cecg subtask
play

CECG Subtask Speaker: Jiaqing Liu School of Information Renmin - PowerPoint PPT Presentation

RUCIR at NTCIR-14 STC-3 CECG Subtask Speaker: Jiaqing Liu School of Information Renmin University of China Author: Xiaohe Li and Zhicheng Dou 2019/6/12 1 STC-3 CECG @ NTCIR-14 Conversation Generation Task Input: post ( = 1


  1. RUCIR at NTCIR-14 STC-3 CECG Subtask Speaker: Jiaqing Liu School of Information Renmin University of China Author: Xiaohe Li and Zhicheng Dou 2019/6/12 1

  2. STC-3 CECG @ NTCIR-14 Conversation Generation Task ๏ƒ˜ Input: post ( ๐‘Œ = ๐‘ฆ 1 ๐‘ฆ 2 โ‹ฏ ๐‘ฆ ๐‘œ ) ๏ƒ˜ Output: response (with fluency and coherence) Post (Given) Response (to be Generated) ็ˆฑ็‹—่ฟ˜ไผšๅš้ฅญ็š„็”ทไบบ๏ผŒๆœ€ๅธ…ไบ†๏ผ ไผšๅš้ฅญ็š„็”ทไบบๆ˜ฏๅพˆๅธ…็š„ๅ•Šใ€‚ The man who cooks and loves dogs is very The man who cooks is handsome. handsome! 2019/6/12 2

  3. STC-3 CECG @ NTCIR-14 Conversation Generation Task ๏ƒ˜ Input: post ( ๐‘Œ = ๐‘ฆ 1 ๐‘ฆ 2 โ‹ฏ ๐‘ฆ ๐‘œ ) ๏ƒ˜ Output: response (with fluency and coherence) Post (Given) Response (to be Generated) ็ˆฑ็‹—่ฟ˜ไผšๅš้ฅญ็š„็”ทไบบ๏ผŒๆœ€ๅธ…ไบ†๏ผ ไผšๅš้ฅญ็š„็”ทไบบๆ˜ฏๅพˆๅธ…็š„ๅ•Šใ€‚ The man who cooks and loves dogs is very The man who cooks is handsome. handsome! No Not Co Conside ider Em Emotion on (Import ortant ant in Co Convers ersati ation) on) 2019/6/12 3

  4. STC-3 CECG @ NTCIR-14 Emotional Conversation Generation ๏ƒ˜ Input: post & emotion category (of response) {Like, Happiness, Anger, Disgust, Sadness, Other} ๏ƒ˜ Output: response (with fluency and coherence & emotional consistency) Emotion Post (Given) Response (to be Generated) (Given) ็ˆฑ็‹—่ฟ˜ไผšๅš้ฅญ็š„็”ทไบบ๏ผŒๆœ€ๅธ…ไบ†๏ผ ๅ–œๆฌข ไผšๅš้ฅญ็š„็”ทไบบๆ˜ฏๅพˆๅธ…็š„ๅ•Šใ€‚ The man who cooks and loves dogs Like The man who cooks is handsome. is very handsome! 2019/6/12 4

  5. STC-3 CECG @ NTCIR-14 Emotional Conversation Generation ๏ƒ˜ Input: post & emotion category (of response) {Like, Happiness, Anger, Disgust, Sadness, Other} ๏ƒ˜ Output: response (with fluency and coherence & emotional consistency) Emotion Post (Given) Response (to be Generated) (Given) ็ˆฑ็‹—่ฟ˜ไผšๅš้ฅญ็š„็”ทไบบ๏ผŒๆœ€ๅธ…ไบ†๏ผ ๅ–œๆฌข ไผšๅš้ฅญ็š„็”ทไบบๆ˜ฏๅพˆๅธ…็š„ๅ•Šใ€‚ The man who cooks and loves dogs Like The man who cooks is handsome. is very handsome! Goal: l: Genera rate te the Respon onse se with h Sp Specia ial l Em Emotion on 2019/6/12 5

