Abstract
Limited reading skills are a severe impediment for participation in our information-based society. Automatic text simplification has been suggested as an assistive technology to improve accessibility, but previous research has largely neglected variation between individual users and has suggested an objective notion of what makes text difficult and what does not.
However, as attested by previous research, readers perceive text difficulty individually and subjectively. Text simplification systems that assume general solutions and do not adjust to their individual users therefore cannot provide optimal solutions to the individual user, or by extension to the entire usership. Their potential is bound by the degree to which the target audience displays different simplification needs.
As a response, this thesis presents work that aims to integrate user information into the text simplification workflow, thus personalizing text simplification. This goal is pursued in two ways: (i) making it possible for users to state explicit simplification needs and preferences which the system, trained once on a static dataset, can then focus on at production time, and (ii) enabling a simplification model to learn from high-level user feedback and behavioral data in order to update its beliefs of a user's literacy profile. As an additional line of work, this thesis explores ways to build robust simplification models from limited training data, sharing information between smaller data sources through multi-task learning.
This work marks the first major effort to the development of text simplification systems that integrate information about individual users and adapt to their specific simplification needs. In personalizing text simplification, this user-focused technology can overcome existing upper bounds of performance and improve accessibility for weak readers.
However, as attested by previous research, readers perceive text difficulty individually and subjectively. Text simplification systems that assume general solutions and do not adjust to their individual users therefore cannot provide optimal solutions to the individual user, or by extension to the entire usership. Their potential is bound by the degree to which the target audience displays different simplification needs.
As a response, this thesis presents work that aims to integrate user information into the text simplification workflow, thus personalizing text simplification. This goal is pursued in two ways: (i) making it possible for users to state explicit simplification needs and preferences which the system, trained once on a static dataset, can then focus on at production time, and (ii) enabling a simplification model to learn from high-level user feedback and behavioral data in order to update its beliefs of a user's literacy profile. As an additional line of work, this thesis explores ways to build robust simplification models from limited training data, sharing information between smaller data sources through multi-task learning.
This work marks the first major effort to the development of text simplification systems that integrate information about individual users and adapt to their specific simplification needs. In personalizing text simplification, this user-focused technology can overcome existing upper bounds of performance and improve accessibility for weak readers.
| Original language | English |
|---|
| Publisher | Department of Computer Science, Faculty of Science, University of Copenhagen |
|---|---|
| Publication status | Published - 2018 |