Dr Rob Stacey True Knowledge Ltd.. Open Domain question answering Semantic query language Structured and Unstructured knowledge acquisition >300.

Reconciling facts: how to check the consistency of facts

created from web crawling

Dr Rob StaceyTrue Knowledge Ltd.

True Knowledge Open Domain question answering Semantic query language Structured and Unstructured

knowledge acquisition >300 million facts 20k+ classes Billions of inferred facts

Answering Questions Who was prime minister of the UK

when Bernie Ecclestone was a teenager?

Answering Questions

Local time questions What is the time in Covent Garden

now?

Local time questions

Is Madonna married?

Answering a question Query processing

Triple representation[london] [is an instance of] [city]

Temporal knowledge represented by “facts about facts” [fact: [“123”]] [applies for timeperiod] [<1970 onwards>]

Richness within entity representation• “parametered” objects

[integer: [“8128”]] [group: [london]; [san francisco]]

* Actually 4 with negative relation

Mining the facts

Achieving 96% accuracy with the freetext of Wikipedia

Reconciliation- What system assessment can do

Accept incoming knowledge Contradict knowledge Make knowledge superfluous

Uses user assessments and scoring to determine which facts are believed

How assessment works Run a negative version of the query

• [married] ~[applies to] [madonna] If the query is unknown the fact is

new to the knowledge base If the result is no then fact is either

superfluous or an endorsement If the result is yes the there is a

contradiction

Fact already proven The assertion may simply be an

existing fact – if so more weight is added to the truth of that fact

If the fact is different then it is superfluous to the system, though still valid is it removes the need for inference.

Contradiction One of two facts must be wrong

Assessment scoring decides which fact to believe

The loser is contradicted and not believed or used in query processing

Thanks &QuestionsDr Rob Stacey - True Knowledge Ltd.