CRII:RI:Toward Socially-diverse Multi-document Summarization

点击翻译按钮获取中文摘要

User-generated social data such as microblogs, discussion posts, and product reviews often contain rich expressions of opinions and thoughts, near real-time information, and diverse perspectives on a wide range of topics from health to socio-political movements. Such data represents many unique perspectives originating from people from many different walks of life and belonging to different social groups. Accessing such data is relatively easy; however, extracting important and meaningful information to get a holistic view from such an overwhelming volume of unstructured, noisy, and redundant data remains extremely challenging. Automatic text summarization models are important computational tools that enable identification of salient pieces of text by condensing several documents into a short, concise summary, with most algorithms focusing on improving the quality of the summaries along dimensions of fluency, factuality, coherence and so on. However, when it comes to summarizing socially diverse data, an important yet overlooked dimension of quality is the breadth of diverse perspectives that a summary encapsulates. This research aims to improve text summarization algorithms for socially rich data and is specifically motivated by applications in natural language processing.This project will leverage social diversity, a unique characteristic of social data, for improving the overall quality of multi-document summaries. The technical aims of the project are divided into two thrusts. The first thrust develops a suite of novel datasets of summaries corresponding to multiple documents from different socially salient groups that can serve as reliable benchmarks for model training and evaluation. The second thrust develops new unsupervised and few-shot learning algorithms for generating high quality extractive and abstractive summaries reflecting a broad spectrum of diverse perspectives. This thrust will develop new methods for integrating diversity constraints into summarization algorithms as well as jointly optimizing summary’s textual quality and social diversity. The novel methods will be complemented by a comprehensive evaluation plan including several summarization metrics of automatic and human evaluation. This research effort will create cross-disciplinary opportunities for bringing together researchers and students from language technology and fairness communities to develop new datasets and algorithms for generating socially diverse summaries.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.