SEQUENTIAL DECISIONS AND PREDICTIONS IN NATURAL LANGUAGE PROCESSING

Natural language processing has achieved great success in a wide range of ap- plications, producing both commercial language services and open-source language tools. However, most methods take a static or batch approach, assuming that the model has all information it needs and makes a one-time prediction. In this disser- tation, we study dynamic problems where the input comes in a sequence instead of all at once, and the output must be produced while the input is arriving. In these problems, predictions are often made based only on partial information. We see this dynamic setting in many real-time, interactive applications. These problems usually involve a trade-off between the amount of input received (cost) and the quality of the output prediction (accuracy). Therefore, the evaluation considers both objectives (e.g., plotting a Pareto curve). Our goal is to develop a formal understanding of sequential prediction and decision-making problems in natural language processing and to propose efficient solutions. Toward this end, we present meta-algorithms that take an existent batch model and produce a dynamic model to handle sequential inputs and outputs. Webuild our framework upon theories of Markov Decision Process (MDP), which allows learning to trade off competing objectives in a principled way. The main machine learning techniques we use are from imitation learning and reinforcement learning, and we advance current techniques to tackle problems arising in our settings. We evaluate our algorithm on a variety of applications, including dependency parsing, machine translation, and question answering. We show that our approach achieves a better cost-accuracy trade-off than the batch approach and heuristic-based decision- making approaches. We first propose a general framework for cost-sensitive prediction, where dif- ferent parts of the input come at different costs. We formulate a decision-making process that selects pieces of the input sequentially, and the selection is adaptive to each instance. Our approach is evaluated on both standard classification tasks and a structured prediction task (dependency parsing). We show that it achieves similar prediction quality to methods that use all input, while inducing a much smaller cost. Next, we extend the framework to problems where the input is revealed incremen- tally in a fixed order. We study two applications: simultaneous machine translation and quiz bowl (incremental text classification). We discuss challenges in this set- ting and show that adding domain knowledge eases the decision-making problem. A central theme throughout the chapters is an MDP formulation of a challenging problem with sequential input/output and trade-off decisions, accompanied by a learning algorithm that solves the MDP.

Structured Approaches for Exploring Interpersonal Relationships in Natural Language Text

Human relationships have long been studied by scientists from domains like sociology, psychology, literature, etc. for understanding people's desires, goals, actions and expected behaviors. In this dissertation we study inter-personal relationships as expressed in natural language text. Modeling inter-personal relationships from text finds application in general natural language understanding, as well as real-world domains such as social networks, discussion forums, intelligent virtual agents, etc. We propose that the study of relationships should incorporate not only linguistic cues in text, but also the contexts in which these cues appear. Our investigations, backed by empirical evaluation, support this thesis, and demonstrate that the task benefits from using structured models that incorporate both types of information. We present such structured models to address the task of modeling the nature of relationships between any two given characters from a narrative. To begin with, we assume that relationships are of two types: cooperative and non-cooperative. We first describe an approach to jointly infer relationships between all characters in the narrative, and demonstrate how the task of characterizing the relationship between two characters can benefit from including information about their relationships with other characters in the narrative. We next formulate the relationship-modeling problem as a sequence prediction task to acknowledge the evolving nature of human relationships, and demonstrate the need to model the history of a relationship in predicting its evolution. Thereafter, we present a data-driven method to automatically discover various types of relationships such as familial, romantic, hostile, etc. Like before, we address the task of modeling evolving relationships but don't restrict ourselves to two types of relationships. We also demonstrate the need to incorporate not only local historical but also global context while solving this problem. Lastly, we demonstrate a practical application of modeling inter-personal relationships in the domain of online educational discussion forums. Such forums offer opportunities for its users to interact and form deeper relationships. With this view, we address the task of identifying initiation of such deeper relationships between a student and the instructor. Specifically, we analyze contents of the forums to automatically suggest threads to the instructors that require their intervention. By highlighting scenarios that need direct instructor-student interactions, we alleviate the need for the instructor to manually peruse all threads of the forum and also assist students who have limited avenues for communicating with instructors. We do this by incorporating the discourse structure of the thread through latent variables that abstractly represent contents of individual posts and model the flow of information in the thread. Such latent structured models that incorporate the linguistic cues without losing their context can be helpful in other related natural language understanding tasks as well. We demonstrate this by using the model for a very different task: identifying if a stated desire has been fulfilled by the end of a story.

