Skip to Main content Skip to Navigation
Theses

De l'analyse d'opinions à la détection des problèmes d'interactions humain-machine : application à la gestion de la relation client

Abstract : This PHD thesis is motivated by the growing popularity of chatbots acting as advisors on corporate websites. This research addresses the detection of the interaction problems between a virtual advisor and its users from the angle of opinion and emotion analysis in the texts. The present study takes place in the concrete application context of a French energy supplier EDF, using EDF chatbot corpus. This corpus gathers spontaneous and rich expressions, collected in "in-the-wild" conditions, difficult to analyze automatically, and still little studied. We propose a typology of interaction problems and annotate a part of the corpus according to this typology. A part of created annotation is used to evaluate the system. The system named DAPI (automatic detection of interaction problems) developed during this thesis is a hybrid system that combines the symbolic approach and the unsupervised learning of semantic representation (word embeddings). The purpose of the DAPI system is to be directly connected to the chatbot and to detect online interaction problems as soon as a user statement is received. The originality of the proposed method is based on : i) taking into account the history of the dialogue ; ii) the modeling of interaction problems as the expressions of user spontaneous opinion or emotion towards the interaction ; iii) the integration of the web-chat and in-the-wild language specificities as linguistic cues for linguistic rules ; iv) use of lexical word embedding (word2vec) learned on the large untagged chatbot corpus to model semantic similarities. The results obtained are very encouraging considering the complexity of the data : F-score = 74.3%.
Complete list of metadata

https://pastel.archives-ouvertes.fr/tel-03383799
Contributor : ABES STAR :  Contact
Submitted on : Monday, October 18, 2021 - 4:58:10 PM
Last modification on : Friday, April 15, 2022 - 1:46:00 PM
Long-term archiving on: : Wednesday, January 19, 2022 - 9:36:16 PM

File

PoltavchenkoMaslowskiManuscrit...
Version validated by the jury (STAR)

Identifiers

  • HAL Id : tel-03383799, version 1

Citation

Irina Poltavchenko. De l'analyse d'opinions à la détection des problèmes d'interactions humain-machine : application à la gestion de la relation client. Traitement du texte et du document. Télécom ParisTech, 2018. Français. ⟨NNT : 2018ENST0030⟩. ⟨tel-03383799⟩

Share

Metrics

Record views

101

Files downloads

69