Instructional prompts are a novel technique that can significantly improve the performance of natural language processing tasks by specifying the task instruction to the language model. This is the first paper that uses instructional prompts to improve performance of the question answering task in biomedical domain. This work makes two significant contributions. Firstly, a question answer dataset of 600K question answer pairs has been developed by using the medical textbook ‘Differential Diagnosis Primary Care’, which contains information on how to diagnose a patient by observing their disease symptoms. Secondly, a question answering language model augmented with instructional prompts has been developed by training on the medical information extracted from the book ‘Differential Diagnosis Primary Care’. Experiments have been conducted to demonstrate that it performs better than a normal question answering model that does not use instructional prompts. Instructional prompts are based on prompt tuning and prefix tuning, which are novel techniques which can help train language model to do specific downstream tasks by keeping majority of model parameters frozen, and only optimizing a small number of continuous task-specific vectors (called the prefixes).

Increasing misinformation in social media channels has become more prevalent since the beginning of the COVID-19 pandemic as countless myths and rumors have circulated over the internet. This misinformation has potentially lethal consequences as many people make important health decisions based on what they read online, thus creating an urgent need to combat it. Although many Natural Language Processing (NLP) techniques have been used to identify misinformation in text, prompt-based methods are under-studied for this task. This work explores prompt learning to classify COVID-19 related misinformation. To this extent, I analyze the effectiveness of this proposed approach on four datasets. Experimental results show that prompt-based classification achieves on average ~13% and ~6% improvement compared to a single-task and multi-task model, respectively. Moreover, analysis shows that prompt-based models can achieve competitive results compared to baselines in a few-shot learning scenario.

Machine learning models can pick up biases and spurious correlations from training data and projects and amplify these biases during inference, thus posing significant challenges in real-world settings. One approach to mitigating this is a class of methods that can identify filter out bias-inducing samples from the training datasets to force models to avoid being exposed to biases. However, the filtering leads to a considerable wastage of resources as most of the dataset created is discarded as biased. This work deals with avoiding the wastage of resources by identifying and quantifying the biases. I further elaborate on the implications of dataset filtering on robustness (to adversarial attacks) and generalization (to out-of-distribution samples). The findings suggest that while dataset filtering does help to improve OOD(Out-Of-Distribution) generalization, it has a significant negative impact on robustness to adversarial attacks. It also shows that transforming bias-inducing samples into adversarial samples (instead of eliminating them from the dataset) can significantly boost robustness without sacrificing generalization.

Systematic Reviews (SRs) aim to synthesize the totality of evidence for clinical practice and are important in making clinical practice guidelines and health policy decisions. However, conducting SRs manually is a laborious and time-consuming process. This challenge is growing due to the increase in the number of databases to search and the papers being published. Hence, the automation of SRs is an essential task. The goal of this thesis work is to develop Natural Language Processing (NLP)-based classifiers to automate the title and abstract-based screening for clinical SRs based on inclusion/exclusion criteria. In clinical SRs, a high-sensitivity system is a key requirement. Most existing methods for SRs use binary classification systems trained on labeled data to predict inclusion/exclusion. While previous studies have shown that NLP-based classification methods can automate title and abstract-based screening for SRs, methods for achieving high-sensitivity have not been empirically studied. In addition, the training strategy for binary classification has several limitations: (1) it ignores the inclusion/exclusion criteria, (2) lacks generalization ability, (3) suffers from low resource data, and (4) fails to achieve reasonable precision at high-sensitivity levels. This thesis work presents contributions to several aspects of the clinical systematic review domain. First, it presents an empirical study of NLP-based supervised text classification and high-sensitivity methods on datasets developed from six different SRs in the clinical domain. Second, this thesis work provides a novel approach to view SR as a Question Answering (QA) problem in order to overcome the limitations of the binary classification training strategy; and propose a more general abstract screening model for different SRs. Finally, this work provides a new QA-based dataset for six different SRs which is made available to the community.

