Global Inference for Summarization Using Integer Linear Programming

摘要是将源文本压缩成较短的版本，同时保留其信息内容的过程。摘要的应用是多种多样的。从快速访问新闻和科学文章到帮助医生收集患者信息和满足浏览器的系统。人类每天都会毫不费力地总结（例如，通过描述讲座、会议或电影的内容），但是自动生成高质量的摘要仍然是一个挑战。困难主要在于任务的性质是复杂的，必须满足许多限制（例如，摘要长度、信息量、连贯性、语法性），并最终需要大规模的文本理解。由于强大的文本理解超出了当前NLP技术的能力，因此今天的大多数工作都集中在提取摘要上。这里的想法是通过识别并随后连接文档中最重要的句子来创建摘要。不需要大量的语言分析，就可以为各种文档创建摘要，而不受风格、文本类型和主题的影响。不幸的是，摘录通常是可读性和文本质量较低的文档。在这个项目中，我们将开发新的单文档摘要模型，脱离句子提取范式。我们将把摘要建模为一个优化问题，并使用整数线性规划（ILP）来找到最适合当前应用程序、任务或用户的摘要。ILP制剂由于两个原因是有利的。首先，它允许我们显式地编码输出摘要必须满足的约束。其次，ILP是一个研究得很好的优化问题，具有在存在许多冲突约束的情况下找到全局最优解的有效算法。该提案旨在通过开发基于ILP框架的新颖且统一的模型来改变摘要范式，该模型能够识别文档中的重要内容并适当地表达。本研究的成功将对文摘及相关领域（如，信息检索），这是通过逐步扩展传统模型无法实现的。

Summarization is the process of condensing a source text into a shorter version while preserving its information content. The applications of summarization are many and varied. From quick access to news and scientific articles to systems that aid physicians in gathering patient information and meeting browsers. Humans summarize on a daily basis and effortlessly (e.g., by describing the contents of a lecture, a meeting or a movie), but producing high quality summaries automatically remains a challenge. The difficulty lies primarily in the nature of the task which is complex, must satisfy many constraints (e.g., summary length, informativeness, coherence, grammaticality) and ultimately requires large-scale text understanding. Since robust text understanding is beyond the capabilities of current NLP technology, most work today focuses on extractive summarization. The idea here is to create a summary simply by identifying and subsequently concatenating the most important sentences in a document. Without a great deal of linguistic analysis, it is possible to create summaries for a wide range of documents, independently of style, text type, and subject matter. Unfortunately, extracts are often documents of low readability and text quality. In this project we will develop novel models for single-document summarization that break away from the sentence extraction paradigm. We will model summarization as an optimisation problem and use integer linear programming (ILP) for finding a summary that is best for the application, task, or user at hand. The ILP formulation is advantageous for two reasons. First, it allows us to explicitly encode the constraints our output summaries must meet. Secondly, ILP is a well studied optimization problem with efficient algorithms for finding a globally optimal solution in the presence of many conflicting constraints. This proposal aims to shift the summarization paradigm by developing novel and unified models based on the ILP framework that are able to identify what is important in a document and express it appropriately. The success of this research will make significant and far-reaching impact on summarization and related areas (e.g., information retrieval) that could not be brought about by incrementally extending conventional models.