TASK- AND DOMAIN-DEPENDENT NATURAL LANGUAGE PROCESSING FOR SUPPORTING SPOKEN DOCUMENT USE

This dissertation investigates how Natural Language Processing (NLP) can improve users’ access to spoken content in multimedia documents. Information seeking in electronic environments is a highly variable process shaped by information-seeking factors. It is shown that incorporating knowledge about these factors into the design process—task- and domain-dependent (TDD) development—leads to effective Spoken Document Support System (SDSS) prototypes. The presented field studies identify Automatic Speech Recognition (ASR) and Sentiment Analysis (SA) as two requisite NLP technologies for supporting information seekers in the studied domain. The dissertation also demonstrates how SA models can be optimized with TDD feedback taken into account. The described optimization techniques rely on task-based evaluation, which is not commonly used in SA and ASR research. The experiments reported here identify important scenarios in which non-task-based evaluation is not correlated with the outcome of task-based appraisal. Although some task-based evaluation of SA, including the one advocated here, is no more resource-intensive than calculating intrinsic measures, others can be more expensive if they involve human-subject experimentation. This work also introduces a semi-automatic TDD measure of ASR quality that simulates human-subject experimentation, minimizing resource consumption without sacrificing empirical grounding.

Department dc:contributor.department

Computer Science

Year dc:date.issued

2018

Author dc:creator

• Kazemian, Siavash

Advisor dc:contributor.advisor

• Penn, Gerald

Handle dc:identifier.uri

http://hdl.handle.net/1807/89897

OAI identifier oai:identifier

oai:utoronto.scholaris.ca:1807/89897