A Multi-Agent System for Building Dynamic Ontologies K evin Ottens - - PowerPoint PPT Presentation

Introduction Dynamo Distributed Clustering Algorithm Multi-Criteria Hierarchy Discussion & Perspectives

A Multi-Agent System for Building Dynamic Ontologies

K´ evin Ottens, Marie-Pierre Gleizes & Pierre Glize

Institut de Recherche en Informatique de Toulouse (IRIT) SMAC team

AAMAS 2007 – May 14–18 2007, Honolulu, Hawai’i, USA.

K´ evin Ottens, Marie-Pierre Gleizes & Pierre Glize — A Multi-Agent System for Building Dynamic Ontologies 1/22

Introduction Dynamo Distributed Clustering Algorithm Multi-Criteria Hierarchy Discussion & Perspectives

Plan

1 Introduction 2 Introducing the Dynamo System 3 Distributed Clustering Algorithm 4 Multi-Criteria Hierarchy 5 Discussion & Perspectives

K´ evin Ottens, Marie-Pierre Gleizes & Pierre Glize — A Multi-Agent System for Building Dynamic Ontologies 2/22

Introduction Dynamo Distributed Clustering Algorithm Multi-Criteria Hierarchy Discussion & Perspectives

Plan

1 Introduction 2 Introducing the Dynamo System 3 Distributed Clustering Algorithm 4 Multi-Criteria Hierarchy 5 Discussion & Perspectives

K´ evin Ottens, Marie-Pierre Gleizes & Pierre Glize — A Multi-Agent System for Building Dynamic Ontologies 3/22

Introduction Dynamo Distributed Clustering Algorithm Multi-Criteria Hierarchy Discussion & Perspectives

Introduction

Current situation Text analysis makes ontology building easier NLP analysis examination is a difficult and slow process Emerging technics based on machine learning Our proposal Keep the user in the production loop Allow the ”Living Design” of ontologies Reorganization following user modifications

K´ evin Ottens, Marie-Pierre Gleizes & Pierre Glize — A Multi-Agent System for Building Dynamic Ontologies 4/22

Introduction Dynamo Distributed Clustering Algorithm Multi-Criteria Hierarchy Discussion & Perspectives

Plan

1 Introduction 2 Introducing the Dynamo System 3 Distributed Clustering Algorithm 4 Multi-Criteria Hierarchy 5 Discussion & Perspectives

K´ evin Ottens, Marie-Pierre Gleizes & Pierre Glize — A Multi-Agent System for Building Dynamic Ontologies 5/22

Introduction Dynamo Distributed Clustering Algorithm Multi-Criteria Hierarchy Discussion & Perspectives

Overview

Architecture

?? Ontologist Interface Multi−Agent System Concept Agent Term Term network Terms Extraction Tool

K´ evin Ottens, Marie-Pierre Gleizes & Pierre Glize — A Multi-Agent System for Building Dynamic Ontologies 6/22

Introduction Dynamo Distributed Clustering Algorithm Multi-Criteria Hierarchy Discussion & Perspectives

Overview

Term Network Produced by Syntex ”Head-Expansion” graph

knowledge engineering from text knowledge engineering

Term contexts used to compute similarity Multi-Agent System Each agent represents a concept of the taxonomy Each agent tries to position itself Based on a condition/action rule set

K´ evin Ottens, Marie-Pierre Gleizes & Pierre Glize — A Multi-Agent System for Building Dynamic Ontologies 7/22

Introduction Dynamo Distributed Clustering Algorithm Multi-Criteria Hierarchy Discussion & Perspectives

Plan

1 Introduction 2 Introducing the Dynamo System 3 Distributed Clustering Algorithm 4 Multi-Criteria Hierarchy 5 Discussion & Perspectives

K´ evin Ottens, Marie-Pierre Gleizes & Pierre Glize — A Multi-Agent System for Building Dynamic Ontologies 8/22

Introduction Dynamo Distributed Clustering Algorithm Multi-Criteria Hierarchy Discussion & Perspectives

Distributed Clustering Algorithm

Local view

Ak−1 Ak An A2 A1 P

...... ......

