Évaluation sémantique d'informations symboliques : la cotation

Abstract : Confidence in information should represent how far one can believe it, how much faith to put in it. Trust is a thriving field of study yet, in general, it tends to measure quality of the process responsible for producing the information rather than advise on whether to believe it or not. In the same way that hearing a fact from a trustworthy source is insufficient to fully believe it, automatic evaluation of trust in an information requires a rich model capable of explicitly puting forward why what it qualifies should or should not be believed. This is the problem we have tackled in our work. From a careful study of an existing representation of confidence, we choose to split the problem in two: the encoding of trust, i.e. how it is represented, and the rules governing its appraisal, i.e. how it is evaluated. We derive the quintessential dimensions participating in the building of trust from the prerequisites imposed on the definition of its encoding. We offer a categorisation of these dimensions which gathers the evaluated criteria according to their object and influence and thus ensures their independence and non-redundancy. We also take great care of ensuring the readability of the measures involved in the assessment by proposing their expression along discrete scales made explicit through the use of linguistic labels. After these dimensions have been selected, we can address the problem of their combination to model the trust-building process. We solve this problem by proposing a philosophy of integration fo the dimensions, that is, we shape the architecture of information scoring. We provide this architecture with a representation as a scoring-chain which highlights the order in which dimensions are considered and the influence they have on the increase or decrease of the confidence evaluation. We also show how the flexibility of our model can be used to represent different user gullibility-postures, an essential adaptability for the modeling of subjective matters. Once these definitions are set, we propose a theoretical formalisation of the scoring process and of its expression, the score. Using the expressiveness of multivalued logics, we choose to set our solutions in this formalism. To reintroduce the important distinction between impossibility of measure and a neutral, yet expressed, measure, we extend this formalism by adding a new truth-degree. Within this new framework of an extended symbolic logic, we define combination operators to represent the entire collection of proposals we offered and formalise credulity-modeling. We then consider the implementation of our model in the extraction and scoring of symbolic information. We first examine the transposition of information scoring to the problem of knowledge extraction from text. We describe successively the scoring of information extraction and that of their fusion, examining for both how the scoring dimensions translate. We then develop a prototype for implementing our model. Finally, we apply both model and prototype to a real-world usecase consisting of the extraction and scoring of a social-network from a corpus of published texts.
Document type :
Theses
Complete list of metadatas

https://hal-univ-paris8.archives-ouvertes.fr/tel-01559975
Contributor : Adrien Revault d'Allonnes <>
Submitted on : Tuesday, July 11, 2017 - 10:18:19 AM
Last modification on : Tuesday, March 26, 2019 - 2:24:04 PM
Long-term archiving on : Thursday, January 25, 2018 - 3:51:48 AM

File

Licence


Public Domain

Identifiers

  • HAL Id : tel-01559975, version 1

Citation

Adrien Revault d'Allonnes. Évaluation sémantique d'informations symboliques : la cotation. Intelligence artificielle [cs.AI]. Laboratoire d'informatique de Paris 6 [LIP6]; Paris 6, 2011. Français. ⟨tel-01559975⟩

Share

Metrics

Record views

189

Files downloads

241