Outline

• The Logos Controlled Vocabulary (LCV)
  Spoke previously about the LCV here at BibleTech 2010. Today i'll focus on how we're mining it for data.
• The LCV corpus
• Learning from the corpus
  • Topic importance
  • References
  • Key terms
  • Topic relationships

Semantic Organization at Logos

... organizing the world's Biblical information

• the Bible Knowledgebase: people, places, things and events described in the Biblical text
  Released Logos 5 last November: using italics for new items since last BibleTech
• the Logos Controlled Vocabulary
• Referent Analysis
• Bible Sense Lexicon
• Timeline
• Preaching Themes

Why Topical Analysis?

• A primary way to explore the text
• Facilitates discovery
• Enables and rewards a large library

The Logos Controlled Vocabulary (LCV)

• LCV = a controlled vocabulary for Biblical studies
  • controlled vocabulary: an organized system for labeling content using words and phrases, to improve search and navigation
  • Supports topic search and exploration, connects other knowledge
  • Mitigates problems of ambiguity, homographs, synonyms, variation in spelling and other kinds of representation
• Provides semantic organization of concepts
  Not keywords
• Organizes content from Bible dictionaries, encyclopedias, and other subject-oriented secondary resources, regardless of how the publisher expressed it.
  LCV as a database of conceptual knowledge is distinct from its use in annotating texts: most of today's talk is about this latter use

LCV Example: Deceit

• Attributes:

  Type: lexical	Name: False
  EN Preferred label: "Deceit"	Extra-biblical: False
  EN Alternate labels: "Deception", "Deceitful", "Deceive", "Connivance", ...	Compound: False
  Importance: 0.4737	Breadth: 0.67

• Relationships:

  Related concepts: none

  Preaching Theme: Honesty

  Logos Topics: http://topics.logos.com/Deceit

Newer LCV Features

• Incorporated Spanish terms and resources
• Added hierarchy (broader/narrower concepts)
• Linked to Preaching Themes, Bible Sense Lexicon
• Disambiguated compound concepts for ambiguous names
• Added Topic Guide to expose more LCV features
• Aligned more resources with the LCV

Connecting Concepts to Terms and References

The LCV Corpus: Overview

• 140k articles (34M words) annotated with concepts
• ~50 "independent" sources
• Varying styles and purposes

Breadth of Article Sources for Concepts

Distribution of Terms for Concepts

Learning Concept → Reference Associations

Ac 18:26	20	2 Ti 4:19	19
Ro 16:3	16	1 Co 16:19	15
Ac 18:18	9	Ac 18:2	8
Ac 18:1-3	3	Ac 18:2-3	3
Ro 16:3-5	3	Ro 16:4	3

42 references, 137 instances

• Content about Biblical subjects typically includes references
  Varies widely depending on the concept as well as the book
• Estimate the association between a concept and references by aggregating counts across all content (voting scheme)
  This also allows us to rank them. Since we normalize references in our markup, they're language independent, so Spanish books help.
• Provide a ranked list of references to explore for a concept
• These estimates should get better as more content is added
  We don't know why a reference is included, but idiosyncrasies tend to wash out with more data. In this example, there are only a few "canonical" references, so the distribution is sharp. That's less true for "fuzzier" concepts.

Learning Reference → Concept Associations

• Reversing this data enables going from a passage to associated concepts
• Not just frequency of association, but distinctiveness
  a concept associated with many passages vs. one associated with only a few
  Overlapping ranges are an equally thorny problem here.
• "Nebai" and "Neh 10:19" are both fairly unique content indicators
• Determining identity can be tricky:
  • Same word? (stemming or morphological analysis)
  • Same reference? (overlapping ranges)
• Word repetition is important evidence for topical relevance: reference repetition is rare.

Learning Term → Concept Associations

• Other high-frequency terms not in this article: "husband", "church", "Christian", "couple", "Ephesus"
• Suggests new concepts that are related. Term → concept data means no dead ends.
• Article context can increase precision.
  Here we still need distinctiveness analysis: Paul is frequent, but throughout. tentmaker is probably more distinctive.
  Many terms are ambiguous absent context.

Concept Relationships through Reference Vectors

• Compare reference vectors to find concepts with similar references
• Example: ointment
  • Top ten ointment references have 133 instances out of 672 (20%)
  • Find other concepts that have Mk 14:3 as a reference
  • Score them against the top ten ointment references

Top Ointment Matches by Reference

• 92 other concepts have Mk 14:3 as a reference

Conclusions

• Developed a large corpus of concept-aligned content
• Demonstrated ways to learn novel associations between terms, references, and concepts
• Future directions:
  • Building systems to generate conceptual associations
  • Collocation analysis
  • Concept annotation of running text