A Neural Network Based Natural Language Translator

1
A Neural Network Based Natural Language Translator

• As Compared to a Dictionary Based Translator

A Project By Will Servos Advised By Prof. Ralph
Morelli
2
The Problem

• Applying Artificial Intelligence to Natural
  Language
• Comparing a Neural Network approach to a standard
  Dictionary-based system

play
spielen
English
Neural Network
German
learn
lernen
3
The Current Situation

• Researching the Problem
• A healthy background of similar problems
• Software to be utilized
• Designing the Networks
• Test networks will be run on a variety of depths
• Input and Output will include modifiers to
  simplify the character values
• The Networks will be tested using a variety of
  lay types
• Setting up the Data
• Data Sets have been collected
• Verbs are grouped into pairs (English-German)
• Only Regular Verbs are Qualified for the list
  initially

4
Implementing JOONE

• Simple GUI
• Interfaces with Spreadsheets
• Gives a semblance of a tangible network
• Tale first input
• Character Conversion

5
SYSTRAN 5.0 in Depth

• Commonly Used
• Trusted by Individuals and Corporations
• Reliable, Consistent, Fast Results
• Library Servers
• Extensive Rule Sets
• Internet portals use generic translation service
  based on the largest possible dictionaries.
• Aimed at gist level translation

6
What Qualifies a Verb. . .

• Regular verbs in English
• Tense Change
• Regular verbs in German
• Tense Change
• Subject Change
• Length in both languages cannot exceed 12
  characters
• Word list 750 verb pairs (English-German)

• dance - tanzen
• kiss - kuessen
• love - lieben
• learn - lernen
• listen - hoeren
• need - brauchen
• play - spielen
• rain - regnen
• remember - erinnern
• search - suchen
• thank - danken
• work - arbeiten

7
Project Status

• Accomplishments
• Acquire JOONE Software
• Acquire SYSTRAN 5.0
• Research Similar Efforts
• Compile Regular Verb List
• Begin Implementing Test Networks

• Expandability
• Adapt for Irregular Verbs
• Train Adapted NN
• Test Adapted NN against SYSTRAN

• To Be Accomplished
• Train NN
• Test NN against SYSTRAN