EVA: An Expert System for Classifying Ancient Greek Vases
1. EVA
An Expert System for Vases of the Antiquity
Martina Trognitz
Deutsches Arch¨ ologisches Institut, Berlin
a
22 October 2013
M. Trognitz (DAI)
EVA
1 / 31
2. What is EVA?
EVA is an expert system for computer aided classification and dating of
ceramics. It represents the application of natural language processing
methods for an archaeological problem.
It was the subject of my master thesis at the University of Heidelberg, in
the department of Computational Linguistics.
M. Trognitz (DAI)
EVA
2 / 31
3. Outline
1
Motivation
2
The problem
3
What is an expert system?
How it works
Properties
4
Implementation
Before implementation
System architecture
The knowledge base
Description texts
5
Discussion
M. Trognitz (DAI)
EVA
3 / 31
4. Motivation
Motivation
Combination of computational linguistics and archaeology
Fill the gap between the number of human experts and amount of
unclassified ceramic
EVA could provide a second opinion and be used as a learning tool
M. Trognitz (DAI)
EVA
4 / 31
5. The problem
Outline
1
Motivation
2
The problem
3
What is an expert system?
How it works
Properties
4
Implementation
Before implementation
System architecture
The knowledge base
Description texts
5
Discussion
M. Trognitz (DAI)
EVA
5 / 31
6. The problem
The problem
Ceramic is a common find at excavations.
Form and decoration depend on various factors and change in the
course of time:
cultural environment
source
taste and fashion
technical achievements
Hence ceramic is used to date archeological find deposits. It serves as
a type fossil.
M. Trognitz (DAI)
EVA
6 / 31
7. The problem
Problem
An ancient Greek vase is a difficult object for the non-expert to come to
terms with. Faced with rows of apparently undifferentiated black, red and
buff pots, he or she is at a loss as to where to begin.
Tom Rasmussen & Nigel Spivey
M. Trognitz (DAI)
EVA
7 / 31
8. The problem
Problem
An ancient Greek vase is a difficult object for the non-expert to come to
terms with. Faced with rows of apparently undifferentiated black, red and
buff pots, he or she is at a loss as to where to begin.
Tom Rasmussen & Nigel Spivey
Solution
Store the knowledge of an expert into an expert system to classify and
date ceramic.
M. Trognitz (DAI)
EVA
7 / 31
9. What is an expert system?
Outline
1
Motivation
2
The problem
3
What is an expert system?
How it works
Properties
4
Implementation
Before implementation
System architecture
The knowledge base
Description texts
5
Discussion
M. Trognitz (DAI)
EVA
8 / 31
10. What is an expert system?
What is an expert system?
It is a program capable of solving
problems similarly as human experts
would do.
It uses knowledge and inference
methods to solve problems.
It can solve complex problems
normally requiring enourmous human
expertise.
Edward Feigenbaum
“Father of expert systems”
M. Trognitz (DAI)
EVA
9 / 31
11. What is an expert system?
Special subject of artificial intelligence
Early systems were developed in the sixties (DENDRAL)
They are used comercially since the eighties
Can be used in a wide range of subjects (MYCIN, PROSPECTOR,
XCON/R1)
M. Trognitz (DAI)
EVA
10 / 31
12. What is an expert system?
Special subject of artificial intelligence
Early systems were developed in the sixties (DENDRAL)
They are used comercially since the eighties
Can be used in a wide range of subjects (MYCIN, PROSPECTOR,
XCON/R1)
Remark
An expert system only works well in a well-defined special field, the
knowledge domain.
M. Trognitz (DAI)
EVA
10 / 31
13. What is an expert system?
How it works
How it works
facts & informations
user interface
user
M. Trognitz (DAI)
expertise
expert system
knowledge base
inference engine
EVA
11 / 31
14. What is an expert system?
How it works
Functionality of the inference engine
inference engine
deductions
facts
working memory
agenda
knowledge base
M. Trognitz (DAI)
EVA
12 / 31
15. What is an expert system?
Properties
Properties of expert systems
high perfomance, results compete with those of human experts
proper response time
robust
understandable
flexible
M. Trognitz (DAI)
EVA
13 / 31
16. What is an expert system?
Properties
Properties of expert systems
high perfomance, results compete with those of human experts
proper response time
robust
understandable
flexible
Remark
The knowledge is explicitly disconnected from the processig part of the
program.
M. Trognitz (DAI)
EVA
13 / 31
17. Implementation
Outline
1
Motivation
2
The problem
3
What is an expert system?
How it works
Properties
4
Implementation
Before implementation
System architecture
The knowledge base
Description texts
5
Discussion
M. Trognitz (DAI)
EVA
14 / 31
18. Implementation
Before implementation
Before implementation
Size of knowledge domain?
“Ancient vases” is reduced to:
attic protogeometric and geometric
Knowledge base?
sort of a rule-based system
based on a decision tree
Inference engine and knowledge base depend on each other
The system is implemented with Python
M. Trognitz (DAI)
EVA
15 / 31
20. Implementation
The knowledge base
The knowledge base
The knowledge base is the heart of the system
A restricted, well-defined knowledge domain is mapped to a
knowledge base
The knowledge representation depends on:
area of application
scope
source of knowledge
functionality of the inference engine
A knowledge engineer transfers the knowledge from an expert to the
knowledge base of the expert system.
M. Trognitz (DAI)
EVA
17 / 31
21. Implementation
The knowledge base
Why a decision tree?
