Effective data curation for frequently asked questions

摘要

Frequently-asked-question (FAQ) systems are effective in operating and reducing costs of IT services. Basically, FAQ data preparation requires data curation of available heterogeneous question-and-answer (QA) data sets and creating FAQ clusters. We identified that the labor intensiveness of data curation is a major problem and that it strongly affects the final FAQ output quality. To deal with this problem, we designed a FAQ creation system with a strong focus on the effectiveness of its data-curation component. We conducted a field study by inspecting two sources: incident reports and a QA forum. The first source of incident reports showed a high F-score of 89.9% (precision: 82.5%, recall: 100%). We also applied the same set of parameters to 300 entries of the QA forum and achieved an F-score of 94.3% (precision: 94.9%, recall: 93.8%).