Natural Intelligence ICCS 2010

Natural Intelligence -

Commonsense Question Answering

with Conceptual Graphs

Fatih Mehmet Güler and Aysenur BirturkDepartment of Computer Engineering, METU

06531, Ankara/TURKEY

[email protected]

[email protected]

Motivation

Massive Knowledge found as Natural Language

Text based Question Answering (no tagging)

Open Domain Question Answering

Address Commonsense Reasoning Problem

Linguistically motivated KRR

Intelligence is the accumulation of knowledge

Integrate State of the Art Tools

Ultimate goal: Getting closer to strong AI

Summary of the System

Natural Language is parsed

Utterances are represented using CGs

Concepts and Relation types are mapped to

Cyc equivalent counterparts

Type hierarchies are computed

Knowledge is accumulated

If the input is a question

Search for answer (projection)

Summary of the System (Cont’d)

NI

NLP

CCG

C&C Tools

KRR

CGs

Cogitant

Commonsense

Open Cyc

Background

Combinatory Categorial Grammar (CCG)

C&C Tools

Conceptual Graphs

Cogitant

Open Cyc

Combinatory Categorial Grammar (CCG)

Lexicalized Theory of Grammar based on

Categorial Grammar ( Steedman 2001).

Functions can be applied or composed

Arguments can be picked up or turned into

functors (Type raising)

Easy for Semantic Representations

Small number of semantically transparent

combinatory rules to combine CCG categories.

Assign semantic representations to the lexical entries

Interpret combinatory rules

CCG parse for “Mr. Hyde ate two

lemmons”

CCG Parse for “Susan knows that Bob

likes Fred”

DRS for “Susan knows that Bob likes

Fred”

C&C Tools

Linguistically Motivated Large-Scale NLP with C&C

and Boxer. (Curran, Clark, Bos, 2007)

C&C Parser

POS Tagging, Supertagging

Parsing, Chunking

Named Entity Recognition

Boxer

Uses CCG parser output

Generates DRS Semantic Representations

Freely available for research

http://svn.ask.it.usyd.edu.au/trac/candc/wiki

C&C Tools

Large Scale NLP is possible with C&C and

Boxer

C&C Parser: state of the art parser for CCG

Boxer: Semantic representations in DRS

Open Cyc

Open source version of Cyc system

Cyc: greatest effort to encode Common Sense

knowledge in machine processable way

500.000 concepts 26.000 relations and 5.000.000

assertions

CycL language similar to Lisp

We use Cyc to map parsed words to common sense

counterparts such as person to #$Person

(disambiguation)

Open Cyc (cont’d)

(#$likesAsFriend #$GeorgeWBush #$AlGore)

#$isa, #$genls

(#$isa #$GeorgeWBush

#$UnitedStatesPresident)

(#$genls #$UnitedStatesPresident #$Person)

Cogitant

Library for Conceptual Graph operations

Supports broad CG operations (Genest &

Salvat, 1998)

Graph representation

Conversion from CGIF

Projection checking

Rule application

Natural Intelligence – Commonsense

Question Answering with CGs

Augment Common Sense knowledge

Modular Approach

Separation of Concerns

State of the art tools

Architecture - Modules

Natural Language Processing (C&C Tools are used for implementation) Convert natural language to CGIF

Reasoning (Cogitant library is used for implementation) CG operations

Common Sense (Open Cyc is used for implementation) Common sense mapping

Storage (Conceptual Graphs are stored in a database) Persistence of CGs

System Definition

User enters a sentence from web interface;

This sentence is converted to CGIF using the NLP module;

CGIF is converted to CGs using the reasoning module;

Support is generated to CGs using the common sense module;

Common sense rules gathered from common sense module are applied to CGs using reasoning module;

CGs are merged to the previous ones using reasoning module;

If the input sentence is a question sentence, same operations take place, except the resulting graph is used to query existing CGs using the reasoning service, and if there are projections from this query graph to previous CGs, results are displayed to the user;

CGs are persisted using the storage module.

Common Sense Mapping

Cyc: (prettyString TERM STRING)

Chain up to #$Thing using #$genls relations

Same for relations using #$genlPreds

Relation hierarchies are converted to forward

rules

#$performedBy -> #$temporallyRelated

Sample Concept Hierarchy

#$Place ->

#$EnduringThing-Localized ->

#$Location-Underspecified ->

#$Thing ->

#$SomethingExisting ->

#$Individual ->

#$Thing ^^

#$Trajector-Underspecified ->

#$Location-Underspecified ^^

#$TemporallyExistingThing ->

#$TemporalThing ->

#$Individual ^^

#$SpatialThing-Localized ->

#$TemporallyExistingThing ^^

#$SpatialThing ->

#$Individual ^^

#$Boundary-Underspecified ->

#$Region-Underspecified ->

#$Location-Underspecified ^^

#$Landmark-Underspecified ->

#$Individual ^^

#$Location-Underspecified ^^

#$SpatialThing-NonSituational ->

#$SpatialThing ^^

#$Individual ^^

#$Location-Underspecified ^^

Conversion to Cogitant Support

Convert Cyc hierarcy to Cogitant support

format

Concept Types

Relation Types

Individuals

Rules

Convert assertions to Cogitant graph format

Apply forward rules

Answering Queries

Significance

Sentences like;

What are the intangible things in this situation?

Was Mr. Hyde there while eating the apples?

Does Mr. Hyde exist after eating the apples?

Do the apples exist after Mr. Hyde ate them?

Deep Natural Language Understanding

State of the art tools

Open domain question answering

Difficulties

Open Cyc API is broken

Does not work in Turkish locale (fixes are sent to

maintainers)

Still, provided API sends one IP packet per character, way

too slow over network

Custom socket API is developed and used over TCP

Custom Lisp functions for generalization hierarchy and

concept mapping

Cogitant problematic

Java API is very limited (compared to C++)

Only works over XML files

Conclusion

Central Integrated Common Sense QAS

CCG for Natural Language Processing

Conceptual Graphs for KRR

Cyc for Common Sense

Future Work

Implement Rule Induction

Backward Chaining (Resolution)

Improve NLP module and Common Sense mapping

Probabilistic Reasoning

Question Answering System (QAS) to be used in;

Education (Learning Management Systems)

Semantic Search (Content Management Systems)

Intelligent Help