The French Correction: When Retrieval Is Harder to Specify than Adaptation

Abstract

A common idea in the eld of case-based reasoning is that the retrieval step can be specied by the use of some similarity measure: the retrieved cases maximize the similarity to the target problem and, then, the adaptation step has to take into account the mismatches between the retrieved cases and the target problem in order to solve this latter. The use of this methodological schema for the application described in this paper has proven to be non ecient. Indeed, designing a retrieval procedure without the precise knowledge of the adaptation procedure has not been possible. The domain of this application is the correction of French sentences: a problem is an incorrect sentence and a valid solution is a correction of this problem. Adaptation consists in solving an analogical equation that enables to execute the correction of the retrieved case on the target problem. Thus, retrieval has to ensure that this application is feasible. The rst version of such a retrieval procedure is described and evaluated: it is a knowledge-light procedure that does not use linguistic knowledge about French.

Analogy-Based Post-treatment of CNN Image Segmentations

“…Second, when the problem and solution spaces are the same (P = S), a single case adaptation can be performed by solving an equation x s :y s ::x tgt :? (see [7]). In the following, analogical proportions are used according to this second way.…”

A review of the development and future challenges of case-based reasoning

“…In text-based case retrieval, Chang et al (2016) integrated natural language processing (NLP) technology into CBR case retrieval agent to support product information retrieval, including filtering mechanism and semantic-based case retrieval agent. Lepage et al (2020) proposed adaptation before retrieval in the CBR process of sentence correction, that is, using the adaptation-guided retrieval method to achieve French correction. Based on nonsymbolic types such as images, Wilkerson et al (2021) proposed a weighting strategy that performs better in feature-intensive spaces to achieve case retrieval, which obtains the features learned from data using DL to supplement the existing knowledge engineering features and learns the feature weights of both through neural networks.…”