Visual question answering (VQA) is a task that answers the questions by giving an image, and thus involves both language and vision methods to solve, which make the VQA tasks a frontier interdisciplinary field. In recent years, as the great progress made in simple question tasks (e.g. object recognition), researchers start to shift their interests to the questions that require knowledge and reasoning. Knowledge-based VQA requires answering questions with external knowledge in addition to the content of images. One dataset that is mostly used in evaluating knowledge-based VQA is OK-VQA, but it lacks a gold standard knowledge corpus for retrieval. Existing work leverages different knowledge bases (e.g., ConceptNet and Wikipedia) to obtain external knowledge. Because of varying knowledge bases, it is hard to fairly compare models' performance. To address this issue, this paper collects a natural language knowledge base that can be used for any question answering (QA) system. Moreover, a Visual Retriever-Reader pipeline is proposed to approach knowledge-based VQA, where the visual retriever aims to retrieve relevant knowledge, and the visual reader seeks to predict answers based on given knowledge. The retriever is constructed with two versions: term based retriever which uses best matching 25 (BM25), and neural based retriever where the latest dense passage retriever (DPR) is introduced. To encode the visual information, the image and caption are encoded separately in the two kinds of neural based retriever: Image-DPR and Caption-DPR. There are also two styles of readers, classification reader and extraction reader. Both the retriever and reader are trained with weak supervision. The experimental results show that a good retriever can significantly improve the reader's performance on the OK-VQA challenge.

Reverse engineers use decompilers to analyze binaries when their source code is unavailable. A binary decompiler attempts to transform binary programs to their corresponding high-level source code by recovering and inferring the information that was lost during the compilation process. One type of information that is lost during compilation is variable names, which are critical for reverse engineers to analyze and understand programs. Traditional binary decompilers generally use automatically generated, placeholder variable names that are meaningless or have little correlation with their intended semantics. Having correct or meaningful variable names in decompiled code, instead of placeholder variable names, greatly increases the readability of decompiled binary code. Decompiled Identifier Renaming Engine (DIRE) is a state-of-the-art, deep-learning-based solution that automatically predicts variable names in decompiled binary code. However, DIRE's prediction result is far from perfect. The first goal of this research project is to take a close look at the current state-of-the-art solution for automated variable name prediction on decompilation output of binary code, assess the prediction quality, and understand how the prediction result can be improved. Then, as the second goal of this research project, I aim to improve the prediction quality of variable names. With a thorough understanding of DIRE's issues, I focus on improving the quality of training data. This thesis proposes a novel approach to improving the quality of the training data by normalizing variable names and converting their abbreviated forms to their full forms. I implemented and evaluated the proposed approach on a data set of over 10k and 20k binaries and showed improvements over DIRE.

Self-Driving cars are a long-lasting ambition for many AI scientists and engineers. In the last decade alone, many self-driving cars like Google Waymo, Tesla Autopilot, Uber, etc. have been roaming the streets of many cities. As a rapidly expanding field, researchers all over the world are attempting to develop more safe and efficient AI agents that can navigate through our cities. However, driving is a very complex task to master even for a human, let alone the challenges in developing robots to do the same. It requires attention and inputs from the surroundings of the car, and it is nearly impossible for us to program all the possible factors affecting this complex task. As a solution, imitation learning was introduced, wherein the agents learn a policy, mapping the observations to the actions through demonstrations given by humans. Through imitation learning, one could easily teach self-driving cars the expected behavior in many scenarios. Despite their autonomous nature, it is undeniable that humans play a vital role in the development and execution of safe and trustworthy self-driving cars and hence form the strongest link in this application of Human-Robot Interaction. Several approaches were taken to incorporate this link between humans and self-driving cars, one of which involves the communication of human's navigational instruction to self-driving cars. The communicative channel provides humans with control over the agent’s decisions as well as the ability to guide them in real-time. In this work, the abilities of imitation learning in creating a self-driving agent that can follow natural language instructions given by humans based on environmental objects’ descriptions were explored. The proposed model architecture is capable of handling latent temporal context in these instructions thus making the agent capable of taking multiple decisions along its course. The work shows promising results that push the boundaries of natural language instructions and their complexities in navigating self-driving cars through towns.

