Abstract

Legal contracts are formal agreements between parties that create legally binding obligations and rights, typically between two parties involved in a transaction. Legal clauses form the fundamental units of discourse of a legal contract, which are individual provisions that set out specific terms or conditions of the agreement, which collectively make up the entire contract. Contracts are typically drafted by lawyers or other legal professionals who have expertise in contract law through a process that involves a series of steps to ensure that the interests of both the involved parties are captured, while also ensuring the legal correctness of the content. The domain specific content in legal contracts is an interesting area for the application of NLP techniques, not only because of the peculiar nature of legalese, but also because of the challenges that lie in processing the long length of the content involved. Compared to the general domain of language such as news, encyclopedic content, stories, social media content or some other domain-specific verticals such as scientific articles, judicial proceedings, the domain of legal contracts has seen much less research when it comes to the application of deep learning techniques in NLP, with most of it focusing on problems in contract understanding and review. However, a study of the application of generative methods to this domain has been severely lacking. This thesis aims to establish a stepping stone towards the AI-aided generation of legal content, by presenting a study of two generation problems. We also focus on involving user customizability in the process, for easy tailoring of legal content with respect to the parties involved. The thesis starts off with a simple exploratory idea focused on a small set of rental agreements. In this pilot study, we built and studied an agreement drafting tool that could match informal user intents to rental agreement clauses. Observations from this study brought forth the challenges that lie in the generation of legal content, and the need for explicit finetuning in the face of the scarcity of supervised data while leveraging the presence of large unlabeled corpus for content generation. Following this, we study and extend a prior work done towards contract drafting on the recommendation of legal clauses. In this work, we experiment with addition of a new clause to an existing, incomplete contract draft and experiment several strategies to study the effect of different informative signals for clause recommendation. While we model the contexts for recommendation using mean-based pooling of representations from BERTbased architectures, we also explore methods to achieve a better modeling with the use of long range transformers and present the difficulties involved. In another major contribution of this thesis, we devise a pipeline for generation of legal clauses from minimal, keyword-level information. We take inspiration from the content planning for story generation vii viii paradigm, and study the application of coherence-based techniques in devising a content plan for the generation of legal clauses. Our approach is centered on the idea of generating a legal clause, given the topic (or the type) of the clause and a few keywords for customization. Our proposed pipeline uses the topic and keyword input to use a lightweight graph-based mechanism for creating a content plan to act as the outline for the clause to be generated. The content plan is used by a transformer-based generative model for producing an appropriate legal clause, interpolating through the content plan keywords. We propose an ordered content plan consisting of clause keywords, ranked with respect to their generic-ness, with the keywords more generic to the topic ranked higher. We show the benefit of this order as opposed to a natural, sequential order. The study also compares the ordered plan-based generation in contrast to baselines involving prompt-based generation and generation from unordered content plans. We also show the robustness of our approach across a broad range of topics. With our techniques based completely on pretrained transformer architectures, we also contribute a small chapter on the effectiveness of these architectures on four other tasks.

Abstract

Wireless sensor networks are a network of small or large, dedicated sensors designed for particular and specialized purposes. These purposes can range from monitoring environmental behavior like temperature, pressure change, etc., to monitoring man-made physical objects and activities like vehicle speed, sound, density in a region, etc. Sensor networks have a vast application in fields that includes environmental monitoring, natural or artificial object tracking, fire and water detection, natural calamity detection, and traffic monitoring. These groups of sensors and actuators collect information and data for the purpose they have been designed for and then send all the information to a central location for processing or general storage in databases for future use. Based on the topologies of the networks and routing protocols, a particular sensor network can be classified as hierarchical-based routing, location-based routing, and flat-based routing. These protocols are very simple and thus are very prone to attacks like Selective forwarding, HELLO flood attack, Sinkhole attack, Sybil attack, Wormholes, replaying routing information, and Acknowledgement spoofing or altering. Also, since one of the properties of these sensors is limited power, the wireless sensor networks consisting of such sensors generally face power shortages or outages altogether. Wireless Sensors that are small and limited battery capacity are widely used for several different applications. For this reason, energy efficiency, as well as the cost of the overall network, are the two main factors influencing the Wireless Sensor Networks (WSN) performance. Low Energy Adaptive Clustering Hierarchy (LEACH) is an energy-efficient hierarchical-based routing protocol. The prime focus of this thesis is to analyze various existing LEACH protocols and to propose a new LEACH Protocol to resolve some of the issues mentioned above related to wireless sensor networks. This thesis presents an optimized cluster head selection mechanism protocol based on a weighted centroid approach. It is a clustering-based routing protocol that minimizes total energy consumption. Our LEACH-WC protocol algorithm outperforms other LEACH protocol algorithms by making nodes offer themselves candidates for high-energy cluster-heads and adapting the corresponding clusters based on the nodes that choose to be cluster-heads at a given time. We have also discussed an exciting, combined cost and energy wastage minimization problem, and an interesting solution is proposed. Along with the improvement in the existing LEACH protocol, we have also discussed an Along with the improvement in the existing LEACH protocol, we have also discussed an approach that indexes the physical location of various entities across the globe. In the future, combining this improvement in LEACH protocol with the proposed indexing system approach and their application may result in an even better and more intelligent transportation system using wireless sensors network.

