MPG's Chapter on Random Indexing

part 1: motivation

if we want to do web-scale, subsetting becomes key. goal: Compute similarity between any pair of resources, or a query and a resource Then we can use a threshold on similarity to do subsetting The problem we attack is subsetting of the largest reason-able semantic re pository: LDSR [cite Naso's iswc] intro to statistical semanics. It does not use ontology emphasis on goal: complex models
- do not scale don't do well, as Stone, Dennis and Kwantes { , 2009 #1362} show
simpler models that use too much info (i. e. the entire web) do not claim to be cognitive models

tasks that are appropriate. word-word. But not so useful here. Only useful when we have no full text definition. doc-doc

we replicate the state of the art we discard some of the state of the art because it's not plausible as a model of cognition

why our gold standard is not gold. See also discussion keyword extraction from sparql queries is imperfect
- method
  - ASK api calls sets up weights. Weights are generated by RI and ESA NN per query
  results
  - timing precision and recall

First stub at evaleating subsetting method is applicable to spreading activation, future deliverable

Why don't we use the state of the art on word-word? We do use state of the art for doc-doc What does it mean when models do better than human agreement?
- human agreement is poor, even in fields where experts are highly paid The average of many noisy judgments is a more reliable estimator. Wisdom of crowds. It's easier to model the average than the individual.

Introduction on the standardization effort for user interests related vocabularies.

Detailed design of user interests based selection plug-in.

A comparative study of 3 different query strategies, which shows the advantage of user interests based selection.

The draft of this chapter can be found at D2-3-2-March 7th-WICI-chapter.doc