A specific vase can be described with a specific set of characteristics.
EG I:
figured:no; shape:amphora; handlePosition:neck; handleForm:band;
body:ovoid; motifs:band of slanting lines
MG II:
figured:no; shape:amphora; handlePosition:neck; handleForm:band;
body:ovoid; motifs:hatched meander, zigzag, dogtooth
IF a specific set of characteristics is given THEN you get a specific
vase.
M. Trognitz (DAI)
EVA
18 / 31
22. Implementation
The knowledge base
Decision tree in EVA
start
figured?
no
shape?
amphora
handlePostion?
neck
...
M. Trognitz (DAI)
shoulder
...
yes
LG!
oinochoe
...
...
...
belly
...
EVA
19 / 31
23. Implementation
The knowledge base
tree.py
The decision tree is built in the module tree.py
Each node in the tree consists of a value, a question related to the
value, a link to the parent node and a list of child nodes.
no
shape?
amphora
handlePostion?
M. Trognitz (DAI)
oinochoe
...
EVA
yes
LG!
...
...
20 / 31
24. Implementation
The knowledge base
Knowledge base details
The knowledge base is stored in a separate text file
root - figured? : yes - LG!
root - figured? : no - shape? : amphora - handlePosition? : neck - ...
root - figured? : no - shape? : amphora - handlePosition? : shoulder -
The full question that is displayed to the user is stored in
dictionaries.py
M. Trognitz (DAI)
EVA
21 / 31
27. Implementation
Description texts
This is a small oinochoe. The handle is
outlined and has a wavy line. The rim is
decorated with three horizontal lines. On
the neck a horizontal row of dots, a
horizontal panel with a lozenge chain and
another row of dots can be seen. The
shoulder is decorated with a dotted snake
and some sparse dots. On the belly are
linked dots. All ornaments are interspersed
by encircling bands. The lower part of the
vase is covered by a thin layer of clay.
M. Trognitz (DAI)
EVA
CVA Oxford 4 (GB, 24) p.
12, plate 30 1-3
23 / 31
28. Implementation
Description texts
This is a small oinochoe. The handle is
outlined and has a wavy line. The rim is
decorated with three horizontal lines. On
the neck a horizontal row of dots, a
horizontal panel with a lozenge chain and
another row of dots can be seen. The
shoulder is decorated with a dotted snake
and some sparse dots. On the belly are
linked dots. All ornaments are interspersed
by encircling bands. The lower part of the
vase is covered by a thin layer of clay.
M. Trognitz (DAI)
EVA
CVA Oxford 4 (GB, 24) p.
12, plate 30 1-3
23 / 31
29. Implementation
Description texts
Some aspects of natural language texts
The form and structure of the texts depend on:
personal style
language skills
knowledge of the subject
Some informations may be missing
M. Trognitz (DAI)
EVA
24 / 31
31. Implementation
Description texts
Information extraction
Question
At which part of the body are the handles attached?
Possible answers
This neck-handled amphora has a thick barred rim.
The handles of this amphora are attached to the neck.
This is a neck-handled amphora with a thick barred rim.
This amphora has a thick barred rim. The handles are on the neck.
M. Trognitz (DAI)
EVA
25 / 31
32. Implementation
Description texts
Information extraction
Question
At which part of the body are the handles attached?
Possible answers
This neck-handled amphora has a thick barred rim.
The handles of this amphora are attached to the neck.
This is a neck-handled amphora with a thick barred rim.
This amphora has a thick barred rim. The handles are on the neck.
Possible patterns to look for
... string -handled ...
... handles verbal phrase with on/to ... string
M. Trognitz (DAI)
EVA
25 / 31
33. Implementation
Description texts
Processing of texts
1: Parsing
Stanford Parser (Klein – Manning 2003)
2: Go through decision tree
Questions are answered automatically with the given text (information
extraction)
It is done by searching for specific patterns
M. Trognitz (DAI)
EVA
26 / 31
34. Implementation
Description texts
Demonstration of EVA
The main module ist called core.py
core.py builds the decision tree with tree.py and a knowledge base
stored in a text file
In dictionaries.py the user friendly questions and answers are stored.
The patterns for information extraction are also stored in
dictionaries.py
M. Trognitz (DAI)
EVA
27 / 31
35. Implementation
Description texts
Demonstration of EVA
The main module ist called core.py
core.py builds the decision tree with tree.py and a knowledge base
stored in a text file
In dictionaries.py the user friendly questions and answers are stored.
The patterns for information extraction are also stored in
dictionaries.py
EVA is in an experimental status.
M. Trognitz (DAI)
EVA
27 / 31
36. Discussion
Outline
1
Motivation
2
The problem
3
What is an expert system?
How it works
Properties
4
Implementation
Before implementation
System architecture
The knowledge base
Description texts
5
Discussion
M. Trognitz (DAI)
EVA
28 / 31
37. Discussion
Drawbacks
time-consuming
Knowledge base and patterns are handcraftet.
abstract
The program does not consider the actual vase (e.g. looks at images). It
only relies on textual descriptions.
M. Trognitz (DAI)
EVA
29 / 31
38. Discussion
Future work
GUI (with example images)
Use a certainty factor to weigh outcome
Expansion of knowledge base
more regions; earlier and later styles
Additional languages
Build knowledge base by means of machine learning
Include image recognition
M. Trognitz (DAI)
EVA
30 / 31