Easy PRAM-based High-performance Parallel Programming with ICE

A poster of this paper will be presented at the 25th International Conference on Parallel Architecture and Compilation Technology (PACT ’16), September 11-15, 2016, Haifa, Israel.

Language-based Techniques for Practical and Trustworthy Secure Multi-party Computations

Secure Multi-party Computation (MPC) enables a set of parties to collaboratively compute, using cryptographic protocols, a function over their private data in a way that the participants do not see each other's data, they only see the final output. Typical MPC examples include statistical computations over joint private data, private set intersection, and auctions. While these applications are examples of monolithic MPC, richer MPC applications move between "normal" (i.e., per-party local) and "secure" (i.e., joint, multi-party secure) modes repeatedly, resulting overall in mixed-mode computations. For example, we might use MPC to implement the role of the dealer in a game of mental poker -- the game will be divided into rounds of local decision-making (e.g. bidding) and joint interaction (e.g. dealing). Mixed-mode computations are also used to improve performance over monolithic secure computations. Starting with the Fairplay project, several MPC frameworks have been proposed in the last decade to help programmers write MPC applications in a high-level language, while the toolchain manages the low-level details. However, these frameworks are either not expressive enough to allow writing mixed-mode applications or lack formal specification, and reasoning capabilities, thereby diminishing the parties' trust in such tools, and the programs written using them. Furthermore, none of the frameworks provides a verified toolchain to run the MPC programs, leaving the potential of security holes that can compromise the privacy of parties' data. This dissertation presents language-based techniques to make MPC more practical and trustworthy. First, it presents the design and implementation of a new MPC Domain Specific Language, called Wysteria, for writing rich mixed-mode MPC applications. Wysteria provides several benefits over previous languages, including a conceptual single thread of control, generic support for more than two parties, high-level abstractions for secret shares, and a fully formalized type system and operational semantics. Using Wysteria, we have implemented several MPC applications, including, for the first time, a card dealing application. The dissertation next presents Wys*, an embedding of Wysteria in F*, a full-featured verification oriented programming language. Wys* improves on Wysteria along three lines: (a) It enables programmers to formally verify the correctness and security properties of their programs. As far as we know, Wys* is the first language to provide verification capabilities for MPC programs. (b) It provides a partially verified toolchain to run MPC programs, and finally (c) It enables the MPC programs to use, with no extra effort, standard language constructs from the host language F*, thereby making it more usable and scalable. Finally, the dissertation develops static analyses that help optimize monolithic MPC programs into mixed-mode MPC programs, while providing similar privacy guarantees as the monolithic versions.

Towards Proactive Context-aware Computing and Systems

A primary goal of context-aware systems is delivering the right information at the right place and right time to users in order to enable them to make effective decisions and improve their quality of life. There are three key requirements for achieving this goal: determining what information is relevant, personalizing it based on the users’ context (location, preferences, behavioral history etc.), and delivering it to them in a timely manner without an explicit request from them. These requirements create a paradigm that we term as “Proactive Context-aware Computing”. Most of the existing context-aware systems fulfill only a subset of these requirements. Many of these systems focus only on personalization of the requested information based on users’ current context. Moreover, they are often designed for specific domains. In addition, most of the existing systems are reactive - the users request for some information and the system delivers it to them. These systems are not proactive i.e. they cannot anticipate users’ intent and behavior and act proactively without an explicit request from them. In order to overcome these limitations, we need to conduct a deeper analysis and enhance our understanding of context-aware systems that are generic, universal, proactive and applicable to a wide variety of domains. To support this dissertation, we explore several directions. Clearly the most significant sources of information about users today are smartphones. A large amount of users’ context can be acquired through them and they can be used as an effective means to deliver information to users. In addition, social media such as Facebook, Flickr and Foursquare provide a rich and powerful platform to mine users’ interests, preferences and behavioral history. We employ the ubiquity of smartphones and the wealth of information available from social media to address the challenge of building proactive context-aware systems. We have implemented and evaluated a few approaches, including some as part of the Rover framework, to achieve the paradigm of Proactive Context-aware Computing. Rover is a context-aware research platform which has been evolving for the last 6 years. Since location is one of the most important context for users, we have developed ‘Locus’, an indoor localization, tracking and navigation system for multi-story buildings. Other important dimensions of users’ context include the activities that they are engaged in. To this end, we have developed ‘SenseMe’, a system that leverages the smartphone and its multiple sensors in order to perform multidimensional context and activity recognition for users. As part of the ‘SenseMe’ project, we also conducted an exploratory study of privacy, trust, risks and other concerns of users with smart phone based personal sensing systems and applications. To determine what information would be relevant to users’ situations, we have developed ‘TellMe’ - a system that employs a new, flexible and scalable approach based on Natural Language Processing techniques to perform bootstrapped discovery and ranking of relevant information in context-aware systems. In order to personalize the relevant information, we have also developed an algorithm and system for mining a broad range of users’ preferences from their social network profiles and activities. For recommending new information to the users based on their past behavior and context history (such as visited locations, activities and time), we have developed a recommender system and approach for performing multi-dimensional collaborative recommendations using tensor factorization. For timely delivery of personalized and relevant information, it is essential to anticipate and predict users’ behavior. To this end, we have developed a unified infrastructure, within the Rover framework, and implemented several novel approaches and algorithms that employ various contextual features and state of the art machine learning techniques for building diverse behavioral models of users. Examples of generated models include classifying users’ semantic places and mobility states, predicting their availability for accepting calls on smartphones and inferring their device charging behavior. Finally, to enable proactivity in context-aware systems, we have also developed a planning framework based on HTN planning. Together, these works provide a major push in the direction of proactive context-aware computing.