  6. STC-3 CECG @ NTCIR-14 Same me Post: st: Different ferent Res espo ponse se with th Diffe ffere rent nt Emotion otion Emotion Post (Given) Response (to be Generated) (Given) ็ˆฑ็‹—่ฟ˜ไผšๅš้ฅญ็š„็”ทไบบ๏ผŒๆœ€ๅธ…ไบ†๏ผ ๅ–œๆฌข ไผšๅš้ฅญ็š„็”ทไบบๆ˜ฏๅพˆๅธ…็š„ๅ•Šใ€‚ The man who cooks and loves dogs Like The man who cooks is handsome. is very handsome! ็ˆฑ็‹—่ฟ˜ไผšๅš้ฅญ็š„็”ทไบบ๏ผŒๆœ€ๅธ…ไบ†๏ผ ๅŽŒๆถ ไฝ†ๆ˜ฏๆˆ‘็œŸ็š„่ฎจๅŽŒ่ฟ™ๆ ท็š„็”ทไบบ๏ผ The man who cooks and loves dogs Disgust But I really hate such a man! is very handsome! ็ˆฑ็‹—่ฟ˜ไผšๅš้ฅญ็š„็”ทไบบ๏ผŒๆœ€ๅธ…ไบ†๏ผ ๆ‚ฒไผค ๅฅฝไผคๅฟƒ๏ผŒๆˆ‘ๆฒก้‡ๅˆฐ่ฟ‡่ฟ™ๆ ท็š„็”ทไบบใ€‚ The man who cooks and loves dogs Sadness So sad, I have never met such a man. is very handsome! 2019/6/12 6

  7. Model Architecture Emotional Response Generate and Select Reranker 4. Reranker Rule-Based Decoder Decoder Decoder Attention Attention Attention Copy Emotion Mechanism Mechanism Mechanism Mechanism Factor Keywords Encoder Encoder Encoder Extraction ร— ๐‘™ 3. Emotional Seq2Seq 1. Rule-Based 2. Multi-Seq2Seq with Fine Tune Post & Emotion 2019/6/12 7

  8. Rule-Based Method Emotional Response Generate and Select Reranker 4. Reranker Rule-Based Decoder Decoder Decoder Attention Attention Attention Copy Emotion Mechanism Mechanism Mechanism Mechanism Factor Keywords Encoder Encoder Encoder Extraction ร— ๐‘™ 3. Emotional Seq2Seq 1. Rule-Based 2. Multi-Seq2Seq with Fine Tune Post & Emotion 2019/6/12 8

  9. Rule-Based Method ๏ƒ˜ Extract the Keyword in the Post (based on NER) ๏ƒ˜ Fill it into the Proper Constructed Template ๆตทๅ—ๆธธๆ˜ฏ็ ด็ญไบ† [ ๆ€’ ][ ๆ€’ ][ ๆ€’ ] Post Hainan tour is ruined [angry] [angry] [angry] ๅ–œๆฌข ๆœ€ๅ–œๆฌข ๆตทๅ— ไบ† Like I like Hainan most. ้ซ˜ๅ…ด ๆƒณๅˆฐ ๆตทๅ— ๅฐฑๅพˆๅผ€ๅฟƒ Happiness I am very happy when I think of Hainan . ็”Ÿๆฐ” ไธๆƒณๅฌๅˆฐ ๆตทๅ— ๏ผŒๅˆซ่ทŸๆˆ‘ๆ๏ผ I donโ€™t want to hear about Hainan , don't mention it to me! Anger ๅŽŒๆถ ่ถ…็บงไธๅ–œๆฌข ๆตทๅ— ๏ผ Disgust Super dislike Hainan ! ๆ‚ฒไผค ๆตทๅ— ไผค้€ไบ†ๆˆ‘็š„ๅฟƒ Sadness Hainan broke my heart 2019/6/12 9

  10. Multi-Seq2Seq Emotional Response Generate and Select Reranker 4. Reranker Rule-Based Decoder Decoder Decoder Attention Copy Emotion Attention Attention Mechanism Mechanism Mechanism Mechanism Factor Keywords Encoder Encoder Encoder Extraction ร— ๐‘™ 3. Emotional Seq2Seq 1. Rule-Based 2. Multi-Seq2Seq with Fine Tune Post & Emotion 2019/6/12 10

