Réutilisation d'entités nommées pour la réponse au courriel
|Advisor:||Lamontagne, Luc D.|
|Abstract:||An automatic e-mail response system is a solution for improving the operations of certain business services, like customers’ services or investor relations. Those services are dealing with a large volume requests coming through e-mail messages, most of them being repetitive. We have decided to explore a CBR approach (Case-Based Reasoning) for this problem. Such an approach makes use of antecedent messages to respond to new incoming e-mails. Requests coming from customers or investors are often redundant; we could select an adequate answer among the archived messages, and then adapt it to make it coherent with the actual context of the new message request. In this project, we address the re-use problem, but more specifically the identification of named entity and their specialized roles. These entities are portions of text strongly depend on the context of the antecedent message, and hence need some adaptation to be re-used. We divide the reuse process in two tasks which are: a) the identification of modifiable portions of an antecedent message; b) the selection of portions to be adapted to build the answer of the request. For first task, we make use of information extraction techniques. But we will concentrate our efforts uniquely on the extraction of named entities and their specializations. For second task we make use of text classification techniques to decide which portions are subject to adaptation. This decision is based on the context of the request, words which compose it. We used different approaches for the two tasks. We tested manual and automatics top-down and bottom-up extraction techniques on an e-mail corpus for the identification of iv modifiable portions extraction task. Manual approach gives us excellent results. But, we notice a degradation of performance for automatic extraction techniques. For the selection of portions to be adapted, we compared made use of association rules and various word representation. Association rules use permits to compress data without degrades results a lot. Globally, results are good and indicate that our approach, desrcibes before, could be applied to our problem.|
|Document Type:||Mémoire de maîtrise|
|Open Access Date:||12 April 2018|
|Collection:||Thèses et mémoires|
All documents in CorpusUL are protected by Copyright Act of Canada.