Abstract

Music is an important art form which has been present for centuries. Lyrics are a crucial part of a song conveying thoughts, messages and also a wide range of emotions. Individuals listen to songs to satisfy their emotional needs. The task of identifying emotions from a given music track has been an active pursuit in the Music Information Retrieval (MIR) community for years. Music emotion recognition has typically relied on acoustic features, social tags, and other metadata to identify and classify music emotions. The role of lyrics in music emotion recognition remains under-appreciated in spite of several studies reporting superior performance of music emotion classifiers based on features extracted from lyrics. In the first study, we use the transformer-based approach model using XLNet as the base architecture, which has not been used to identify emotional connotations of music based on lyrics. Our proposed approach outperforms existing methods for multiple datasets. We also used a robust methodology to enhance web crawlers’ accuracy for extracting lyrics. There are no datasets of Indian language songs that contain both valence and arousal manual ratings of lyrics. We present a new manually annotated dataset of Telugu songs’ lyrics collected from Spotify with valence and arousal annotated on a discrete scale. A fairly high interannotator agreement was observed for both valence and arousal. Subsequently, we create two music emotion recognition models by using two classification techniques to identify valence, arousal and respective emotion quadrant from lyrics. Support vector machine (SVM) with term frequency-inverse document frequency (TF-IDF) features and fine-tuning the pre-trained XLMRoBERTa (XLM-R) model were used for valence, arousal and quadrant classification tasks. Fine-tuned XLMRoBERTa performs better than the SVM by improving macro-averaged F1-scores of 54.69%, 67.61%, 34.13% to 77.90%, 80.71% and 58.33% for valence, arousal and quadrant classifications, respectively, on 10-fold cross-validation. In addition, we compare our lyrics annotations with Spotify’s annotations of valence and energy (same as arousal), which are based on entire music tracks. We also compare the performance emotion recognition models on the original, translated, transliterated texts. The implications of our findings are discussed. We make the dataset publicly available with lyrics, annotations and Spotify IDs. We also used the XLM-R model to identify emotional connotations in Indian language song lyrics datasets of Hindi and Telugu. We also conducted a perceptual validation study on misclassified lyrics. Lastly, we conduct a study to understand the individual differences between the preferences of music with and without lyrics

Abstract

Data on the internet is growing at ever increasing rates. People share content with each other (or the world at once) over social media platforms such as the likes of Twitter, and consume content from online media outlets such as news agencies like CNN and Dailymail. Gone are the days of monolithic text-only blogs. Online content today generally encompasses multiple modes, or modalities, of communication coming together to convey information. These modalities include text along with images, audio and video. Machine learning models have long been able to capture and understand images and text as separate entities. The first neural network to be used on images dates to before the advent of internet itself. However, only recently, have machine learning researchers started to adopt the notion that multiple modalities can be understood better in a shared setting and under a common architecture, not as isolated black boxes. In this thesis, we attempt to use data-driven approaches towards understanding and improving the interaction of modalities within neural network architectures. The first problem we tackle is that of fake news detection in tweets and posts on microblogging websites such as Weibo. Existing works on the problem focus on independently encoding the different modalities and there is a lack of emphasis on shared learning. Our model attempts to generate richer embeddings through a combination of embeddings generated from pre-trained models. We managed to achieve results that beat the state of the art architecture on the problem statement, and the work was accepted as a full paper in the ASONAM 2021 conference. The second problem we tackle is that of image-aided summarization. While text summarization is a problem that has existed for an eternity, it is not enough to condense information in modality rich sources like news articles. Our model tries to generate textual summaries of articles that demonstrate overlap with image content present in the article, along with selecting the most relevant image from the entire article. We make use of multimodal information retrieval models such as OSCAR to aid in the intermixing of modal information. The third problem we dive into in this thesis is that of content recommendation. Undertaken as a project at LinkedIn, in this problem we try to improve the ranking of LinkedIn Learning content for each user. Since user history is causal, we enable use of time as a modality through making use of techniques such as Time2Vec and train ranking models jointly to enable better vi vii representation of user history and prediction of future action. Through this methodology, we were able to create a strong recommendation system. In the fourth problem, we take a look at the availability of multimodal datasets in Indic languages. To enable and enrich research in this domain in an Indic setting, we try to create the first authentic (not translated) dataset of Image-Text pairs in 11 Indian languages. We use deep-learning based caption filtration techniques to prune down the Samanantar dataset, and then use a query simplification algorithm to create queries to download images related to those sentences. Our work enables the creation of large multimodal models such as CLIP and OSCAR within an Indian setting.