  11. Seq2Seq with Attention + Attention ๐‘‘ 3 ็š„็กฎ ๐‘ 3,1 ๐‘ 3,2 ๐‘ 3,3 ๐‘ 3,4 ๆ˜ฏ็š„ ๅฆ‚ๆญค totally Yeah true <EOS> โ„Ž 1 โ„Ž 2 โ„Ž 3 โ„Ž 4 ๐‘ก 1 ๐‘ก 2 ๐‘ก 3 ๐‘ก 4 ้‚ฃ ๅฐๅญ ็œŸ ๆ˜ฏ็š„ ็š„็กฎ ้…ท <SOS> ๅฆ‚ๆญค That guy is really Yeah totally cool true Decoder Alignment Encoder Model Image Reference: Qiu et al., 2017. Alime chat: A sequence to sequence and rerank based chatbot engine. ACL 2017. 2019/6/12 11

  12. Multi-Seq2Seq with Fine Tune ๏ƒ˜ Train Different Seq2Seq Models for Different Emotions All Data: Encoder Attention Decoder Post Response 2019/6/12 12

  13. Multi-Seq2Seq with Fine Tune ๏ƒ˜ Multi-Seq2Seq: One Model for One Emotion Category All Data: Encoder Attention Decoder Post Response Response Like Data: Encoder Attention Decoder Post (Like) Response Happy Data: Encoder Attention Decoder Post (Happy) Response Anger Data: Decoder Post Encoder Attention (Anger) Response Disgust Data: Encoder Attention Decoder Post (Disgust) Response Sad Data: Decoder Post Encoder Attention (Sad) 2019/6/12 13

  14. Emotional Seq2Seq Emotional Response Generate and Select Reranker 4. Reranker Rule-Based Decoder Decoder Decoder Attention Attention Attention Copy Emotion Mechanism Mechanism Factor Mechanism Mechanism Keywords Encoder Encoder Encoder Extraction ร— ๐‘™ 3. Emotional Seq2Seq 1. Rule-Based 2. Multi-Seq2Seq with Fine Tune Post & Emotion 2019/6/12 14

  15. Emotional Seq2Seq ๏ƒ˜ Idea: Increase the Probability of Emotional Words ๏ƒ˜ Emotional Seq2Seq: One Model for Many Emotion Categories Attention Mechanism Encoder Post Copy-Net Decoder Response Mechanism Emotion Emotion Category Factor ๐‘— Reference: Zhou et al., 2018. Emotional chatting machine: Emotional conversation generation with internal and external memory. AAAI 2018. 2019/6/12 15

  16. Emotional Seq2Seq ๏ƒ˜ Implicit Method: Emotion Factor by Emotion Embedding ( ๐‘“ ๐‘— ) ๐‘ก ๐‘ข = GRU decoder ๐‘ก ๐‘ขโˆ’1 ; [๐‘ง ๐‘ขโˆ’1 , ๐‘‘ ๐‘ข , ๐‘“ ๐‘— ] Attention Mechanism Encoder Post Copy-Net Decoder Response Mechanism Emotion Emotion Category Factor ๐‘— ๐‘“ ๐‘— Reference: Zhou et al., 2018. Emotional chatting machine: Emotional conversation generation with internal and external memory. AAAI 2018. 2019/6/12 16

  17. Emotional Seq2Seq ๏ƒ˜ Explicit Method: Adding Copy Probability of Emotional Word in Emotional Dictionary ( E ) Built by Clustering ๐‘„ ๐‘ง ๐‘ข |๐‘ก ๐‘ข = ๐‘„ ๐‘๐‘ ๐‘— ๐‘ง ๐‘ข |๐‘ก ๐‘ข + ๐‘„ ๐‘“๐‘›๐‘ ๐‘ง ๐‘ข |๐‘ก ๐‘ข , ๐น ๐‘œ๐‘๐‘œ โˆ’ ๐‘“๐‘›๐‘๐‘ข๐‘—๐‘๐‘œ๐‘๐‘š ๐‘ฅ๐‘๐‘ ๐‘’ ๐‘“๐‘›๐‘ ๐‘ง ๐‘ข |๐‘ก ๐‘ข , ๐น = แ‰Š 0, ๐‘„ softmax ๐น๐‘‹ ๐‘“ ๐‘ก ๐‘ข , ๐‘“๐‘›๐‘๐‘ข๐‘—๐‘๐‘œ๐‘๐‘š ๐‘ฅ๐‘๐‘ ๐‘’ Attention Mechanism Encoder Post Copy-Net Decoder Response Mechanism Reference: Zhou et al., 2018. Emotional chatting machine: Emotional conversation generation with internal and external memory. AAAI 2018. 2019/6/12 17

