Cross-referencing Using Fine-grained Topic Modeling

Proceedings of NAACL-HLT 2019, pages 3978–3987 Minneapolis, Minnesota, June 2 - June 7, 2019. c⃝2019 Association for Computational Linguistics 3978 Cross-referencing using Fine-grained Topic Modeling Jeffrey Lund, Piper Armstrong, Wilson Fearn, Stephen Cowley, Emily Hales, Kevin Seppi Computer Science Department Brigham Young University {jefflund, piper.armstrong, wfearn, scowley4, emilyhales, kseppi}@byu.edu Abstract Cross-referencing, which links passages of text to other related passages, can be a valu- able study aid for facilitating comprehension of a text. However, cross-referencing requires first, a comprehensive thematic knowledge of the entire corpus, and second, a focused search through the corpus specifically to find such useful connections. Due to this, cross- reference resources are prohibitively expen- sive and exist only for the most well-studied texts (e.g. religious texts). We develop a topic-based system for automatically produc- ing candidate cross-references which can be easily verified by human annotators. Our sys- tem utilizes fine-grained topic modeling with thousands of highly nuanced and specific top- ics to identify verse pairs which are topically related. We demonstrate that our system can be cost effective compared to having annota- tors acquire the expertise necessary to produce cross-reference resources unaided. 1 Introduction Cross-references are references within or between bodies of text that can help elaborate upon or clar- ify that text. They can be a useful tool for deep understanding and can also be used to analyze the relational structure of a text. In contrast to a word concordance which simply shows passages which share a common keyword, cross-references often include links which do not necessarily share the same keywords, but are still related topically. The existence of a thorough and complete cross- reference resource can facilitate better scholarship of a text and help readers to quickly find clarify- ing information or see repeated themes throughout a text. Compared to language tasks such as part-of- speech tagging, producing cross-reference anno- tations is much more labor intensive. To pro- duce cross-references, annotators must become in- timately familiar with a text in order to note that a particular passage is related to another passage they happen to recall. This level of familiarity and expertise with a particular text typically re- quires the annotator to spend a great deal of time studying and reading. Possibly for this reason, ex- pansive cross-references have only been produced for the most well-studied texts, such as the Bible. Other texts, such as academic textbooks, may in- clude indices or other similar references, but these tend to be sparse, focusing on a small number of keywords rather than linking each individual pas- sage with other relevant passages. The process of creating such a resource can be expensive and time consuming. For exam- ple, the Bible1 published by The Church of Je- sus Christ of Latter-day Saints includes numer- ous cross-references and topic-based categories. These cross-references took hundreds of volun- teers thousands of hours over seven years to pro- duce (Anderson, 1979). This process involved collecting more than 19,900 manually curated en- tries from volunteers, and then editing and refining those references with …

Content truncated to 500 words