1. Rule-based Reasoning in Semantic Text Analysis
Ivan Rygaev
Laboratory of Computational Linguistics Kharkevich Institute for Information Transmission Problems RAS, Moscow, Russia
RuleML+RR
London, July 2017
Ivan Rygaev | RuleML+RR 2017

2. SemETAP Semantic Text Analyzer
Using Winograd Schemas for Evaluation of Implicit Information Extraction Systems
SemETAP Semantic Text Analyzer
ETAP-3 is a powerful linguistic processor
Syntactic parser
Machine translation
Paraphrasing, UNL etc.
SemETAP semantic analyzer is a part of ETAP-3
Translates an original sentence to a language-independent semantic representation in a formal language.
Applies logical rules (concept definitions) to infer new knowledge.
Ivan Rygaev | RuleML+RR 2017

3. Ivan Rygaev | RuleML+RR 2017
Using Winograd Schemas for Evaluation of Implicit Information Extraction Systems
Text Understanding
The ultimate goal
to achieve near-human understanding of the text
How can we measure understanding?
The amount of inferences that can be made out of the text
How can we test inferences?
Questioning
SemETAP is able to answer questions for which there is no direct answer in the original text
Ivan Rygaev | RuleML+RR 2017

4. Ivan Rygaev | RuleML+RR 2017
Using Winograd Schemas for Evaluation of Implicit Information Extraction Systems
An example
Input sentence:
John sold an umbrella to Peter
Questions (easy for humans):
Who bought the umbrella? (Peter)
What did John give to Peter? (the umbrella)
What did John get? (money)
Who owns the umbrella? (Peter)
Where is the knowledge?
In the meaning of the words sell, buy, give, get and own.
Ivan Rygaev | RuleML+RR 2017

5. Semantic Analysis Steps
Using Winograd Schemas for Evaluation of Implicit Information Extraction Systems
Semantic Analysis Steps
Syntactic Tree
Basic Semantic Structure
Words are translated to semantic concepts and syntactic relations to semantic roles (roughly)
Enhanced Semantic Structure
Concept definitions applied to extend the semantic graph
Ivan Rygaev | RuleML+RR 2017

6. Basic Semantic Structure
Using Winograd Schemas for Evaluation of Implicit Information Extraction Systems
Basic Semantic Structure
hasGivenName (Human_1, "John") hasAgent (Selling_1, Human_1) hasAgent2 (Selling_1, Human_2) hasGivenName (Human_2, "Peter") hasObject (Selling_1, Umbrella_1)
Ivan Rygaev | RuleML+RR 2017

7. Concept Definition Example
Using Winograd Schemas for Evaluation of Implicit Information Extraction Systems
Concept Definition Example
Rule Selling: Selling(?selling) Implication: hasAgent(?selling, ?seller) & Agent(?seller) & hasAgent2(?selling, ?buyer) & Agent(?buyer) & hasObject(?selling, ?thing) & hasPrice(?selling, ?money) & CurrencyMeasure(?money) & Buying(?buying) & hasAgent(?buying, ?buyer) & hasAgent2(?buying, ?seller) & hasObject(?buying, ?thing) & hasPrice(?buying, ?money)
Ivan Rygaev | RuleML+RR 2017

8. Ivan Rygaev | RuleML+RR 2017
Using Winograd Schemas for Evaluation of Implicit Information Extraction Systems
Rules Logic
Selling is Buying with swapped arguments
Buying is Exchange for money
Exchange is two mutual Givings
Giving implies change of the owner
Getting is Giving with slightly different argument roles.
Ivan Rygaev | RuleML+RR 2017

9. Ivan Rygaev | RuleML+RR 2017
Using Winograd Schemas for Evaluation of Implicit Information Extraction Systems
Rule Features
Conjunctive existential rules compatible with Datalog±
Almost necessarily contain new variables in the rule head
Applied in the forward chaining manner (chase)
A number of techniques to guarantee termination
Restricted chase: do not add exactly the same subgraph that already exists
Functional relations simplify checking for existence
Depth limit: do not apply the rule if no variables from the rule head would map to the original individuals from the text
Ivan Rygaev | RuleML+RR 2017

10. Question Processing Steps
Using Winograd Schemas for Evaluation of Implicit Information Extraction Systems
Question Processing Steps
Basic semantic structure of the question is built
Individuals corresponding to wh-words are marked in a special way
The semantic structure is used as a pattern for SPARQL query
The query is run against the enhanced semantic structure of the text
Returned wh-word individuals are displayed to the user
Ivan Rygaev | RuleML+RR 2017

11. Ivan Rygaev | RuleML+RR 2017
Using Winograd Schemas for Evaluation of Implicit Information Extraction Systems
Implementation
Originally we used
Virtuoso DB to store the semantic graphs
SPARQL insert queries to apply the rules
Easy to write but not efficient
Too much overhead in the query execution
Almost no control on the query performance optimization
Recently we switched to RDFox reasoner (Oxford)
Very efficient in-memory RDF storage and reasoner
Optimized for chase
Meets all the requirements mentioned above
Ivan Rygaev | RuleML+RR 2017

12. Ivan Rygaev | RuleML+RR 2017
Using Winograd Schemas for Evaluation of Implicit Information Extraction Systems
Desired Features
Probabilistic uncertainty
Supported: on the level of individuals – epistemic modality
Desired: on the level of propositions/atoms/triples
Disjunctive uncertainty
Desired: allow disjunction in the rule head
Negation
Supported: on the level of individuals – Negation concept
Universal quantifier
Desired: at least in the context of negation
Ivan Rygaev | RuleML+RR 2017

13. Ivan Rygaev | RuleML+RR 2017
Using Winograd Schemas for Evaluation of Implicit Information Extraction Systems
Future Work
Restricted chase is not enough
Depends on the sequence of rule application
Sometimes can create duplicates
Solution: need to look for a noncontradicting subgraph rather than exactly the same one (coreference problem)
Depth limit might be too strong
Can potentially eliminate useful inferences
Solution: use query rewriting to increase the depth
Benchmarking
Such rules can be used to test the reasoners performance
Ivan Rygaev | RuleML+RR 2017

14. Ivan Rygaev | RuleML+RR 2017
Using Winograd Schemas for Evaluation of Implicit Information Extraction Systems
Thank you!
Ivan Rygaev | RuleML+RR 2017