  18. Re-Ranker Emotional Response Generate and Select Reranker 4. Reranker Rule-Based Decoder Decoder Decoder Attention Attention Attention Copy Emotion Mechanism Mechanism Mechanism Mechanism Factor Keywords Encoder Encoder Encoder Extraction ร— ๐‘™ 3. Emotional Seq2Seq 1. Rule-Based 2. Multi-Seq2Seq with Fine Tune Post & Emotion 2019/6/12 18

  19. Re-Ranker ๏ƒ˜ Given the Response Set: How to select the Best Response? Rule-Base Method Generated Response 1 Multi-Seq2Seq Generated Responses Beam Width Emotional Seq2Seq Generated Responses Beam Width 2019/6/12 19

  20. Re-Ranker ๏ƒ˜ Given the Response Set: How to select the Best Response? ๏ƒ˜ Metrics: Emotional Consistency & Coherence & Fluency ๏ƒ˜ Rank by Metrics: Emotion Score + Coherence Score Rule-Base Method Generated Response 1 Multi-Seq2Seq Generated Responses Beam Width Ranked by Generated Probability Emotional Seq2Seq Generated Responses Beam Width 2019/6/12 20

  21. Re-Ranker ๏ƒ˜ Emotion Score based on Emotional Dictionary ๏ƒ˜ Explicit Emotional Word: High Score ๏ƒ˜ Implicit Emotional Word: Low Score ๏ƒ˜ Degree Word (e.g., very, a little, not) : ๏ƒ˜ Strengthen or Weaken or Reverse Emotion Score ไฝ ็œ‹ไธŠๅŽปไธๅคชๅฅฝใ€‚ Post You don't look very good. ๆ‚ฒไผค ๆˆ‘ๆ˜จๆ™šๅคฑ็œ ไบ†ใ€‚ ๏Œ Sadness I lost sleep last night. ๆ‚ฒไผค ๆ˜จๆ™šๅคฑ็œ ไบ†๏ผŒๆˆ‘ๅฅฝ้šพ่ฟ‡ใ€‚ ๏Š Sadness I was so sad about insomnia last night. Dictionary Reference: ๅพ็ณๅฎ , ๆž—้ธฟ้ฃž , ๆฝ˜ๅฎ‡ , ไปปๆƒ  , ้™ˆๅปบ็พŽ : ๆƒ…ๆ„Ÿ่ฏๆฑ‡ๆœฌไฝ“็š„ๆž„้€  . ๆƒ…ๆŠฅๅญฆๆŠฅ 27(2), 180 โ€“ 185 (2008). 2019/6/12 21

  22. Re-Ranker ๏ƒ˜ Coherence Score ๏ƒ˜ The Term Similarity between Response and Post ๏ƒ˜ Count the Number of Same Term (to be improved) ๆˆ‘่Žทๅฅ–ไบ†ใ€‚ Post I won the prize. ้ซ˜ๅ…ด ๆˆ‘ไธบไฝ ๆ„Ÿๅˆฐๅพˆๅผ€ๅฟƒใ€‚ ๏Œ Happiness I am so happy for you. ้ซ˜ๅ…ด ๆˆ‘ไธบไฝ ่Žทๅฅ–่€Œๆ„Ÿๅˆฐๅผ€ๅฟƒใ€‚ ๏Š Happiness I am very happy that you won the prize. 2019/6/12 22

Recommend


More recommend