2021 · Peer-reviewed journal article

Performance Trade-Off for Bert Based Multi-Domain Multilingual Chatbot Architectures

We tested whether separate chatbots could share one multilingual BERT model while keeping their own sets of possible answers.

Research question

Do thirteen chatbot tasks need thirteen models?

Evidence and scope

A visitor asking how to get a virtual card and a visitor asking about an unrelated subject need different answers. Our classifier shares the language model across chatbots and uses the active chatbot to narrow the possible responses. The evaluation included banking questions covering 77 intents.

Across the study we tested 13 classification tasks and 333 classes in Turkish, English and German. In the five-model comparison, the separate models used 10 GB of memory in total. The shared model used 2 GB, with a small loss of accuracy.

Primary source

Contribution
Davut Emre Taşar · First and corresponding author
Source attribution
Davut Emre Taşar, Şükrü Ozan, Seçilay Kutal, Oğuzhan Ölmez, Semih Gülüm, Fatih Akca, Ceren Belhan
Publication status
Peer-reviewed journal article
Publication or platform
Journal of Artificial Intelligence and Data Science
Identifier
IZ:JA77UY82ER

Sources

Research areas · Language

Related work