Logo des Repositoriums
 

German Question Answering in CRM Systems

dc.contributor.authorSchäfter, Sophie
dc.contributor.authorZylowski, Thorsten
dc.date.accessioned2021-12-14T10:56:54Z
dc.date.available2021-12-14T10:56:54Z
dc.date.issued2021
dc.description.abstractWe present a Question Answering (QA) approach within the context of Customer Relationship Management (CRM) systems in German which we applied to several use cases. However, we aim to generalize the suggested approach, so it can be applied to other domains. In this course, the subjects of tabular QA, extractive QA and Frequently Asked Questions (FAQ) are examined. Moreover, we reveal our findings regarding fine-tuning transformer models for QA and propose an automatic labelling mechanism that can be integrated into QA systems in order to simplify the creation of training data. We evaluated on various CRM-related data sources. The evaluation of the fine-tuned extractive QA pipeline resulted in an F1-score of 76.93 %. A qualitative analysis by domain experts showed that the tabular QA pipeline with translation and column name mapping leads to an accuracy of 40 %.en
dc.identifier.doi10.18420/informatik2021-100
dc.identifier.isbn978-3-88579-708-1
dc.identifier.pissn1617-5468
dc.identifier.urihttps://dl.gi.de/handle/20.500.12116/37604
dc.language.isoen
dc.publisherGesellschaft für Informatik, Bonn
dc.relation.ispartofINFORMATIK 2021
dc.relation.ispartofseriesLecture Notes in Informatics (LNI) - Proceedings, Volume P-314
dc.subjectNatural Language Processing
dc.subjectQuestion Answering
dc.subjectInformation Retrieval
dc.subjectCustomer Relationship Management
dc.titleGerman Question Answering in CRM Systemsen
gi.citation.endPage1220
gi.citation.startPage1213
gi.conference.date27. September - 1. Oktober 2021
gi.conference.locationBerlin
gi.conference.sessiontitleWorkshop: Künstliche Intelligenz für kleine und mittlere Unternehmen (KI-KMU 2021)

Dateien

Originalbündel
1 - 1 von 1
Vorschaubild nicht verfügbar
Name:
N1-6.pdf
Größe:
581.01 KB
Format:
Adobe Portable Document Format