18

The Concept Convert free text into a structure that captures objects, relations, existential, propositional and temporal logic – the whole thing, no excuses

Embed Size (px)

Citation preview

The Concept

• Convert free text into a structure that captures objects, relations, existential, propositional and temporal logic – the whole thing, no excuses.

• Resolve the anaphora.• Build a query in the same structure using free

text• Run the query against the structure using

structure matching and Constraint Reasoning

A Query Is Built from Free Text

Note the logical spine binding all the relations in the sentence

Close Integration

The search query is built by the same text to structure converter used on the source text

with the same abilities - multiple sentences in a discourse- anaphora resolution so the query can be built

up naturally- switching of interrogative form into

declarative form so there will be a direct match

- use of knowledge already gained as reading proceeds

The Variables Are Made Free

Enron is fixed

The Smallest Start Set Is Generated

A good place to start

Search Behaviour – Use What Is There

This relation searches for other ToOwn relations which match it in terms of its parameters – it finds what is already there. ToOwn happens to be transitive, so it expands from its start point – that is, if A owns B and B owns C then A owns C

Propagation Continues

Searching Behaviour – Go Active

The Detail relation takes its input, and then finds everything connected to the input to create its output – it searches actively around what it is given, rather than being limited to children

Search Behaviour - Chameleon

This relation finds what is there, but must become every other possible relation in the process of finding the information –

• It was omitted• The report doesn’t mention it• We didn’t include it• I wanted to put it in, but Fred said no

Getting From One Relation To Another

SynonymMaps, synonyms and antonyms are used to expand the search to find all relations which may convey the information

Propagation Direction Switches Back

Some Numbers

1

32

175

7

7

Pruning flows back and forth – starting at 1, going to 32, to 175, then back to 7Some relations generate large numbers of possibilities, other relations prune them

Consistent Reasoning

There were thirty two subsidiaries found initially, but omission of information found on only a few of those – the excess subsidiaries are pruned, so that all sets of objects throughout the query structure are consistent

Consistent reasoning allows information to flow in any direction, and propagation continues until everything is consistent

In this case, the consistent reasoning model is being dynamically created from the text of the query as soon as it is entered

Propagation Completes And Values Returned

Similarities with SQL

• Some similarities with an SQL query – very many differences

• The query can be at multiple levels- “Does John intend to own…”

• The relation can be a variable• Passing from one relation to another – Omit to Exclude

to false Include to false Mention• The query is being expanded within the structure to

catch every possible way the required information could have been said

• The query can go active• Pruning of alternatives can flow back and forth, unlike

SQL

Why Was What

Queries which match structure directly

What – the relation is matched against other relations to find out what “what” is

Was – A logical (or existential) is sought, by matching relations

Why – “why” becomes “what causes”, and matches the relation it controls

Why Do It This Way

• Conversion of the text into a resolved structure speeds up searching

• The points where the query anchors are the points where search should occur – it doesn’t read the text, it searches around the nodes in the network where it already is, or moves along links to anaphora far away in the text

• It can handle an explosion of possibilities at close range