Integrating Methods for Mining and Summarizing Text Conversations

由于互联网革命，大量的个人和组织不断从事电子邮件 交流、博客、短信和其他社交媒体活动。与此同时， 技术实现了许多口头对话（例如，面对面的会议）进行自动转录。这些趋势的最终结果是，人类的书面形式的对话数据正在以惊人的速度积累。 虽然已经提出了几种计算方法来分析，挖掘和总结这些非正式的，会话的文件，他们一直在四个关键方面的限制。首先，大多数关于摘要会话的研究都集中在提取系统上，其中摘要只是输入会话中的话语的子集。其次，大多数提出的方法主要是为会议和电子邮件开发的，很少有其他会话模式的工作。第三，这些方法大多是有监督的，需要在标注语料库上进行训练。最后，目前的会话挖掘技术是非常特定于任务的，例如主题建模，对话行为建模和提取会话结构的方法在很大程度上是独立开发的。 在本提案中，我们计划通过开发以下内容来解决所有这些限制： - 摘要系统将越来越抽象，更紧密地反映人们自然产生和期望的内容。 - 将数据和见解从一种对话方式转移到另一种对话方式的方法;例如，从会议和电子邮件到博客和微博。 - 半监督和无监督机器学习技术，可以通过利用大量现成的未标记会话数据轻松应用于博客和推文。 一个用于集成不同文本挖掘任务的框架。特别地，我们将研究对话如何 建模、主题建模和情感分析可以同时且相互依赖地执行。

Due to the Internet Revolution, large numbers of individuals and organizations constantly engage in email exchanges, blogging, texting and other social media activities. At the same time, improvements in speech technology enable many spoken conversations (e.g., face -to-face meetings) to be automatically transcribed. The net result of these trends is that human conversational data-in written forms-are accumulating at a phenomenal rate. Although several computational methods have been proposed to analyze, mine and summarize these informal, conversational documents, they have been limited in four key ways. First, the majority of research on summarizing conversations has been focusing on extractive systems, where a summary is simply a subset of the utterances in the input conversation. Second, most of the proposed methods have been mainly developed for meetings and emails, with little work on other conversational modalities. Third, most of these methods are supervised ones, which need to be trained on annotated corpora. Finally, current conversation mining techniques are very task-specific; for instance methods for topic modeling, for dialog act modeling and for extracting the conversational structure have been developed largely independently. In this proposal, we plan to address all these limitations by developing: - Summarization systems that will be increasingly abstractive, reflecting more closely what people naturally produce and expect. - Methods to transfer data and insights from one conversational modality to another; for instance, from meetings and emails to blogs and micro-blogs. -Semi-supervised and unsupervised machine learning techniques that can be easily applied to blogs and tweets by leveraging large amounts of readily available unlabeled conversational data. -A framework for integrating the different text mining tasks. In particular, we will study how dialog act modeling, topic modeling and sentiment analysis can be performed simultaneously and interdependently.