Abstract

Malaria, which is spread by the female anopheles mosquito, is a highly fatal disease that affects many parts of the world, with up to 0.4 million deaths reported worldwide. The detection of malaria infection levels is based on vital gene expressions. Experts quantify malaria parasite-infected RBCs and classify their life cycle stages at the macroscopic level in order to make informed decisions. Several computational approaches have recently been proposed to avoid the dimensionality problem and produce accurately predicted results. Our study presents a theoretical framework to select diagnosis markers and drug targets by implementing ML techniques on sc-RNA-seq data. The main objective is to select the top-ranked genes from the scRNA-seq profiles at different stages of the Plasmodium falciparum (Pf) life cycle inside infected RBC. We employ a supervised learning algorithm coupled with feature selection algorithms to extract the most relevant genes to predict the life cycle stages of Pf inside RBC. The first stage of modeling is to optimize the quality of data from the dataset (5066 features) by removing the irrelevant features. Genetic Algorithm (GA) based search technique is popularly used for feature selection and dealing with high dimensionality datasets. This reduced subset (378) is further utilized in the second stage of high accuracy multi-class classification. In this work, a GA-based dimensionality reduction technique is used on single-cell transcriptomics to obtain an optimised subset of features from a larger data set. To separately transform the selected elements into a lower dimension, features are chosen based on their class variants, taking into account increased efficiency and accuracy. We constructed the protein-protein interaction network (PPIN) of these genes and performed topological analysis using the Search Tool for the Retrieval of Interacting Genes/ Proteins database (STRING 11.0 b) and Gephi software to provide hierarchies according to the importance of the genes in the network. Various topological measures are estimated to evaluate the node characteristics in the PPINs, including degree, between centrality, eccentricity, closeness centrality, eigenvector centrality, and clustering coefficient. Proteins having a high degree and betweenness centrality tend to assert more control over the network function. We also performed gene ontology analysis to determine the role of proteins in the parasite’s life cycle progression. For the multi-class classification of the life cycle of malaria parasite based on oriented gradients and local binary pattern features, a three-pronged approach employing the multi-class Support Vector Machine (SVM), Logistic Regression (LR), and Random Forest (RF) techniques are used. On using these 378 features, RF performed best with a classification accuracy of 92% while SVM had a 91% vi vii accuracy and LR gave 88% accuracy. By merely using the 378 features, we achieved similar or better performance scores for all four classes, across all three models. Further, randomly chosen features from our dataset of 378 were also evaluated using the SVM, LR, and RF models. We achieved an accuracy of 81%, 79%, and 80% for the three respective models. This proves the robustness of the features selected using the GA-based approach. The proposed research methodology can be likely used for improved malaria diagnosis and drug targets.

