Unsupervised Relation Extraction From Web

Vempati Anurag Sai
Bhavishya Mittal
Supervisor: Dr. Amitabha Mukerjee

Abstract:

Extracting knowledge from web is an important NLP problem. Since any information worthwhile is mostly of the SPO (Subject, Predicate and Object) form, we will be aiming to extract such tuples from the huge web data available today. We also aim to build a query module that can search for the most relevant tuple and return it. The framework involves three main modules: A relation extractor, Learning Module and a Query Module. Dependency graph obtained from the Stanford parser is used to extract the tuples. An SVM is trained to learn to distinguish trustworthy tuples from the non-trustworthy to reduce the computational complexity involved in using a parser on each and every sentence. The classifier also helps in removal of noise which crept in due to the parser. Similar tuples are grouped together for obtaining a frequency count. During the query process, all the tuples with certain amount of overlap with the query are returned and ranked according to their relevance to the query.

Some tuples extracted by the "relation exractor" module for input sentence "The American Civil War, also known as the War between the States or simply the Civil War, was a civil war fought from 1861 to 1865 in the United States after several Southern slave states declared their secession and formed the Confederate States of America":

Trusted and Untrusted tuples extracted by the "relation exractor"
Click on the images for expanded view

Some of the trsutworthy tuples identified by the classifier after training on the AIMed corpus with 1955 sentences from MEDLINE ab- stracts annotated with gene/protein names and protein-protein interactions:

('Injections', 'transmit', 'infections')
('green tea', 'had', 'significant chemopreventive effects')
('zinc deficiency', 'affects', 'cognitive development')
('Exposure to paclitaxel for 8 h or less', 'produced', 'biphasic survival curves')
('a low fat diet', 'reduces', 'DDT bioaccumulation')
('Raman', 'comprised', 'a near infrared laser')
('asthmatic patients', 'contained', 'high levels of IgE antibody')
('poxvirus', 'causes', 'tumors')
('The gastrointestinal tract of numerous animal species', 'contains', 'melatonin')
('osteoblasts', 'secrete', 'the primary endocrine source')
('A58', 'disrupts', 'reverse transcription')
('PPAR alpha agonists', 'repress', 'human fibrinogen gene expression')
('Human coronary artery', 'expressed', 'MCP-1 mRNA')
('dihydropyrimidine dehydrogenase', 'affect', 'tumor sensitivity and resistance')
('the use of dexamethasone', 'reduce', 'delivery of chemotherapeutic agents')
('The beta-lactamase characterized from strain SLO74', 'named', 'OCH-1')

Some of the top results for the query WHAT energize kinases. Notice that the system returned tuples with 'stimulate' in the relation field because it's a synonym of 'energize':

Query results
Click on the images for expanded view

Some of the top results for the query WHAT cause WHAT. Notice that the system returned tuples with 'induce', 'makes' and 'have' in the relation field because they are all synonyms of 'cause':

Query results
Click on the images for expanded view

Documentation:

PROPOSAL    PRESENTATION    POSTER    FINAL REPORT    CODES

REFERENCES:

  1. Banko, Michele, et al. "Open Information Extraction from the Web." IJCAI. Vol. 7. 2007.
  2. Fader, Anthony, Stephen Soderland, and Oren Etzioni. Identifying relations for open information extraction. Proceedings of the Conference on Empirical Methods in Natural Language Processing. Association for Computational Linguistics, 2011.
  3. Dan Klein and Christopher D. Manning. 2003. Accurate Unlexicalized Parsing. Proceedings of the 41st Meeting of the Association for Computational Linguistics, pp. 423-430.
  4. Marie-Catherine de Marneffe, Bill MacCartney and Christopher D. Manning. 2006. Generating Typed Dependency Parses from Phrase Structure Parses. In LREC 2006.
  5. Jython libraries for Stanford Parser by Viktor Pekar. https://github.com/vpekar/stanford-parser-in-jython
  6. Python implementation of Dijkstras algorithm by David Eppstein UC Irvine, 4 April 2002. http://code.activestate.com/recipes/577343-dijkstras-algorithm-for-shortest-paths
  7. ftp://ftp.cs.utexas.edu/pub/mooney/bio-data/ courtesy: Raymond J.Mooney, University of Texas at Austin