| Popis: |
The number of books available online is increasing, but user interfaces may not be taking full advantage of advances in machine learning techniques that could help users navigate, explore, discover and understand interesting and useful content in books. Using a group of ten students and over one thousand crowdsourced judgments, we conducted multiple user studies to evaluate topics and related passages in books, all learned by topic modeling. Using ten books, selected from humanities (e.g. Plato's Republic), social sciences (e.g. Marx's Capital) and sciences (e.g. Einstein's Relativity), and four different evaluation experiments, we show that users agree that the learned topics are coherent and important to the book, and related to the automatically generated passages. We show how crowdsourced evaluations are useful, and can complement more focused evaluations using students who have studied the texts. This work provides a framework for (1) learning topics and related passages in books, and (2) evaluating those learned topics and passages, and moves one step toward automatic annotation to support topic navigation of books. |