Abstract

The objective of this study is to examine the interdependence between scene representation and robust autonomy of mobile robots such as quadrotors and cars. It addresses crucial issues with the motion planning and SLAM components in the navigation system. Voxel Grids built from range sensors have been a popular choice of world representation for motion planning. The planners query the map for distance to collision and gradients, the planner uses these measurements to generate a smooth and safe trajectory through trajectory optimization. However, these methods assume the presence of zero sensor noise, such an assumption is severely violated when the maps are built through RGBD cameras, the noise in the point clouds are transferred to the occupancy maps through a non-linear transform and is further propagated to the distance and gradient field which results in a non-parametric noise. The performance of the deterministic planners that rely on such inaccurate gradients was found to decrease. As an antidote we propose CCO-VOXEL the very first chance-constrained optimization (CCO) algorithm that can compute trajectory plans with probabilistic safety guarantees in real-time directly on the voxel-grid representation of the world. We leverage the notion of Hilbert Space embedding of distributions and Maximum Mean Discrepancy (MMD) and gradient free Cross-Entropy Method for obtaining its minimum. We also show how a combination of low-dimensional feature embedding and pre-caching of Kernel Matrices of MMD allows us to achieve real-time performance in simulations as well as in implementations on on-board commodity hardware that controls the quadrotor flight. Voxel maps are not robust to outliers and fail to capture the world geometry when built from a highly noisy and sparse point cloud, such as those generated from triangulation through monocular visualinertial SLAM (VI-SLAM). Inaccurate maps like these can significantly decrease the performance of the motion planner. We propose UrbanFly , a quadrotor planning framework for navigation in a planar urban high-rise environment. The core importance of UrbanFly is its ability to handle sparse and noisy point clouds built from VI-SLAM. Based on the insight that the sky-scrappers are planar, we derived an Uncertainty-Integrated Cuboidal (UIC) representation of the world and developed an uncertaintyaware planner that can interpret and leverage the benefits of UIC to generate safe trajectories in real time. UrbanFly demonstrates a 3.5 times improvement in safety metrics compared to state-of- the-art algorithms. Points clouds generated from LiDAR need a lot of memory to store and transfer and lack structure, making it hard to extract spatial features. This makes it challenging to deploy lidar based Loop Detection and Closure (LDC) for distributed collaborative SLAM. To address this, we present FinderNet, a 6-DOF vi vii distributed LDC system. We develop a novel DEM representation to reveal the underlying geometric structure of the point clouds, and an auto-encoder decoder structure to compress the point clouds. LDC is performed in the compressed space, allowing deployment in a multi-agent setting. Our approach provides view-invariance through a process of canonicalization and differentiable alignment, which is different from existing methods that rely on feature aggregation or overlap estimation. FinderNet has been evaluated on real-world datasets such as Kitti, GPR, and Oxford Robot Car.

Abstract

In this work, we propose IndoLayout, a novel real- time approach for generating high-quality occupancy maps from an RGB image for indoor scenes. Such occupancy maps are often crucial for pathplanning and mapping in indoor environments but are often built using only information contained in the ego view. In contrast, our approach also predicts occupancy values beyond immediately visible regions from just a monocular image, leveraging learnt priors from indoor scenes. Hence, our proposed network can produce a hallucinated, amodal scene layout that includes areas occluded in the RGB image, such as a navigable floor behind a desk. Specifically, we propose a novel architecture that uses self-attention and adversarial learning to vastly improve the quality of the predicted layout. We evaluate our model on several photorealistic indoor datasets and outperform previous relevant work on all metrics that measure layout quality, including newly adopted ones. Finally, we demonstrate the effectiveness of our method by showing significant improvements on the Point Goal navigation task over similar approaches using IndoLayout.

Abstract

Social media has experienced significant growth in the past decade and has enabled people to connect with people all over the world, have increased access to information and have an opportunity to express themselves and join like-minded communities. However, hate speech and online harassment are significant problems, with around two-thirds of adults under 30 having experienced some form of online harassment. Therefore, it becomes essential to regulate content on social media and it needs to be done automatically due to the large volume of daily content. In this thesis, we attempt to solve the above-mentioned problems by building best-in-class classifiers for novel datasets. One way to tackle harmful content online is to have a positive reinforcement approach and encourage positive and supportive messages. We propose a Hope Speech Detection model trained on a first-of-a-kind hope speech dataset. In the first approach, we used contextual embeddings to train classifiers using logistic regression, random forest, SVM, and LSTM based models. The second approach used a majority voting ensemble of 11 models obtained by fine-tuning pre-trained transformer models. Our model ranks first in terms of F1 score in the English language. While supporting and boosting positive content online is helpful, there should also be a distinction made between content that is positive and content that seems positive but encourages emotion suppression. Over the past few years, there has been a growing concern around toxic positivity on social media, a phenomenon where positivity is used to minimize one’s emotional experience. In this thesis, we create a dataset for toxic positivity classification from Twitter and an inspirational quote website. We then perform benchmarking experiments using various text classification models and show the suitability of these models for the task. While there are many hate speech classifiers trained on a generic hate speech definition, there is a lack of datasets that focus on homophobia and transphobia. In this thesis, we describe our approach to classify homophobia and transphobia in social media comments. We used an ensemble of transformer based models to build our classifier. Our classifier ranks 1st in terms of F1 score

