Generating interactive multimedia summaries of evaluative text.

许多组织和个人每天都面临着管理大型文本数据语料库的挑战。一个重要的应用是评价性文本，即表示对一个实体的评价为正面或负面的任何文件。评估信息对决策具有战略意义。例如，许多网站收集了大量消费者对消费电子产品的在线评论。尽管这些文献对产品设计者、规划者和制造商具有很大的战略价值，但有效地处理这些信息仍然是一项复杂和耗时的任务。另一个例子是，许多用户在出国旅游之前，会在预订之前浏览旅行日志和在线用户对酒店的评论。然而，有太多的评估网站，用户被评论淹没。在这些和许多其他类似的情况下，自动化解决方案有可能极大地减少跟上不断增长的评估文献所需的成本和时间。我们不要求用户一次只阅读一条评论，而是建议以多粒度的方式忠实地总结文本评论，以便用户可以根据自己的需求和时间限制选择合适的粒度级别。虽然关于如何从文本中提取评价信息已经做了大量的工作，但很少有人关注如何有效地总结提取的信息并呈现给用户。在以前的工作中，我们已经开始探索这个问题，在这个项目中，我们建议从不同的方向扩展我们的工作。总的来说，我们的目标是消除简单化的假设，增加摘要过程输入的复杂性，并改进其输出的质量。我们方法的一个独特的、与众不同的方面是它是多媒体和交互的。我们的目标是传达用图形提取的所有知识，指出自然语言中最重要的发现，并为从其中提取知识的语料库的交互探索提供支持。

Many organizations and individuals are faced daily with the challenge of managing large corpora of text data. One important application is evaluative text, i.e. any document expressing an evaluation of an entity as either positive or negative. Evaluative information is strategic to decision making.  For instance, many websites collect large quantities of online customer reviews of consumer electronics. Although this literature can be of great strategic value to product designers, planners and manufacturers, the effective processing of this information remains a complex and time-consuming task. As another example, before traveling abroad, many users browse through travel-logs and online user reviews on hotels before they make reservations. However, there are so many evaluation web sites that users are awash in reviews.In these and many other similar situations, an automated solution has the potential to greatly reduce both the cost and time necessary to keep up with the growing amount of evaluative literature. Instead of asking the user to read through one review at a time, we propose to summarize the text reviews faithfully in a multi-granularity fashion, so that the user can pick the appropriate granularity level depending on her needs and time constraints.While there has been substantial work on how evaluative information can be extracted from text, little attention has been paid to how the extracted information can be effectively summarized and presented to users. In previous work we have started to explore this problem and in this project we propose to extend our work in different directions. In general, we aim at removing simplifying assumptions, at increasing the sophistication of the summarization process input, and at improving the quality of its output. A unique, distinguishing aspect of our approach is that it is multimedia and interactive. We aim to convey all the knowledge extracted with graphics, to point out the most important findings in natural language, and also to provide support for the interactive exploration of the corpus from which the knowledge was extracted.