Upgrade to Pro — share decks privately, control downloads, hide ads and more …

DADgraph: A Discourse-aware Dialogue Graph Neural Network for Multiparty Dialogue Machine Reading Comprehension

DADgraph: A Discourse-aware Dialogue Graph Neural Network for Multiparty Dialogue Machine Reading Comprehension

by Jiaqi Li, Ming Liu, Zihao Zheng, Heng Zhang, Bing Qin, Min-Yen Kan, Ting Liu

Paper at: https://arxiv.org/abs/2104.12377
Video at: https://youtu.be/FDM82HdCQUY

Multiparty Dialogue Machine Reading Comprehension (MRC) differs from traditional MRC as models must handle the complex dialogue discourse structure, previously unconsidered in traditional MRC. To fully exploit such discourse structure in multiparty dialogue, we present a discourse-aware dialogue graph neural network, DADgraph, which explicitly constructs the dialogue graph using discourse dependency links and discourse relations. To validate our model, we perform experiments on the Molweni corpus, a large-scale MRC dataset built over multiparty dialogue annotated with discourse structure. Experiments on Molweni show that our discourse-aware model achieves statistically significant improvements compared against strong neural network MRC baselines.

14da6ebc2e909305afdb348e7970de81?s=128

wing.nus

July 21, 2021
Tweet

Transcript

  1. 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
  2. MRC SQuAd, RACE, NarrativeQA, CoQA, QuAC etc.

  3. Multiparty Dialogue MRC A dialog from Molweni dataset. jimcooncat: installing

    acroread gives me a 404 on maverick -- what to do ? 𝑈1 jrib: where are you installing acroread from ? 𝑈2 elfranne: people in the same local network ? 𝑈3 llutz: not network , on local computer 𝑈4 elfranne: so its only available for `` localhost '' and not others on the same local network 𝑈5 jimcooncat: thank you , i had forgot to update 𝑈6 llutz: yes , `` other users on localhost ‘‘ 𝑈7 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.
  4. Background • Discourse structure of multiparty dialogue. jimcooncat: installing acroread

    gives me a 404 on maverick -- what to do ? 𝑈1 jrib: where are you installing acroread from ? 𝑈2 elfranne: people in the same local network ? 𝑈3 llutz: not network , on local computer 𝑈4 elfranne: so its only available for `` localhost '' and not others on the same local network 𝑈5 jimcooncat: thank you , i had forgot to update 𝑈6 llutz: yes , `` other users on localhost ‘‘ 𝑈7 𝑈1 𝑈2 𝑈3 𝑈4 𝑈5 𝑈6 𝑈7 Q-Elab Expl. QAP QAP Ack. Ack. Q-Elab
  5. Background • Discourse structure has been successfully applied to QA

    and MRC. Sachan et al. Learning Answer-Entailing Structures for Machine Comprehension. EMNLP 2015.
  6. Background • Graph structure has been proven to effectively represent

    dialogs. Hu et al. GSN: A Graph-Structured Network for Multi-Party Dialogues. IJCAI 2019.
  7. Hypothesis Discourse structure informs multiparty dialogue MRC performance in modeling

    long-term dependencies.
  8. Method: DADgraph Sequential Context Encoding Discourse Graph Modeling MRC 𝑢1

    GRU 𝑔1 𝑢2 GRU 𝑔2 𝑢3 GRU 𝑔3 𝑢4 GRU 𝑔4 𝑢5 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. ℎ1 ℎ2 ℎ3 ℎ4 ℎ5 ℎ1 ℎ2 ℎ3 ℎ4 ℎ5 𝑤1 𝑤𝑖 𝑤𝑚 𝑤𝑗 … … … 𝑞 Interaction 𝑡𝑖 predict 𝑆𝑁𝐴 𝑆𝑖,𝑗 𝑡1 𝑡𝑚 … … 𝑡𝑗 … 𝑤𝑖: ith word in dialog; q: question
  9. Method • Sequential context modelling • The pretrained BERT model

    is used to represent the utterance to get 𝑢𝑖. • The sequence structure of dialogue is modeled by GRU. • Finally, the utterance representation 𝑔𝑖 of fusion context is obtained; Sequential Context Encoding 𝑢1 GRU 𝑔1 𝑢2 GRU 𝑔2 𝑢3 GRU 𝑔3 𝑢4 GRU 𝑔4 𝑢5 GRU 𝑔5
  10. Method • Discourse graph modelling • The utterance representation 𝑔𝑖

    obtained in the previous step is taken as the input; • GCN is used to model discourse structure of dialogue. • The updated representation ℎ𝑖 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. ℎ1 ℎ2 ℎ3 ℎ4 ℎ5
  11. Method • MRC module • First, let the ℎ𝑖 attend

    with each word in the input dialogue. • Weighted sum of utterances representations and obtains 𝑡𝑖. • Predict the probability of the span (i, j) via 𝑡𝑖 and 𝑡𝑗. 𝑡𝑖 𝑡1 𝑡𝑚 … … 𝑡𝑗 … 𝑞 ℎ1 ℎ2 ℎ3 ℎ4 ℎ5 + 𝑤1 𝑤𝑖 𝑤𝑚 𝑤𝑗 … … … ∙ WS 𝑓1 𝑓𝑖 𝑓𝑚 𝑓𝑗 … … … ∙ 𝑐1 𝑐𝑖 𝑐𝑚 𝑐𝑗 … … … WS:weighted sum
  12. Experiments • 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.
  13. Experiments • 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
  14. Experiments • 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
  15. Experiments: case study

  16. Conclusion • Propose DADgraph model for multiparty dialogue MRC task.

    • Prove the discourse structure can help understand the dialogue. Thank you!