Tweet
Tweet

Slide 1

Slide 1 text

Jiaqi Li, Ming Liu, Zihao Zheng, Heng Zhang, Bing Qin, Min-Yen Kan, Ting Liu July 21 IJCNN 2021 DADgraph: A Discourse-aware Dialogue Graph Neural Network for Multiparty Dialogue Machine Reading Comprehension Harbin Institute of Technology National University of Singapore

Slide 2

Slide 2 text

MRC SQuAd, RACE, NarrativeQA, CoQA, QuAC etc.

Slide 3

Slide 3 text

Multiparty Dialogue MRC A dialog from Molweni dataset. jimcooncat: installing acroread gives me a 404 on maverick -- what to do ? π‘ˆ1 jrib: where are you installing acroread from ? π‘ˆ2 elfranne: people in the same local network ? π‘ˆ3 llutz: not network , on local computer π‘ˆ4 elfranne: so its only available for `` localhost '' and not others on the same local network π‘ˆ5 jimcooncat: thank you , i had forgot to update π‘ˆ6 llutz: yes , `` other users on localhost β€˜β€˜ π‘ˆ7 Q1: Why does jimcoonact meet the error? A1: forgot to update Q2: Where does llutz install acroread? A2: on local computer Q3: How did erUSUL create a new partiton table? A3: NA.

Slide 4

Slide 4 text

Background β€’ Discourse structure of multiparty dialogue. jimcooncat: installing acroread gives me a 404 on maverick -- what to do ? π‘ˆ1 jrib: where are you installing acroread from ? π‘ˆ2 elfranne: people in the same local network ? π‘ˆ3 llutz: not network , on local computer π‘ˆ4 elfranne: so its only available for `` localhost '' and not others on the same local network π‘ˆ5 jimcooncat: thank you , i had forgot to update π‘ˆ6 llutz: yes , `` other users on localhost β€˜β€˜ π‘ˆ7 π‘ˆ1 π‘ˆ2 π‘ˆ3 π‘ˆ4 π‘ˆ5 π‘ˆ6 π‘ˆ7 Q-Elab Expl. QAP QAP Ack. Ack. Q-Elab

Slide 5

Slide 5 text

Background β€’ Discourse structure has been successfully applied to QA and MRC. Sachan et al. Learning Answer-Entailing Structures for Machine Comprehension. EMNLP 2015.

Slide 6

Slide 6 text

Background β€’ Graph structure has been proven to effectively represent dialogs. Hu et al. GSN: A Graph-Structured Network for Multi-Party Dialogues. IJCAI 2019.

Slide 7

Slide 7 text

Hypothesis Discourse structure informs multiparty dialogue MRC performance in modeling long-term dependencies.

Slide 8

Slide 8 text

Method: DADgraph Sequential Context Encoding Discourse Graph Modeling MRC 𝑒1 GRU 𝑔1 𝑒2 GRU 𝑔2 𝑒3 GRU 𝑔3 𝑒4 GRU 𝑔4 𝑒5 GRU 𝑔5 GCN 𝑔1 𝑔2 𝑔3 𝑔4 𝑔5 Links between utterances from same speaker. Links between utterances from different speaker. Elaboration relation. Question-answer pair relation. Acknowledgement relation. …… There are many other discourse relations not in the graph. β„Ž1 β„Ž2 β„Ž3 β„Ž4 β„Ž5 β„Ž1 β„Ž2 β„Ž3 β„Ž4 β„Ž5 𝑀1 𝑀𝑖 π‘€π‘š 𝑀𝑗 … … … π‘ž Interaction 𝑑𝑖 predict 𝑆𝑁𝐴 𝑆𝑖,𝑗 𝑑1 π‘‘π‘š … … 𝑑𝑗 … 𝑀𝑖: ith word in dialog; q: question

Slide 9

Slide 9 text

Method β€’ Sequential context modelling β€’ The pretrained BERT model is used to represent the utterance to get 𝑒𝑖. β€’ The sequence structure of dialogue is modeled by GRU. β€’ Finally, the utterance representation 𝑔𝑖 of fusion context is obtained; Sequential Context Encoding 𝑒1 GRU 𝑔1 𝑒2 GRU 𝑔2 𝑒3 GRU 𝑔3 𝑒4 GRU 𝑔4 𝑒5 GRU 𝑔5

Slide 10

Slide 10 text

Method β€’ Discourse graph modelling β€’ The utterance representation 𝑔𝑖 obtained in the previous step is taken as the input; β€’ GCN is used to model discourse structure of dialogue. β€’ The updated representation β„Žπ‘– of fusion discourse structure is obtained; GCN Discourse Graph Modeling 𝑔1 𝑔2 𝑔3 𝑔4 𝑔5 Links between utterances from same speaker. Links between utterances from different speaker. Elaboration relation. Question-answer pair relation. Acknowledgement relation. …… There are many other discourse relations not in the graph. β„Ž1 β„Ž2 β„Ž3 β„Ž4 β„Ž5

Slide 11

Slide 11 text

Method β€’ MRC module β€’ First, let the β„Žπ‘– attend with each word in the input dialogue. β€’ Weighted sum of utterances representations and obtains 𝑑𝑖. β€’ Predict the probability of the span (i, j) via 𝑑𝑖 and 𝑑𝑗. 𝑑𝑖 𝑑1 π‘‘π‘š … … 𝑑𝑗 … π‘ž β„Ž1 β„Ž2 β„Ž3 β„Ž4 β„Ž5 + 𝑀1 𝑀𝑖 π‘€π‘š 𝑀𝑗 … … … βˆ™ WS 𝑓1 𝑓𝑖 π‘“π‘š 𝑓𝑗 … … … βˆ™ 𝑐1 𝑐𝑖 π‘π‘š 𝑐𝑗 … … … WS：weighted sum

Slide 12

Slide 12 text

Experiments β€’ Dataset: Molweni corpus. Train Dev Test Total Dialogs 8,771 883 100 9,754 Utterances 77,374 7,823 845 86,042 Questions 24,682 2,513 2,871 30,066 Li et al. Molweni: A Challenge Multiparty Dialogues-based Machine Reading Comprehension Dataset with Discourse Structure. COLING 2020.

Slide 13

Slide 13 text

Experiments β€’ Main results EM F1 BiDAF 22.9 39.8 DocQA 42.5 56.0 BERT 45.3 58.0 DialogueRNN 45.4 60.9 DialogueGCN 45.7 61.0 DADgraph (Our) 46.5 61.5 Human performance 64.3 80.2

Slide 14

Slide 14 text

Experiments β€’ Ablation results EM F1 DADgraph 46.5 61.5 - w/o discourse relations 44.9 60.6 - w/o discourse structure 44.7 60.5

Slide 15

Slide 15 text

Experiments: case study

Slide 16

Slide 16 text

Conclusion β€’ Propose DADgraph model for multiparty dialogue MRC task. β€’ Prove the discourse structure can help understand the dialogue. Thank you!