Abstract

The discovery of new molecules and materials helps expand the horizons of novel and innovative real-life applications. In the pursuit of finding molecules with desired properties, chemists have traditionally relied on experimentation and recently on combinatorial methods to generate new substances often complimented by computational methods. The sheer size of the chemical space makes it infeasible to search through all possible molecules exhaustively. This calls for fast and efficient methods to navigate the chemical space to find substances with desired properties. This class of problems is referred to as inverse design problems. There is a variety of inverse problems in chemistry encompassing various subfields like drug discovery, retrosynthesis, structure identification etc. Recent developments in modern machine learning (ML) methods have shown great promise in being able to tackle problems of this kind. This has helped in making major strides in all key phases of molecule discovery ranging from in silico candidate generation to their synthesis with focus on small organic molecules. Optimization techniques like Bayesian optimization, reinforcement learning, attention-based transformers, deep generative models like variational autoencoders and generative adversarial networks form a robust arsenal of methods. The first chapter of this thesis summarizes the development of deep learning to tackle a wide variety of inverse design problems in chemistry towards the quest for synthesizing small organic compounds with purpose. Spectroscopy is the study of how matter interacts with electromagnetic radiations of specific frequencies that has led to several monumental discoveries in science. The spectra of any particular molecule is highly information-rich; while structure to spectra is straightforward using computational methods, the inverse relation of spectra to the corresponding molecular structure is still an unsolved problem. Nuclear Magnetic Resonance (NMR) spectroscopy is one such critical technique in the scientists’ toolkit to characterise small organic molecules to biomolecular structures like proteins and nucleic acids. In the second half of the thesis, a novel machine learning framework is proposed that attempts to solve this inverse problem by navigating the chemical space to find the correct structure given an NMR spectra. The proposed framework uses a combination of online Monte-Carlo-Tree-Search (MCTS) and a set of offline trained Graph Convolution Networks to build a molecule iteratively from scratch. Our method is able to predict the correct structure of the molecule ∼ 80% of the time in its top 3 guesses. We believe that the proposed framework is a significant step in solving the inverse design problem of NMR spectra to molecule that would be a significant step forward in high-throughput molecular synthesis

Abstract

In the recent years we have seen an enormous increase in the utilisation of Artificial Intelligence (AI), and the healthcare sector has not been exempt from this transformation. Machine learning systems have demonstrated unmatched success in the healthcare sector thanks to recent developments in digitalization and the massive influx of biomedical data. This could be a game changer for the sector. The rise of cardiovascular diseases across the globe has made electrocardiograms (ECGs) a crucial modality for diagnosis, owing to their non-invasive nature and simplicity. However, gathering 12-lead ECG data is an arduous task outside clinical settings. Wearables can collect an ECG with fewer leads than the standard 12 or the Photoplethysmogram (PPG) data, but medical professionals and conventional ECG analysis software find this data challenging to interpret. To address this issue, ECG reconstruction has been proposed. The second chapter provides a summary of the current applications of AI in diagnosis, prognosis, and therapy, along with its implications for combating the COVID-19 pandemic. The chapter also identifies obstacles to AI’s widespread adoption in the healthcare sector and suggests remedies to help usher in a more intelligent medical future. A unique single-lead to multi-lead ECG reconstruction method is suggested in the third chapter of this thesis employing a modified Attention U-Net framework. Our model, which was solely trained on lead II of ECG, is capable of replicating the remaining 11 leads of the conventional 12-lead ECG with a Pearson correlation of 0.805, a mean square error of 0.0122, and an R-squared value of 0.639. Moreover, a single combined model is used to reconstruct all 11 leads simultaneously, improving performance and simultaneously reducing the computational resources needed for training compared to current literature in the field. The model’s ability for real-life use was also demonstrated by training a deep learning model for multi-disease classification using actual 12-lead ECG data and testing it on both original and reconstructed 12-lead ECG signals. Comparable classification accuracies for both original and reconstructed signals suggest that the proposed model can preserve diagnostically relevant artefacts. The fourth chapter proposes a new approach using a Wasserstein generative adversarial network to convert PPG signals into single lead ECG signals. Unlike previous studies, we utilize longer 3-second PPG segments to generate a complete 3-second ECG signal in a single step. We also address a significant issue in earlier studies where same patient’s data was used for both model training and testing. To address this limitation, patient-specific data segmentation for the train and test set was employed to obtain more reliable results. In summary, this thesis proposes solutions for multi-lead ECG reconstruction and PPG to ECG conversion, aiming to bridge the divide between easily collected heart monitoring data and easily interpretable heart monitoring data.