Generation of textual summaries at different target reading levels: summarizing line graphs for visually impaired users
Date
2016
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
University of Delaware
Abstract
This work is concerned with the generation of text at different reading levels by tailoring the generated text to fit the reading level that is appropriate to the reader. The technique is employed in the context of conveying the high-level messages of information graphics in online popular media in order to allow access to such media by people who are blind or visually impaired, as well as by systems with limitations on screen sizes or bandwidth where images are not convenient. The contributions of this work aim to avoid commonly placed rule-based methodologies and to improve different phases of the NLG pipeline.
The methodologies and techniques proposed by this work were employed in the context of the SIGHT system, which provided textual summaries of simple bar charts. In this thesis, we handle single line graphs and have made significant contributions to several modules of the NLG pipeline, including: content determination, text structuring, aggregation and lexicalization.
Texts in popular media are written to target readers at different reading levels – some of the text is rather simple (geared toward 4th grade readers, for example), while other text is quite sophisticated (geared toward college-level readers, for example). We found that text that was geared toward a reading level that did not match the reader was difficult for that reader to understand. Thus, we attempt to produce a summary whose writing sophistication matches that of the article in which the graphic appears. Two of the phases of the NLG system are crucial to achieving generation at different reading levels: aggregation and lexicalization.
The methodologies and techniques developed in the context of this work were evaluated by generating summaries of line graphs present in online popular media. Summaries generated at different reading levels were evaluated by both automatic reading level assessment tools and by readers at different reading levels. The appropriateness of the summary context was evaluated by people with visual impairments who were asked to answer important questions about the presented line graphs.