SES-RELATED DIFFERENCES IN WORD LEARNING: EFFECTS OF COGNITIVE INHIBITION AND WORD LEARNING

Socioeconomic status (SES) influences language and cognitive development, with discrepancies particularly noticeable in vocabulary development. This study examines how SES-related differences impact the development of syntactic processing, cognitive inhibition, and word learning. 38 4-5-year-olds from higher- and lower-SES backgrounds completed a word-learning task, in which novel words were embedded in active and passive sentences. Critically, unlike the active sentences, all passive sentences required a syntactic revision. Measures of cognitive inhibition were obtained through a modified Stroop task. Results indicate that lower-SES participants had more difficulty using inhibitory functions to resolve conflict compared to their higher-SES counterparts. However, SES did not impact language processing, as the language outcomes were similar across SES background. Additionally, stronger inhibitory processes were related to better language outcomes in the passive sentence condition. These results suggest that cognitive inhibition impact language processing, but this function may vary across children from different SES backgrounds

Statistical Analysis of Text Summarization Evaluation

This dissertation applies statistical methods to the evaluation of automatic summarization using data from the Text Analysis Conferences in 2008-2011. Several aspects of the evaluation framework itself are studied, including the statistical testing used to determine significant differences, the assessors, and the design of the experiment. In addition, a family of evaluation metrics is developed to predict the score an automatically generated summary would receive from a human judge and its results are demonstrated at the Text Analysis Conference. Finally, variations on the evaluation framework are studied and their relative merits considered. An over-arching theme of this dissertation is the application of standard statistical methods to data that does not conform to the usual testing assumptions.

Trace Oblivious Program Execution

The big data era has dramatically transformed our lives; however, security incidents such as data breaches can put sensitive data (e.g. photos, identities, genomes) at risk. To protect users' data privacy, there is a growing interest in building secure cloud computing systems, which keep sensitive data inputs hidden, even from computation providers. Conceptually, secure cloud computing systems leverage cryptographic techniques (e.g., secure multiparty computation) and trusted hardware (e.g. secure processors) to instantiate a “secure” abstract machine consisting of a CPU and encrypted memory, so that an adversary cannot learn information through either the computation within the CPU or the data in the memory. Unfortunately, evidence has shown that side channels (e.g. memory accesses, timing, and termination) in such a “secure” abstract machine may potentially leak highly sensitive information, including cryptographic keys that form the root of trust for the secure systems. This thesis broadly expands the investigation of a research direction called trace oblivious computation, where programming language techniques are employed to prevent side channel information leakage. We demonstrate the feasibility of trace oblivious computation, by formalizing and building several systems, including GhostRider, which is a hardware-software co-design to provide a hardware-based trace oblivious computing solution, SCVM, which is an automatic RAM-model secure computation system, and ObliVM, which is a programming framework to facilitate programmers to develop applications. All of these systems enjoy formal security guarantees while demonstrating a better performance than prior systems, by one to several orders of magnitude.