Herself’s Artificial Intelligence

There are many problems to be worked through involving language and AI. Many breakthroughs in AI will follow a system that can actually understand language, and it is pretty high on the cool things list. There are two major parts to machine language: speech recognition and generation; and natural language understanding and generation.

Speech recognition uses neural nets, hidden Markov models, Bayesian net works and other tools to tease out what a person is saying and figure out how to pronounce words when responding or reading text to a user. Language understanding uses regular expressions, as in ELIZA, to create responses to the user. State machines and First Order Predicate Calculus have had some successes here. Some work has been done with nets and word vectors also.

Language understanding programs have four main inter-related areas to juggle: syntax; semantics; inference; generation.
Syntax is the structural relationship between words.
Semantics is the meaning of the words.
Inference concerns what is meant or desired from the request or conversation.
Generation is the ability of the software to create a suitable response.

Ambiguity is a large part of the problem, the same word can mean different things in different contexts, the same sentence can mean different things as well. For example: I made her walk. Could be I forced her to stride or I poured cement and made her a walkway.

A language is a set of sentences that may be used to convey information. English contains about 800,000 words, including technical terms, although no one individual probably knows more than 60,000 words. Most written English uses less than 10,000 words and college educated people use about 5,000 words in conversation. Grammar is the syntax or sequences that make up the sentences. Syntax as well as the words used convey information. An interesting point is that where ever people have formed groups a language has developed.

See also:
Hidden Markov Models
Bayesian logic
inkblot earth: Company claims to have solved natural language parsing
AI Qual Summary: Natural Language

Perl Eliza program, source code via Perl for Artificial intelligence