Belief Change for Better Multi-Source Information Analysis – BE4musSIA

WewilluseanddeveloptoolscomingfromKnowledgeRepresentation and Reasoning (KR) in order to perform an analysis of pieces of information coming from different sources. We will also use tools coming from Game Theory and Social Choice Theory for the strategic and (group) decision aspects. To perform this multi-source information analysis, we will in particular use methods coming from belief revision and belief merging, that formalize rational belief change, as well as inconsistency measures, which allows us to measure the extent to which certain pieces of information are in conflict with each other.

WP1: Belief revision

We have proposed several generalizations of these belief revision operators, which makes it possible to broaden the spectrum of operators available to the community, and to define and understand the operators we need for the realization of this project.

A first work was to explore the link between iterated change operators and fusion operators, which allowed us to define the first commutative change operators, as well as to emphasize the fact that the epistemic states usually used in iterated revision must have a more complex structure than a simple pre-order.

A second work was to study how to define revision operators on paraconsistent logics. This is important work on several aspects. First it provides a generalization of classical operators. But it also allows for a general reflection on what revision and expansion are when some models of the agent's beliefs are more important than others.

WP2: Belief Merging

For belief merging we were interested in the ability of belief merging operators to identify the truth from the collected opinions. More precisely, we showed which operators were maximum likelihood estimators for certain standard noise functions.

We also studied the links between belief merging operators and voting methods, in particular Borda's voting method. We have proposed several merging operators inspired by this voting method.


WP5: Application to query answering on multiple ontologies

We have proposed merging operators for ontologies.
We also have several publications on the use of embeddings to help provide context, enabling better understanding, and better learning for several applications.

WP1: Belief revision

We are currently working on the proof that the canonical representation of the epistemic states used for iterated change is the ordinal conditional functions.

We are also working on a generalization of expansion operators, in all cases where some models of the agent's beliefs are more important than others.

WP2: Belief Merging

We are currently working on the problem of fusion when the sources have different reliabilities.

WP3: Inconsistency measures and conflict measures


We are currently working on the classification of a certain number of measures of semantic inconsistencies, thanks to a general framework allowing to find the main measures of semantic inconsistencies as particular cases.

WP4: Analysis of multi-source information

This theme is that of the thesis of Quentin Elsaesser, recruited as part of this chair on this subject. The first results are promising, and it seems that the method we are working on is more efficient than existing methods for the evaluation of information sources on the web.

Nicolas Schwind, Sébastien Konieczny: Non-Prioritized Iterated Revision: Improvement via Incremental Belief Merging. KR 2020: 738-747
I. Bloch, S. Blusseau, R. Pino Pérez, E. Puybareau and G. Tochon. On Some Associations Between Mathematical Morphology and Artificial Intelligence. Discrete Geometry and Mathematical Morphology (LNCS 12708). pp. 457-469, 2021.
Kun Yan, Chenbin Zhang, Jun Hou, Ping Wang, Zied Bouraoui, Shoaib Jameel, Steven Schockaert: Inferring Prototypes for Multi-Label Few-Shot Image Classification with Word Vector Guided Attention, AAAI 2022
Rana Alshaikh, Zied Bouraoui, Steven Schockaert: Hierarchical Linear Disentanglement of Data-Driven Conceptual Spaces. IJCAI 2020: 3573-3579
Na Li, Zied Bouraoui, José Camacho-Collados, Luis Espinosa Anke, Qing Gu, Steven Schockaert: Modelling General Properties of Nouns by Selectively Averaging Contextualised Embeddings. IJCAI 2021: 3850-3856
Kun Yan, Zied Bouraoui, Ping Wang, Shoaib Jameel, Steven Schockaert: Aligning Visual Prototypes with BERT Embeddings for Few-Shot Learning. ICMR 2021: 367-375
Nicolas Schwind, Sébastien Konieczny; Ramón Pino Pérez: On paraconsistent belief revision in LP. AAAI 2022.
Zied Bouraoui, Sébastien Konieczny, Truong-Thanh Ma, Ivan Varzinczak: Model-based Merging of Open-Domain Ontologies. ICTAI 2020: 29-34
Patricia Everaere, Sébastien Konieczny, Pierre Marquis: Belief Merging Operators as Maximum Likelihood Estimators. IJCAI 2020: 1763-1769
Patricia Everaere, Chouaib Fellah, Sébastien Konieczny, Ramón Pino Pérez: Borda, Cancellation and Belief Merging. KR 2021: 291-300
A. Mata Diaz, Ramón Pino Pérez. Merging Epistemic States and Manipulation. ECSQARU 2021: 457-470

Suppose that you receive a continous flow of (possibly conflicting) pieces of information from several sources of (initially) unknown reliability. From these pieces of information you need at all times to i) form your opinion ii) evaluate the reliability of the sources.

This is the case for instance:

- if you use several media (newspapers, journals, etc) to form your own opinion on a topic
- if several sensors (of unknown or variable reliability) provide information to a robot
- if some people provide secret pieces of information to a journalist or an intelligence agency
- if one wants to query several websources (of unknown or variable reliability) on the web

The aim of this project is to study the two aforementioned tasks, and to perform them conjointly in order to obtain the best possible evaluation, that is crucial for all the scenarios listed above.

We will use and develop tools coming from Knowledge Representation and Reasoning (KR) in order to perform an analysis of pieces of information coming from different sources. We will also use tools coming from Game Theory and Social Choice Theory for the strategic and (group) decision aspects.

To perform this multi-source information analysis, we will in particular use methods coming from belief revision and belief merging, that formalize rational belief change, as well as inconsistency measures, which allows us to measure the extent to which certain pieces of information are in conflict with each other.

To the best of our knowledge, there is little work on evaluating the reliability of different sources based on confronting pieces of information coming from them. Basically one can for instance suspect a source which frequently provides pieces of information in disagreement with most of the other sources to be wrong (based on some kind of Condorcet’s jury theorem arguments). But we aim at defining different classes of such conflict-based reliability measures.

But we do not want to isolate this task (reliability assessment) from the one of forming a coherent view of the world from the conflicting pieces of information (the merging task). So the originality of this project is in addressing these two tasks simultaneously. Indeed we think that these two tasks have to be performed together in order to obtain optimal results. In particular, we want to study some kind of joint convergence process, where we use the result of the merging task in order to assess the reliability of the sources. This reliability will then be used to improve the result of the merging, that will be used to improve the reliability assessment, and so on and so forth.

The multi-source information analysis frameworks will depend of the amount of information our agent can keep from the information sources. We will study the cases where she can not keep any information; where she can store one formula for each source, representing a summary of the information provided so far by this source; and where she can store the whole message history.

So, as an academic level, the outputs of this projects will be improvements of the theories of belief change, belief merging, and inconsistency measures, as well as the definition and development of tools for multi-source information analysis.

Towards valorisation, we will apply these results to application cases. In particular we will focus on web applications, in order to reason with multiple ontologies, and to answer to user queries using several (possibly conflicting) web ressources.