Principled Approaches to Automatic Text Summarization

Automatic text summarization is a particularly challenging Natural Language Processing (NLP) task involving natural language understanding, content selection and natural language generation. In this thesis, we concentrate on the content selection aspect, the inherent problem of summarization which is controlled by the notion of information Importance. We present a simple and intuitive formulation of the summarization task as two components: a summary scoring function θ measuring how good a text is as a summary of the given sources, and an optimization technique O extracting a summary with a high score according to θ. This perspective offers interesting insights over previous summarization efforts and allows us to pinpoint promising research directions. In particular, we realize that previous works heavily constrained the summary scoring function in order to solve convenient optimization problems (e.g., Integer Linear Programming). We question this assumption and demonstrate that General Purpose Optimization (GPO) techniques like genetic algorithms are practical. These GPOs do not require mathematical properties from the objective function and, thus, the summary scoring function can be relieved from its previously imposed constraints. Additionally, the summary scoring function can be evaluated on its own based on its ability to correlate with humans. This offers a principled way of examining the inner workings of summarization systems and complements the traditional evaluations of the extracted summaries. In fact, evaluation metrics are also summary scoring functions which should correlate well with humans. Thus, the two main challenges of summarization, the evaluation and the development of summarizers, are unified within the same setup: discovering strong summary scoring functions. Hence, we investigated ways of uncovering such functions. First, we conducted an empirical study of learning the summary scoring function from data. The results show that an unconstrained summary scoring function is better able to correlate with humans. Furthermore, an unconstrained summary scoring function optimized approximately with GPO extracts better summaries than a constrained summary scoring function optimized exactly with, e.g., ILP. Along the way, we proposed techniques to leverage the small and biased human judgment datasets. Additionally, we released a new evaluation metric explicitly trained to maximize its correlation with humans. Second, we developed a theoretical formulation of the notion of Importance. In a framework rooted in information theory, we defined the quantities: Redundancy, Relevance and Informativeness. Importance arises as the notion unifying these concepts. More generally, Importance is the measure that guides which choices to make when information must be discarded. Finally, evaluation remains an open-problem with a massive impact on summarization progress. Thus, we conducted experiments on available human judgment datasets commonly used to compare evaluation metrics. We discovered that these datasets do not cover the high-quality range in which summarization systems and evaluation metrics operate. This motivates efforts to collect human judgments for high-scoring summaries as this would be necessary to settle the debate over which metric to use. This would also be greatly beneficial for improving summarization systems and metrics alike.