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.

DIY Archives: Enhancing Access to Collections via Free, Open-Source Platforms

Presentation from the MARAC conference in Boston, MA on March 18-21, 2015. S. 24 - DIY Archives: Enhancing Access to Collections via Free, Open-Source Platforms

Event-Code Interaction Directed Test Cases

The Graphical User Interface (GUI) is an integral component of contemporary computer software. A stable and reliable GUI is necessary for correct functioning of software applications. Comprehensive verification of the GUI is a routine part of most software development life-cycles. The input space of a GUI is typically large, making exhaustive verification difficult. GUI defects are often revealed by exercising parts of the GUI that interact with each other. It is challenging for a verification method to drive the GUI into states that might contain defects. In recent years, model-based methods, that target specific GUI interactions, have been developed. These methods create a formal model of the GUI’s input space from specification of the GUI, visible GUI behaviors and static analysis of the GUI’s program-code. GUIs are typically dynamic in nature, whose user-visible state is guided by underlying program-code and dynamic program-state. This research extends existing model-based GUI testing techniques by modelling interactions between the visible GUI of a GUI-based software and its underlying program-code. The new model is able to, efficiently and effectively, test the GUI in ways that were not possible using existing methods. The thesis is this: Long, useful GUI testcases can be created by examining the interactions between the GUI, of a GUI-based application, and its program-code. To explore this thesis, a model-based GUI testing approach is formulated and evaluated. In this approach, program-code level interactions between GUI event handlers will be examined, modelled and deployed for constructing long GUI testcases. These testcases are able to drive the GUI into states that were not possible using existing models. Implementation and evaluation has been conducted using GUITAR, a fully-automated, open-source GUI testing framework.

Measurement and Simulation of Suppression Effects in a Buoyant Turbulent Line Fire

An experimental and numerical study of turbulent fire suppression is presented. For this work, a novel and canonical facility has been developed, featuring a buoyant, turbulent, methane or propane-fueled diffusion flame suppressed via either nitrogen dilution of the oxidizer or application of a fine water mist. Flames are stabilized on a slot burner surrounded by a co-flowing oxidizer, which allows controlled delivery of either suppressant to achieve a range of conditions from complete combustion through partial and total flame quenching. A minimal supply of pure oxygen is optionally applied along the burner to provide a strengthened flame base that resists liftoff extinction and permits the study of substantially weakened turbulent flames. The carefully designed facility features well-characterized inlet and boundary conditions that are especially amenable to numerical simulation. Non-intrusive diagnostics provide detailed measurements of suppression behavior, yielding insight into the governing suppression processes, and aiding the development and validation of advanced suppression models. Diagnostics include oxidizer composition analysis to determine suppression potential, flame imaging to quantify visible flame structure, luminous and radiative emissions measurements to assess sooting propensity and heat losses, and species-based calorimetry to evaluate global heat release and combustion efficiency. The studied flames experience notable suppression effects, including transition in color from bright yellow to dim blue, expansion in flame height and structural intermittency, and reduction in radiative heat emissions. Still, measurements indicate that the combustion efficiency remains close to unity, and only near the extinction limit do the flames experience an abrupt transition from nearly complete combustion to total extinguishment. Measurements are compared with large eddy simulation results obtained using the Fire Dynamics Simulator, an open-source computational fluid dynamics software package. Comparisons of experimental and simulated results are used to evaluate the performance of available models in predicting fire suppression. Simulations in the present configuration highlight the issue of spurious reignition that is permitted by the classical eddy-dissipation concept for modeling turbulent combustion. To address this issue, simple treatments to prevent spurious reignition are developed and implemented. Simulations incorporating these treatments are shown to produce excellent agreement with the experimentally measured data, including the global combustion efficiency.