A1

Steps

1 Evaluating similarity and ”votes” 2 Partitioning and intermediate layer creation 3 Parent change

K´ evin Ottens, Marie-Pierre Gleizes & Pierre Glize — A Multi-Agent System for Building Dynamic Ontologies 9/22

Introduction Dynamo Distributed Clustering Algorithm Multi-Criteria Hierarchy Discussion & Perspectives

Distributed Clustering Algorithm

Local view

Ak−1 Ak An A2 A1 P P’

...... ......

P’ P’

Steps

1 Evaluating similarity and ”votes” 2 Partitioning and intermediate layer creation 3 Parent change

K´ evin Ottens, Marie-Pierre Gleizes & Pierre Glize — A Multi-Agent System for Building Dynamic Ontologies 9/22

Introduction Dynamo Distributed Clustering Algorithm Multi-Criteria Hierarchy Discussion & Perspectives

Distributed Clustering Algorithm

Local view

Ak−1 Ak An A2 A1 P P’

...... ......

Steps

1 Evaluating similarity and ”votes” 2 Partitioning and intermediate layer creation 3 Parent change

K´ evin Ottens, Marie-Pierre Gleizes & Pierre Glize — A Multi-Agent System for Building Dynamic Ontologies 9/22

Introduction Dynamo Distributed Clustering Algorithm Multi-Criteria Hierarchy Discussion & Perspectives

Distributed Clustering Algorithm

Global View

Ak−1 Ak An A2 A1 P ...... ...... Ak−1 Ak An A2 A1 P P’ ...... ...... P’ P’ Ak−1 Ak An A2 A1 P P’ ...... ...... Ak−1 Ak An A2 A1 P P’ ...... ...... Ak−1 Ak An A2 A1 P ...... ...... Ak−1 Ak An A2 A1 P P’ ...... ...... P’ P’

K´ evin Ottens, Marie-Pierre Gleizes & Pierre Glize — A Multi-Agent System for Building Dynamic Ontologies 10/22

Introduction Dynamo Distributed Clustering Algorithm Multi-Criteria Hierarchy Discussion & Perspectives

Experimental Complexity Results

20000 40000 60000 80000 100000 120000 140000 160000 180000 10 20 30 40 50 60 70 80 90 100 Amount of comparisons Amount of input terms

• 1. Distributed algorithm (on average, with min and max)
• 2. Logarithmic polynomial
• 3. Centralized algorithm

Average complexity: O(n2log(n)) Maximum variance: around 5%

K´ evin Ottens, Marie-Pierre Gleizes & Pierre Glize — A Multi-Agent System for Building Dynamic Ontologies 11/22

Introduction Dynamo Distributed Clustering Algorithm Multi-Criteria Hierarchy Discussion & Perspectives

Qualitative Point of View

Automated run Permanent view on the built hierarchy Allow to obtain a ”first draft” User modification No algorithm adjustment required Dynamicity, revision of the structure

K´ evin Ottens, Marie-Pierre Gleizes & Pierre Glize — A Multi-Agent System for Building Dynamic Ontologies 12/22

Introduction Dynamo Distributed Clustering Algorithm Multi-Criteria Hierarchy Discussion & Perspectives

Plan

1 Introduction 2 Introducing the Dynamo System 3 Distributed Clustering Algorithm 4 Multi-Criteria Hierarchy 5 Discussion & Perspectives

K´ evin Ottens, Marie-Pierre Gleizes & Pierre Glize — A Multi-Agent System for Building Dynamic Ontologies 13/22

Introduction Dynamo Distributed Clustering Algorithm Multi-Criteria Hierarchy Discussion & Perspectives

Head Coverage Rules

Intended Behavior

Observations Similarity can’t be always computed for term pairs Humans have specific heuristics for low-level structuring Goal Take care of those terms Implement a similar heuristic Parent Adequacy Function The best parent for C is the P agent that maximizes a(P, C). When an agent C is unsatisfied by its parent P, it evaluates a(Bi, C) with all its brothers (noted Bi) the one maximizing a(Bi, C) is then chosen as the new parent.

