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 ( 𝑌 = 𝑦 1 𝑦 2 ⋯ 𝑦 𝑜 )  Output: response (with fluency and coherence) Post (Given) Response (to be Generated) 爱狗还会做饭的男人，最帅了！ 会做饭的男人是很帅的啊。 The man who cooks and loves dogs is very The man who cooks is handsome. handsome! 2019/6/12 2

STC-3 CECG @ NTCIR-14 Conversation Generation Task  Input: post ( 𝑌 = 𝑦 1 𝑦 2 ⋯ 𝑦 𝑜 )  Output: response (with fluency and coherence) Post (Given) Response (to be Generated) 爱狗还会做饭的男人，最帅了！ 会做饭的男人是很帅的啊。 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

STC-3 CECG @ NTCIR-14 Emotional Conversation Generation  Input: post & emotion category (of response) {Like, Happiness, Anger, Disgust, Sadness, Other}  Output: response (with fluency and coherence & emotional consistency) Emotion Post (Given) Response (to be Generated) (Given) 爱狗还会做饭的男人，最帅了！ 喜欢 会做饭的男人是很帅的啊。 The man who cooks and loves dogs Like The man who cooks is handsome. is very handsome! 2019/6/12 4

STC-3 CECG @ NTCIR-14 Emotional Conversation Generation  Input: post & emotion category (of response) {Like, Happiness, Anger, Disgust, Sadness, Other}  Output: response (with fluency and coherence & emotional consistency) Emotion Post (Given) Response (to be Generated) (Given) 爱狗还会做饭的男人，最帅了！ 喜欢 会做饭的男人是很帅的啊。 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

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) 爱狗还会做饭的男人，最帅了！ 喜欢 会做饭的男人是很帅的啊。 The man who cooks and loves dogs Like The man who cooks is handsome. is very handsome! 爱狗还会做饭的男人，最帅了！ 厌恶 但是我真的讨厌这样的男人！ The man who cooks and loves dogs Disgust But I really hate such a man! is very handsome! 爱狗还会做饭的男人，最帅了！ 悲伤 好伤心，我没遇到过这样的男人。 The man who cooks and loves dogs Sadness So sad, I have never met such a man. is very handsome! 2019/6/12 6

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 × 𝑙 3. Emotional Seq2Seq 1. Rule-Based 2. Multi-Seq2Seq with Fine Tune Post & Emotion 2019/6/12 7

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 × 𝑙 3. Emotional Seq2Seq 1. Rule-Based 2. Multi-Seq2Seq with Fine Tune Post & Emotion 2019/6/12 8

Rule-Based Method  Extract the Keyword in the Post (based on NER)  Fill it into the Proper Constructed Template 海南游是破灭了 [ 怒 ][ 怒 ][ 怒 ] Post Hainan tour is ruined [angry] [angry] [angry] 喜欢 最喜欢 海南 了 Like I like Hainan most. 高兴 想到 海南 就很开心 Happiness I am very happy when I think of Hainan . 生气 不想听到 海南 ，别跟我提！ I don’t want to hear about Hainan , don't mention it to me! Anger 厌恶 超级不喜欢 海南 ！ Disgust Super dislike Hainan ! 悲伤 海南 伤透了我的心 Sadness Hainan broke my heart 2019/6/12 9

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 × 𝑙 3. Emotional Seq2Seq 1. Rule-Based 2. Multi-Seq2Seq with Fine Tune Post & Emotion 2019/6/12 10

Seq2Seq with Attention + Attention 𝑑 3 的确 𝑏 3,1 𝑏 3,2 𝑏 3,3 𝑏 3,4 是的 如此 totally Yeah true <EOS> ℎ 1 ℎ 2 ℎ 3 ℎ 4 𝑡 1 𝑡 2 𝑡 3 𝑡 4 那 小子 真 是的 的确 酷 <SOS> 如此 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

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

Multi-Seq2Seq with Fine Tune  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

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 × 𝑙 3. Emotional Seq2Seq 1. Rule-Based 2. Multi-Seq2Seq with Fine Tune Post & Emotion 2019/6/12 14

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

Emotional Seq2Seq  Implicit Method: Emotion Factor by Emotion Embedding ( 𝑓 𝑗 ) 𝑡 𝑢 = GRU decoder 𝑡 𝑢−1 ; [𝑧 𝑢−1 , 𝑑 𝑢 , 𝑓 𝑗 ] Attention Mechanism Encoder Post Copy-Net Decoder Response Mechanism Emotion Emotion Category Factor 𝑗 𝑓 𝑗 Reference: Zhou et al., 2018. Emotional chatting machine: Emotional conversation generation with internal and external memory. AAAI 2018. 2019/6/12 16

Emotional Seq2Seq  Explicit Method: Adding Copy Probability of Emotional Word in Emotional Dictionary ( E ) Built by Clustering 𝑄 𝑧 𝑢 |𝑡 𝑢 = 𝑄 𝑝𝑠𝑗 𝑧 𝑢 |𝑡 𝑢 + 𝑄 𝑓𝑛𝑝 𝑧 𝑢 |𝑡 𝑢 , 𝐹 𝑜𝑝𝑜 − 𝑓𝑛𝑝𝑢𝑗𝑝𝑜𝑏𝑚 𝑥𝑝𝑠𝑒 𝑓𝑛𝑝 𝑧 𝑢 |𝑡 𝑢 , 𝐹 = ቊ 0, 𝑄 softmax 𝐹𝑋 𝑓 𝑡 𝑢 , 𝑓𝑛𝑝𝑢𝑗𝑝𝑜𝑏𝑚 𝑥𝑝𝑠𝑒 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

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 × 𝑙 3. Emotional Seq2Seq 1. Rule-Based 2. Multi-Seq2Seq with Fine Tune Post & Emotion 2019/6/12 18

Re-Ranker  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

Re-Ranker  Given the Response Set: How to select the Best Response?  Metrics: Emotional Consistency & Coherence & Fluency  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

Re-Ranker  Emotion Score based on Emotional Dictionary  Explicit Emotional Word: High Score  Implicit Emotional Word: Low Score  Degree Word (e.g., very, a little, not) :  Strengthen or Weaken or Reverse Emotion Score 你看上去不太好。 Post You don't look very good. 悲伤 我昨晚失眠了。  Sadness I lost sleep last night. 悲伤 昨晚失眠了，我好难过。  Sadness I was so sad about insomnia last night. Dictionary Reference: 徐琳宏 , 林鸿飞 , 潘宇 , 任惠 , 陈建美 : 情感词汇本体的构造 . 情报学报 27(2), 180 – 185 (2008). 2019/6/12 21

Re-Ranker  Coherence Score  The Term Similarity between Response and Post  Count the Number of Same Term (to be improved) 我获奖了。 Post I won the prize. 高兴 我为你感到很开心。  Happiness I am so happy for you. 高兴 我为你获奖而感到开心。  Happiness I am very happy that you won the prize. 2019/6/12 22