Anti-Semitic Semantic Search

Nazis cause Jews, Arabs cure them, hookers are best treated with steroids and Adobe Photoshop, and one of the benefits of diabetes is copying documents. Narcissism is treatable by both aggression and Freud, but the latter may cause arthritis and great anguish and destroy your career. Marijuana can "assess psychometric accuracy" and help you "achieve wider social acceptance".

This is just a sampling of the medical advice dished up by HealthBase, a new semantic search engine evidently released into public beta prematurely. As TechCrunch reported last week, HealthBase drew some minor controversy for listing "Jew" as a cause of AIDS. Treatments include "Arab", "gene", alcohol and coarse salt. In an analysis of the benefits and drawbacks, "Pros of Jew" include "endorse extrinsic religiosity" and "facilitate Americanization", while cons include "enemy", "plague" and "kill Christian". On a more paradoxical note, the causes include "Nazi" and "Hitler".

While NetBase, developer of HealthBase, works frantically to improve the engine, bloggers and forum posters such as myself are finding more comical errors.

The concept seems simple enough. Extract the parts of speech and spellings from an English dictionary and the synonyms from a thesaurus. Load up some scholarly journals and Wikipedia articles. Index them by groups of synonymous words. When you need a list of treatments for lung cancer, search "treat", "treatment" and their synonyms near "lung cancer" and its synonyms. Then, do a little grammar checking to guess what word in the sentence is the name of the treatment.

The problem is that so many words have multiple meanings. The link between Jews and AIDS originated, according to TechCrunch, with text like:

Hispano-Visigothic king Egica accuses the Jews of aiding the Muslims, and sentences all Jews to slavery.

HealthBase must think "accuses" always means "accuses of causing" and not bother with the difference between "AIDS" and "aiding". I bet the internal parsing looks something like this: "(Someone) (accuses) (Jew(s)) [of causing] (AIDS) (to the Muslims); (irrelevant clause)."

The "hookers" example seems to have been caused by citations of authors named Hooker.

Another weakness of HealthBase is that it seems to ignore words such as "prevention", "reducing" and "eliminating". That's how it concludes sugar is a treatment for both ADHD and obesity, while SARS reduces depression. It also can't tell whether a given noun in specific enough; thus "herb" is indicated as a treatment for ADHD.

Semantic processing of prose is a promising field, but it's got a long way to go if HealthBase is the state of the art. Gone are the days when artificial intelligence was no quack.

Anti-Semitic Semantic Search