Referring Expression Comprehension (REC) is an important area of research in Natural Language Processing (NLP) and vision domain. It involves locating an object in an image described by a natural language referring expression. This task requires information from both Natural Language and Vision aspect. The task is compositional in nature as it requires visual reasoning as underlying process along with relationships among the objects in the image. Recent works based on modular networks have

displayed to be an effective framework for performing visual reasoning task.

Although this approach is effective, it has been established that the current benchmark datasets for referring expression comprehension suffer from bias. Recent work on CLEVR-Ref+ dataset deals with bias issues by constructing a synthetic dataset

and provides an approach for the aforementioned task which performed better than the previous state-of-the-art models as well as showing the reasoning process. This work aims to improve the performance on CLEVR-Ref+ dataset and achieve comparable interpretability. In this work, the neural module network approach with the attention map technique is employed. The neural module network is composed of the primitive operation modules which are specific to their functions and the output is generated using a separate segmentation module. From empirical results, it is clear that this approach is performing significantly better than the current State-of-theart in one aspect (Predicted programs) and achieving comparable results for another aspect (Ground truth programs)

Information Retrieval (IR) is the task of obtaining pieces of data (such as documents or snippets of text) that are relevant to a particular query or need from a large repository of information. IR is a valuable component of several downstream Natural Language Processing (NLP) tasks, such as Question Answering. Practically, IR is at the heart of many widely-used technologies like search engines.

While probabilistic ranking functions, such as the Okapi BM25 function, have been utilized in IR systems since the 1970's, modern neural approaches pose certain advantages compared to their classical counterparts. In particular, the release of BERT (Bidirectional Encoder Representations from Transformers) has had a significant impact in the NLP community by demonstrating how the use of a Masked Language Model (MLM) trained on a considerable corpus of data can improve a variety of downstream NLP tasks, including sentence classification and passage re-ranking.

IR Systems are also important in the biomedical and clinical domains. Given the continuously-increasing amount of scientific literature across biomedical domain, the ability find answers to specific clinical queries from a repository of millions of articles is a matter of practical value to medics, doctors, and other medical professionals. Moreover, there are domain-specific challenges present in the biomedical domain, including handling clinical jargon and evaluating the similarity or relatedness of various medical symptoms when determining the relevance between a query and a sentence.

This work presents contributions to several aspects of the Biomedical Semantic Information Retrieval domain. First, it introduces Multi-Perspective Sentence Relevance, a novel methodology of utilizing BERT-based models for contextual IR. The system is evaluated using the BioASQ Biomedical IR Challenge. Finally, practical contributions in the form of a live IR system for medics and a proposed challenge on the Living Systematic Review clinical task are provided.

One of the measures to determine the intelligence of a system is through Question Answering, as it requires a system to comprehend a question and reason using its knowledge base to accurately answer it. Qualitative word problems are an important subset of such problems, as they require a system to recognize and reason with qualitative knowledge expressed in natural language. Traditional approaches in this domain include multiple modules to parse a given problem and to perform the required reasoning. Recent approaches involve using large pre-trained Language models like the Bidirection Encoder Representations from Transformers for downstream question answering tasks through supervision. These approaches however either suffer from errors between multiple modules, or are not interpretable with respect to the reasoning process employed. The proposed solution in this work aims to overcome these drawbacks through a single end-to-end trainable model that performs both the required parsing and reasoning. The parsing is achieved through an attention mechanism, whereas the reasoning is performed in vector space using soft logic operations. The model also enforces constraints in the form of auxiliary loss terms to increase the interpretability of the underlying reasoning process. The work achieves state of the art accuracy on the QuaRel dataset and matches that of the QuaRTz dataset with additional interpretability.