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.

DEVELOPMENT OF AN OFF-LINE RAINFLOW COUNTING ALGORITHM WITH TEMPORAL PARAMETERS AND DATA REDUCTION

Rainflow counting methods convert a complex load time history into a set of load reversals for use in fatigue damage modeling. Rainflow counting methods were originally developed to assess fatigue damage associated with mechanical cycling where creep of the material under load was not considered to be a significant contributor to failure. However, creep is a significant factor in some cyclic loading cases such as solder interconnects under temperature cycling. In this case, fatigue life models require the dwell time to account for stress relaxation and creep. This study develops a new version of the multi-parameter rainflow counting algorithm that provides a range-based dwell time estimation for use with time-dependent fatigue damage models. To show the applicability, the method is used to calculate the life of solder joints under a complex thermal cycling regime and is verified by experimental testing. An additional algorithm is developed in this study to provide data reduction in the results of the rainflow counting. This algorithm uses a damage model and a statistical test to determine which of the resultant cycles are statistically insignificant to a given confidence level. This makes the resulting data file to be smaller, and for a simplified load history to be reconstructed.

Essays on Information Frictions, Prices and Unemployment

I investigate the effects of information frictions in price setting decisions. I show that firms' output prices and wages are less sensitive to aggregate economic conditions when firms and workers cannot perfectly understand (or know) the aggregate state of the economy. Prices and wages respond with a lag to aggregate innovations because agents learn slowly about those changes, and this delayed adjustment in prices makes output and unemployment more sensitive to aggregate shocks. In the first chapter of this dissertation, I show that workers' noisy information about the state of the economy help us to explain why real wages are sluggish. In the context of a search and matching model, wages do not immediately respond to a positive aggregate shock because workers do not (yet) have enough information to demand higher wages. This increases firms' incentives to post more vacancies, and it makes unemployment volatile and sensitive to aggregate shocks. This mechanism is robust to two major criticisms of existing theories of sluggish wages and volatile unemployment: the flexibility of wages for new hires and the cyclicality of the opportunity cost of employment. Calibrated to U.S. data, the model explains 60% of the overall unemployment volatility. Consistent with empirical evidence, the response of unemployment to TFP shocks predicted by my model is large, hump-shaped, and peaks one year after the TFP shock, while the response of the aggregate wage is weak and delayed, peaking after two years. In the second chapter of this dissertation, I study the role of information frictions and inventories in firms' price setting decisions in the context of a monetary model. In this model, intermediate goods firms accumulate output inventories, observe aggregate variables with one period lag, and observe their nominal input prices and demand at all times. Firms face idiosyncratic shocks and cannot perfectly infer the state of nature. After a contractionary nominal shock, nominal input prices go down, and firms accumulate inventories because they perceive some positive probability that the nominal price decline is due to a good productivity shock. This prevents firms' prices from decreasing and makes current profits, households' income, and aggregate demand go down. According to my model simulations, a 1% decrease in the money growth rate causes output to decline 0.17% in the first quarter and 0.38% in the second followed by a slow recovery to the steady state. Contractionary nominal shocks also have significant effects on total investment, which remains 1% below the steady state for the first 6 quarters.