K´ evin Ottens, Marie-Pierre Gleizes & Pierre Glize — A Multi-Agent System for Building Dynamic Ontologies 14/22

Introduction Dynamo Distributed Clustering Algorithm Multi-Criteria Hierarchy Discussion & Perspectives

Managing Several Criteria

Guidelines

How? Keeping it simple

Local criteria Nominal values for those criteria

Use cooperation heuristic Cooperation Minimizing non-cooperation Priority system

Determine the current problems Find the most urgent one Try to fix it

K´ evin Ottens, Marie-Pierre Gleizes & Pierre Glize — A Multi-Agent System for Building Dynamic Ontologies 15/22

Introduction Dynamo Distributed Clustering Algorithm Multi-Criteria Hierarchy Discussion & Perspectives

Managing Several Criteria

Actual Implementation

Minimize non cooperation µH(A): ”head coverage” non cooperation degree of A µB(A): ”brotherhood” non cooperation degree of A µM(A): ”message” non cooperation degree of A µ(A) = max(µH(A), µB(A), µM(A)) Take care of the worst problem first µ(A) = µH(A) → Try to find a better parent µ(A) = µB(A) → Improve structuring through clustering µ(A) = µM(A) → Process other agent message

K´ evin Ottens, Marie-Pierre Gleizes & Pierre Glize — A Multi-Agent System for Building Dynamic Ontologies 16/22

Introduction Dynamo Distributed Clustering Algorithm Multi-Criteria Hierarchy Discussion & Perspectives

Experimental Complexity Revisited

5000 10000 15000 20000 25000 10 20 30 40 50 60 70 80 90 100 Amount of messages Amount of input terms

• 1. Dynamo, all rules (on average, with min and max)
• 2. Distributed clustering only (on average)
• 2. Cubic polynomial

Average complexity: O(n3) Maximum variance: around 0.6%

K´ evin Ottens, Marie-Pierre Gleizes & Pierre Glize — A Multi-Agent System for Building Dynamic Ontologies 17/22

Introduction Dynamo Distributed Clustering Algorithm Multi-Criteria Hierarchy Discussion & Perspectives

Plan

1 Introduction 2 Introducing the Dynamo System 3 Distributed Clustering Algorithm 4 Multi-Criteria Hierarchy 5 Discussion & Perspectives

K´ evin Ottens, Marie-Pierre Gleizes & Pierre Glize — A Multi-Agent System for Building Dynamic Ontologies 18/22

Introduction Dynamo Distributed Clustering Algorithm Multi-Criteria Hierarchy Discussion & Perspectives

Discussion

Advantages of our approach Easier system/ontologist coupling Possible distribution on a network Current limitations Results tend to depend on the add order Tend to produce binary trees only (except on leaves)

K´ evin Ottens, Marie-Pierre Gleizes & Pierre Glize — A Multi-Agent System for Building Dynamic Ontologies 19/22

Introduction Dynamo Distributed Clustering Algorithm Multi-Criteria Hierarchy Discussion & Perspectives

Perspectives

Concerning knowledge engineering Get closer to a taxonomy tree Find non taxonomic relations Concerning multi-agent systems Improve the clustering algorithm

Remove the dependency on add order Optimize

Test this algorithm in other domains

K´ evin Ottens, Marie-Pierre Gleizes & Pierre Glize — A Multi-Agent System for Building Dynamic Ontologies 20/22

Introduction Dynamo Distributed Clustering Algorithm Multi-Criteria Hierarchy Discussion & Perspectives

Works in Progress, Conclusion

In progress... Taxonomy production

Tree pruning Not only binary tree

Evaluate the system on more corpora Conclusion Evolving structure is possible in this field Performances are acceptable More efforts needed...

K´ evin Ottens, Marie-Pierre Gleizes & Pierre Glize — A Multi-Agent System for Building Dynamic Ontologies 21/22

Introduction Dynamo Distributed Clustering Algorithm Multi-Criteria Hierarchy Discussion & Perspectives

Questions ?

K´ evin Ottens

• ttens@irit.fr

K´ evin Ottens, Marie-Pierre Gleizes & Pierre Glize — A Multi-Agent System for Building Dynamic Ontologies 22/22