CS-EE 481 Spring 2005 1Founders Day, 2005 University of Portland School of Engineering Project Pocket Gopher Conversational Learning Agent Team Josh Jones.

CS-EE 481 Spring 2005

Founders Day, 2005 1University of Portland School of Engineering

Project Pocket GopherConversational Learning Agent

TeamJosh Jones

Philip Little

Tiffany Takahara

AdvisorDr. Rylander

Industry RepresentativeMs. Glenna Van Duzer

Harland Financial Solutions



Agenda

• Introduction Philip

• Background Philip

• Methods Tiffany

• Results Josh

• Conclusions Tiffany

• Demonstration



Introduction

Thanks to Dr. Rylander, Ms. Glenna Van Duzer and Dr. Ward for all of their help.



Introduction

• Objective: Allow computers to process human language.

• Provide a foundation for future programs that ‘understand’ and respond to written language.

• Provide a sample application that uses this foundation.



Background

• Loebner Prize Contest, Artificial Intelligence• Initial vision: Design a program which can

converse normally with a human.• Actual scope: A program which can present

the grammatical breakdown of a sentence.• Application areas: User interfaces, search

technology, tech support, software development, etc.



Background

• Parsing written language: A hard problem.

• No, I mean a really hard problem.

• Entire research field devoted to this area.



The CLA (Conversational Learning Agent)

• The user types an English sentence into the CLA, to which the CLA gives a response.

• The CLA’s response is determined by its current “personality module.”

• The default personality module recognizes the grammatical structure of a sentence.



The CLA (cont.)

• Words and grammar rules can be added to the CLA’s data file.– (the “learning” in Conversational Learning

Agent)

• Definitions are associated with each word.



Methods

Rapid Application Development Model

• 3 main parts: Underlying Architecture (Josh), GUI (Tiffany), Parser (Philip).

• Modeling Phase done as a group

• Construction done individually

• Integration & Deployment done as a group



Methods

Underlying Architecture (Josh)

• Lets the program access word definitions.

• Hash table for fast word lookup.

• Allows new words to be added to datafile.



Methods

GUI (Tiffany)

• Implemented in C++.

• Designed for Windows XP.

• Chat interface, Word Editor, and Rules Editor.

• Integration with personality module and parsing back-end.



Methods

Parser (Philip)

• Researched existing algorithms.

• Built parser back-end around modified algorithm.

• Designed public interface & integrated



User

Personality Module

Parser Word Graph

Sentence

Response

Sentence

Parse Tree

GUI Chat WindowSentence

Definitions

Response

Architecture



Architecture

• Consists of 4 main objects interacting• GUI

– The interface which the user interacts with the program.

– Three modes of interaction

• Word Graph– Internal database of recognized words and their

definitions.



Architecture (Cont.)

• Personality Module– Determines responses based on grammatical

structure of a sentence and its meaning.– Functionality decided by developer– May update Word Graph

• Parser– Analyzes the sentence and identifies

grammatical structure



Implementation

• Parsing isn’t as easy as it sounds!

• Consider the sentence “Book that flight.”

1. Book. “It must be a noun!”

2. That. “A determinant?”

3. Flight. “Another noun?”

• “Book” is a verb in this context.



Implementation

• Solution: Earley Algorithm

• Only tests grammatical correctness

• Solution: Modify to include parse tree info



Improvements

• Make properties/relationships within words more dynamic– The different properties definitions can have is

currently hard-coded. Use a data file for easier modification if other representations are needed.

• Allow removal of definitions.– Current design does not like forgetting its vocabulary.

• Handle numbers



Conclusions

• The CLA currently provides a method of quickly identifying various parts of English sentences.

• The CLA provides the means to be easily extended with further capabilities– Learning new words, relationships based on user input.

– Providing different kinds of responses.

• An exciting step into the problem of natural language recognition.



Demonstration

• Inputting various sentences, showing how the CLA responds.

• Adding and modifying word information.

• Adding and modifying grammar rules.



Any Questions?

Thank You!

(Not exactly a pocket gopher…)

CS-EE 481 Spring 2005 1Founders Day, 2005 University of Portland School of Engineering Project Pocket Gopher Conversational Learning Agent Team Josh Jones.

Documents

founders day

group slide

word editor

underlying architecture

written language

human language

default personality

current personality