• Eliza (Simon Laven Chatterbot page) Developed by Wizenbaum simulated Rogerian therapy, there is no real understanding, only a lot of clever tricks.
• patern matching (primary and secondary)
• mechanical rewriting of statements into questions
• key word vocabulary (e.g. mother, hate, love) used to focus attention
• rememebered concept (when deadends are reached)

• Early investigators thought that grammar (or syntax) was all you needed to make an automatic langauge translation system work. This was found to be wrong you also need to consider semantics.
• time flies like an arrow
• fruit flies like a banana

• Context free grammar can be represented and parsed by finite state machines.
• These may be extended to form recusive transition networks (RTN), which might be thought of as FSM with networks labeling the arcs. This allows the nodes to call subroutines (context sensative grammars).
• Agumented transition networks (ATN) add registers to RTN to store partially deveoped parse trees, allow conditional execution of arcs, and attach actions capable of modifying the data structure returned. This means that ATN's are Turing machines.

Roger Shank's approach to representing deep meaning.

• I gave the man a book

ATRANS – Abstract transfer (give)

PTRANS – Physical transfer (go)

MTRANS – Mental transfer (tell)

P= Past tense PP = object or picture producer

0 = Object case relation AA = Action modifier (aider)

R = Recipients Act = Action

PA = Picture modifier (aider)

How can conceptual dependency facilitate reasoning?

1. Fewer inference rules.
2. Many inferences are contained in the representation
3. Initial representation of the sentence will have holes. Plugging holes serves as focusing a subject for future sentences.

1. Long time to decompose knowledge into primitive actions
2. Conceptual dependency is good for representing, this may not be good for all kinds of knowledge.

How does Natural Language Processing (NLP) fit into user interface design work?

1. In natural language interface (NLI) queries are open eneded prompts like "what do you want to do?" Which gives very little support to guide user actions
2. Might be used in museuem type applications to allow natural language queries (NLQ) against relational databases. NLQ is parsed and translated to standard SQL.
3. Search engine searches against text database (e.g. find cases where tenents sure landlords. May allow modified SQL type query to search against simple domain model.
4. Natural language text generation. Reports might be written automatically from lab data or output might make use of computer generated speech to help the visually impaired.
5. Text-based adventure games. Restricted domains, make parsing a little easier than continuous speech recognition. Parser is a little more sophiaticated than Eliza.

The classic test of machine intelligence is to have a person communicate electronically with two entities (one a machine and the other a person). If the inquisitor is unable to determine who is the machine and who is the person "true artificial intelligence" has been exhibited. Each year a restricted version Turing competition is held at the Boston Computer Museum. Some progress is evident, but more work needs to be done.