced by climate change. The global warming would increase the cooling energy consumption of present-day residential buildings leading to increase in total energy consumption. The consumption rate can be reduced to the levels less than the current energy consumption by deploying future energy efficient retrofit strategies for present-day buildings. The present-day energy efficient buildings can reduce their consumption by half in future whereas conventional buildings with low-thermal envelope and energy in-efficient appliances could reduce by more than half by retrofitting in the future. The facades of high-rise buildings are widely exposed to solar radiation leading to large solar ingress. However, the BIPV system deployment would not only reduce the heat ingress from the envelope but also convert some part of the solar radiation into electricity.

Abstract

In this work, we try to advance knowledge about differences in neural processing of music between musicians and non-musicians, both through our findings and methodologically. Music is omnipresent in our daily lives. Music listening has been found to have many positive benefits on health, and musicbased interventions have been shown to be an effective form of alternative and complimentary therapy to aid treatment for various disorders. Importantly, it has been suggested that an individual’s musical background and musicality play a key role in modulating the successful acquisition of the benefits music has to offer. Musical training has been shown to engender structural and functional changes in the brain due to intense repetitive sensory-motor demands. In addition to engendering enhanced sensitivity in the processing, representation, and discrimination of sounds and music, musical training has also been associated with transfer effects such as enhanced cognitive function in speech and language processing, motor abilities, attention and memory, etc. Music is thus an ideal tool to study brain adaptation and plasticity, and musicians an ideal group to study brain changes driven by experience, especially when juxtaposed with a non-musicians group. Thus, along with understanding the neural underpinnings of music feature processing, it is also essential to look at how musical training modulates the functionality of the brain by investigating group level differences between musicians and non-musicians. Our work is a step in that direction. A distinctive aspect of the work presented is the use of the naturalistic paradigm wherein participants listen to entire pieces of music without performing any task while being scanned and renders itself an attractive and more ecologically valid approach to studying music perception as opposed to the hitherto controlled paradigms. In our work, we look at the differences between musicians and non-musicians from two different viewpoints - from a segregated viewpoint (which aims to identify local regions that are specialized for a particular task) in our first study, and from an integrated viewpoint (which allows for identification and characterization of interactions between brain regions that allow for integrated functioning) in our second study. Group differences were identified in both the studies thereby revealing differences in neural encoding as a function of musical expertise. Musical training and expertise is likely to enable top-down analytic processing in the musicians, especially of the more complex and higher-level aspects and features of music, leading to differences in listening and processing strategies caused predominantly by their varied backgrounds. This resulted in lesser homogeneity at the group level for the musicians.On the other hand, the non-musicians exhibited greater within group consistencies, indicating primarily sensory, bottom-up processing. In our first study which takes a segregated approach, we investigate the differences between musicians and non-musicians in the encoding of acoustic features encompassing musical timbre, rhythm, and tonality in a continuous music listening paradigm. 18 musicians and 18 non-musicians were scanned using functional magnetic resonance imaging (fMRI) while listening to three 8-minute long musical pieces representing different musical styles. Acoustic features corresponding to timbre, rhythm, and tonality were computationally extracted from the stimuli and correlated with brain responses. Overall, non-musicians exhibited broader regions of correlations, implying greater within group similarities in musical feature processing, especially in the auditory and default mode network-related regions. Musicians demonstrated significant correlations in regions possessing adaptations to music processing due to training, in addition to greater involvement of limbic and reward regions in response to rhythm, indicative of greater affective and analytic processing. However, as a group, they did not exhibit large regions of consistent correlation patterns, especially in processing high-level features, which could be attributed to differences in processing strategies arising out of their varied training and background. For our second study which takes an integrated approach, we note that studies from a segregated viewpoint and studies involving network analysis have indicated the presence of large-scale brain networks involved in the processing of music and have highlighted the differences between musicians and non-musicians. However, network analysis studies of functional connectivity in the brain during music listening have thus far focused solely on static network analysis. Dynamic Functional Connectivity (DFC) studies have lately been found useful in uncovering meaningful, time-varying functional connectivity information in both resting-state and task-based experimental settings. We examin

Abstract

Current mobile technology does not utilize the spectrum efficiently. This problem is going to increase in future because of increase in wireless devices. Cognitive Radio (CR) is one of the solutions to exploit the underutilized spectrum. It is key enabler for next generation mobile technology 5G with applications such as efficient spectrum utilization and interference management. Spectrum sensing is an important component of cognitive radio networks (CRNs) as it provides spectrum awareness required for cognitive processing. In this thesis, we first investigate a distributed and heterogeneous CRN, comprising of secondary users (SUs) employing either energy detector (ED) or autocorrelation detector (AD) to detect the pres- ence or absence of an orthogonal frequency division multiplexing (OFDM) based primary user (PU). For the considered heterogeneous cooperative spectrum sensing (CSS), the optimal soft combining rule is derived. The performance of this optimal fusion rule and different hard combining schemes such as OR, AND and MAJORITY is presented for the case when the noise variance is exactly known. Later, the effect of noise uncertainty is also presented. The proposed heterogeneous CSS is shown to com- bine the excellent performance of the EDs (when the noise variance is exactly known) and robustness of the ADs to the noise uncertainty. However, the proposed soft combining based fusion rule assumes independence between the observations which may not hold in practical scenarios. Incorporating dependence between observations in fusion rule for CSS scheme forced us to look for alternative concepts and theories beyond standard concepts used in the context of spectrum sensing. Copula theory provides a new dimension to the picture. It enables us to incorporate dependence in the fusion process. Using the concept of copula theory, we forged a novel copula-based fusion rule for spectrum sensing. We investigate a distributed detection model where SUs employ ADs for the detection of a PU. In the presence of a PU, it is assumed that the observations across different SUs and subsequently the decision statistics are dependent. For the fusion of these dependent statistics, different copulas such as t-copula, Gaussian, Clayton and Gumbel are employed. In the presence of dependence among decision statistics, significant improvement in detection performance is observed while using copula theory instead of the traditional assumption of independence. Simulation results are presented to show the superiority of copula-based spectrum sensing.

