Concept graphS
FOR
Learning and SEARCH
Ashish Dubey
B.Tech CSE, 2013-14
What is a concept graph?
Representation of semantic relationships in natural language text.
- Concepts in a document become nodes
- Relationships become the edges connection concepts.
Example
Extracted from an article about tree. Clearly shows the interesting attributes.
HOW CAN THEY be USED?
- Easy understanding of an article.
- Concept discovery
- Helps in ideation
- Natural language search - answer finding
Existing VARIANTS
- Google's Knowledge Graph
- Facebook Graph Search
- Luma7
- Mindmap tools
project aim
- Build a concept graph engine
- Should be re-usable
- Build applications on top of it - Visualization and Search
existing work
- Lot of manual tools for concept map construction
- Not many automatic tools
- No re-usable frameworks
EXISTING RESEARCH
-
A survey of concept map mining techniques by Zubrinic, K. et al (2012)
A semi-automatic concept map extraction and evaluation framework by Jorge J. Villalon and Rafael A. Calvo (2010)
APPROACH
- Use efficient information extraction techniques for extraction of concepts and relationships.
- Store the extracted concepts and relations.
- Use the data to build apps like search.
Information extraction
- The most challenging aspect of the project
- Overall efficiency relies on the underlying techniques
- Couple of state of the art NLP libraries like OpenNLP, StanfordCoreNLP, etc
- No single library satisfies the complete requirement
INFORMATION EXtraction - cont
- <subject - relation - predicate> can be treated as concept - <relation - concept - triple>
- Options:
- Syntactic extraction
- POS tags
- NLTK
- Statistical parsers
- Treebanks
- MaltParser, StanfordParser
- Challenges:
- Anaphora resolution, enablers, etc
Information extraction - CONT
Solutions to the problem:
- Information extraction libraries by NLP scientists at UWa like OLLIE
- Based on MaltParser
- Outputs relationship triples
- Anaphora resolution:
- StanfordCoreNLP's coreference resolution system
Post-extraction
- Concepts and relations are stored in a graph database
- Neo4J
- Data exposed through a web API
- Application front-ends can interact with the API to get data.
Concept visualization app
(DEMO)
FUTURE: SEMANTIC SeARCH
- Natural language queries - often questions
- Extraction of answers from the concept graph
- Search modules
- Query processing
- Graph data adapters
- Ranking of results
Thank you
titled
By Ashish Dubey
titled
- 1,087