What’s in an Index: Extracting Domain-specific Knowledge Graphs from Textbooks

Abstract

A typical index at the end of a textbook contains a manuallyprovided vocabulary of terms related to the content of the textbook. In this paper, we extend our previous work on extraction of knowledge models from digital textbooks. We are taking a more critical look at the content of a textbook index and present a mechanism for classifying index terms according to their domain specificity: a core domain concept, an in-domain concept, a concept from a related domain, and a concept from a foreign domain. We link the extracted models to DBpedia and leverage the aggregated linguistic and structural information from textbooks and DBpedia to construct and prune the domain-specific knowledge graphs. The evaluation experiments demonstrate (1) the ability of the approach to identify (with high accuracy) different levels of domain specificity for automatically extracted concepts, (2) its cross-domain robustness, and (3) the added value of the domain specificity information. These results clearly indicate the improved quality of the refined knowledge graphs and widen their potential applicability.

Extraction of Knowledge Models from Textbooks

“…For example, the term cox's theorem 14 is a statistical theorem, it is not part of any of the ten textbooks, but it belongs to the target domain. Using Google Books, this term was found in three textbooks with particular topics: uncertainty theory 15 , statistical evidence measurement 16 , and universal artificial intelligence 17 . All three books mention the term, but it appears only in the index section of the last one, which illustrates the rarity of the term.…”

“…A set of 40 textbooks on university-level Statistics written in English, available in both PDF and EPUB versions, and containing Tables of Content and Index sections have been selected. Additionally, to validate the approach in other domains, five textbooks have been added to the test set for each of the following domains: computer science, history, and literature 9 . This experiment has focused on three different outcomes from the first three stages of the approach (the last stage is the construction of the textbook knowledge model itself).…”

“…Additionally, they have been automatically mapped to the ISI Multilingual Glossary of Statistical Terms (ISI Glossary) 11 . The ISI Glossary translates more than 3500 statistical terms (and synonyms) into 31 languages 12 . As a result, the two original models have formed a unified reference set of 1611 terms.…”

“…The heading and subheading are nouns or noun phrases that represent one unit of knowledge relevant to the textbook 7 , the locators are the associated page references, and crossreferences are references to other index terms. LOCATOR_ELEMENT identifies five different types of page references: single roman number (e.g., V), range of roman numbers (e.g., V-VIII), single page number (e.g., 45), range of page numbers (e.g., [15][16][17], and endnotes (e.g., 25N3). Each type is first identified using a dedicated regular expression, and then normalized to a sequence of page numbers: roman number references are discarded, all the page numbers from the ranges are generated, and note numbers are discarded.…”

“…Using the information from E, it is possible to enrich the original textbook model to incorporate additional semantic information and create connections to the LOD using RDF relationships. Since the TEI Guidelines do not provide a mechanism to integrate RDF with TEI 11 , RDFa 12 attributes are added to the TEI attribute class att.global.linking: @about, @property and @resource [235,264]. Those attributes allow for representing RDF tuples: @about for the subject, @property predicate and @resource object.…”