Abstract

Object recognition deals with the recognition of objects belonging to various categories such as aeroplane, bird, chair etc. A more challenging sub-problem is fine-grained object recognition wherein objects belonging to the subordinate categories in a same entry level category are to be recognized; such as recognition of bird species and car models. It is one of the cornerstones in object recognition due to its potential in augmenting human vision. In the recent years, fine-grained object recognition using Image based Recognition (IR) has seen tremendous progress with the advent of convolutional neural networks. However, IR is cost-sensitive as it is excessively picture-dependent and fails without the adequate number of quality pictures. On the other hand, a picture-independent recognition approach called Question-answering based recognition (QR) has been investigated in which the visual semantic attributes of objects are exploited by posing them as questions to a human-in-the-loop. The responses received to the questions are then utilized for recognizing the test object. However, QR entails heavy human labor as it relies entirely on the human input for recognition. Subsequently, hybrid human-computer recognition called Image and Question answering based recognition (IQR) has been proposed which utilizes both IR and QR jointly for more effective recognition. Still, many of the works in IQR are picture-dependent making them cost-sensitive. The few picture-independent works under IQR which use binary/multiple choice question answering are slow and fail to produce the desirable performance as needed by the present recognition applications. To this end, in this thesis, we propose an improved batch-based question answering method for making the IQR recognition cost-effective, i.e., efficient and picture-independent. We refer to the proposed approach as Recognition via Image and Batch based local Question answering (RIBQ). Unlike the existing picture-independent recognition methods, the proposed method poses a batch of multiple discriminating questions at each time-step to the human-in-the-loop for distinguishing the probable classes. As a part of this, we introduce a cluster based local feature extraction method for finding the best set of N highly discriminating questions to pose to the human-in-the-loop at each time-step. At each time-step, the proposed method first clusters the probable classes into N clusters. From each of the obtained clusters, we mine an attribute whose expected presence/absence is exclusively predominant in that cluster using the proposed concept of locality degree of an attribute. The mined set of attributes from the clusters called as cluster-centric local attributes are then posed as questions to the human-in-the-loop. Based on the responses received, the class probability estimates are re-computed and the probable classes to be considered in the next time-step are determined. After a preset number of time-steps, the most probable class of the target object is returned as the final prediction. In this way, we make the recognition process much faster and robust by modelling localized information obtained from the responses at each time-step. When pictures are available, for more qualitative recognition, we make the human-in-the-loop and vision supplement each other by facilitating the plugin of vision-based predictions into the framework. Experiments on three challenging datasets show significant performance improvement with respect to accuracy and computation time as compared to the existing schemes. The results demonstrate that the proposed concept of batch-based local question answering identifies potential discriminating questions in an efficient manner. They in-turn facilitate processing and knowledge acquisition at a faster rate as compared to the questions posed in the existing question answering approaches, leading to better recognition. Hence, the proposed method achieves effective recognition regardless of the availability of pictures, making it cost-effective. Overall, the proposed approach provides scope for developing efficient information extraction methodologies in all human-in-the-loop based recognition systems.

Abstract

Social media sites such as Twitter and Facebook have become people’s main channels for connecting and voicing their opinions with other individuals and the community. With the growth of social media platforms in India and people expressing themselves in multiple languages and the open and casual nature of these sites leads to more people voicing themselves in their native language, resulting in a greater amount of code-mixed content that is currently lacking in the annotated set of data. Through the access to public opinion on almost every subject, we can collect a huge amount of user data that could prove useful to different businesses, making tasks such as opinion mining, sentiment analysis and opinion mining even more essential. Therefore, recently the task of understanding users in low-resource languages has become the most researched task. With the penetration of the Internet and its proliferation into multilingual societies, the linguistic diversity of the real world is now reflected in online communities. Limiting solutions to a set of major languages is no longer viable. Nevertheless, the proliferation is relatively recent, and hence the amount of available data in many widely-spoken languages is inadequate. Manual annotation of data is a hu- mongous task and often expensive. In this thesis, we aim towards working on these widely-spoken languages to understand users better when they express themselves in their native language on social media platforms. Understanding people broadly means understanding their feeling, opinion, stance towards a particular target, and thus it brings in the tasks of stance and humor detection. We are mainly focusing on text classification which refers to the process of determining text categories based on the text content under a given classification system. We present new solutions for stance detection and the other for detection of humor in English-Hindi code-mixed tweets. The tweets for stance detection are collected for the target topic - ‘Demonetisation.’ Deep neural network models have been widely used in the field of natural language processing (NLP). Convolutional Neural Network’s (CNN’s) and Recurrent neural networks (RNN’s), which are able to process sequences of variable length, are known methods for sequence modeling tasks. Long-short-term memory (LSTM) is one category of RNNs and has achieved exceptional text classification efficiency. Nonetheless, due to the high dimensionality and sparse text data and the dynamic semantics of the natu- ral language, text classification poses demanding challenges — a novel and integrated architecture that incorporates a bidirectional LSTM (BiLSTM) to solve the above problems. A novel and unified archi- tecture containing a bidirectional LSTM (BiLSTM), attention mechanism, and a convolutional layer are proposed in this paper in order to solve the above problems.