Seoul National University

Samsung Sds Co., Ltd.

Jaewoo Ahn

Kyoungwon Park

Gunhee Kim

Seongho Joe

Youngjune Gwon

Hoyoung Kang

A method for training a language model for multimodal dialog includes generating a first token sequence by tokenizing one or more text messages included in multimodal dialog data, generating a second token sequence by inserting, based on an appearance position of an image included in the multimodal dialog data, one or more image presence tokens and one or more image non-presence tokens into the first token sequence, and replacing one or more tokens among a plurality of tokens generated by the tokenization with a mask token, generating one or more feature vectors for the image, and training an artificial neural network-based language model to predict, based on the second token sequence and the one or more feature vectors, an appearance position of the image in the multimodal dialog data.

Growing community of inventors

Seoul, South Korea

Jaewoo Ahn