Generation Sedigheh Mahdavi, Aijun An, Heidar Davoudi Davoudi, - PowerPoint PPT Presentation

Question-Worthy Sentence Selection for Question Generation Sedigheh Mahdavi, Aijun An, Heidar Davoudi Davoudi, Marjan Delpisheh, Emad Gohari Gohari York University

Automatic Question Generation ❑ Teaching machines to ask questions What if there is an engine If there is an engine overheat Model warning , but no steam is seen or overheat warning but no steam is seen or heard ? heard , the problem may not be too serious . 2

Application  Chat bot interfaces and smart conversational systems  Educational environments  Reading comprehension assessment 3

Application  iNAGO Inc. Company  Chat bot interfaces and AI conversational assistance  Car Manual 4

Neural Question Generation ❑ Seq2Seq models What if there is an engine If there is an engine overheat Seq2SeqModel overheat warning but no warning , but no steam is seen or steam is seen or heard ? heard , the problem may not be too serious . 5

Paragraph-level Neural Question Generation 6

Challenge of Car Manual ❑ A small set of training data ❑ Lots of irrelevant sentences 7

Why not summarization methods? This section provides a brief overview about some of the important features that may or may not be on your specific vehicle. for more detailed information, refer to each of the features which can be found later in this owner’s manual. The remote keyless entry (rke) transmitter is used to remotely lock and unlock the doors from up to 60m (197ft) away from the vehicle. press to unlock the driver door. press unlock-symbol again within three seconds to unlock all remaining doors . press to lock all doors. lock and unlock feedback can be personalized. see vehicle personalization. 8

Car manual data example 9

Feature-based Question-worthy Sentence Extraction ❑ Ranking features obtained from summarization methods Question-worthy context (C) ❑ POS-tag features: (verbs in a sentence, (nouns, adjectives, adverbs,… ) A classifier ❑ Tf/idf ❑ The length of a sentence context-based and sentence- based features ❑ The stop words ratio The corresponding passage of the input sentence 10

Experiments 11

Experiments 12

The proposed Model Question generation model Generated question ( Q) Feature-based question-worthy sentence selection LSTM Question-worthy Question-worthy Decoder context (C) context (C) LSTM Attention A classifier A classifier LSTM context-based and sentence- context-based and sentence- LSTM based features based features The input sentence (S) LSTM Encoder The corresponding passage of The corresponding passage of the input sentence the input sentence LSTM 13

Experiments 14

Experiments 15

Thank you for your attention