towards-building-more-intelligent-chatbots-MinlieHuang

Size: px

Start display at page:

Download "towards-building-more-intelligent-chatbots-MinlieHuang"

像能
5 years ago
Views:

1 Towards Building More Intelligent Chatting Machines Minlie Huang ( 黄民烈 ) CS Department, Tsinghua University aihuang@tsinghua.edu.cn Joint Work with Prof. Xiaoyan Zhu 1

2 Good Examples User: 最幸福的人不是拿到了最好的东西, 而是学会去珍惜手上已经拥有的人 Machine: 是的, 所以我也要珍惜自己的幸福 User: 时间啊, 你这把杀猪刀 Time is a killer Machine: 我是被杀的 I am the one who was killed. User: 怕孤獨, 又捨不下自己的小自由 Machine: 我想去流浪, 但没有人陪我去 2

3 Good Examples User: 你若安好, 便是晴天 Machine: 这句话我爱听, 谢谢你的祝福, 我会继续努力的! User: 雪, 飘舞人, 行走树, 纯美天, 苍茫城, 安静 Snow dancing, people walking. Beautiful trees, boundless sky, quiet city. Machine: 好美的意境 So beautiful scene. User: 欲享受选择之快乐必先经历选择之痛苦 Machine: 是的, 人生就是这样, 人生的意义就是选择 User: 有些命里遇见, 从相识到相知, 到最后的相忘于江湖, 似乎都是命中注定 Machine: 缘来缘去, 这就是人生 3

4 Bad Examples 4 User: 其实, 他没有那么爱你 Ohh, he does not love you that much. Machine: 幸福的一天开始了 A happy day starts. User: 回北京了! 很冷很高兴! Back to Beijing now. So cold so happy! Machine: 什么时候来北京啊? When will you come to Beijing? User: 会做饭的男人是最性感的男人 Mans who can cook are the most sexy. Machine: 会做饭的男人是最性感的女人 Mans who can cook are the most sexy women.

5 Bad Examples User: 猫叔羊毛毡手工作品, 太美好了! Machine: 这是什么品种的猫啊 User: 不得不夸一句广东的饭确实好吃! Machine: 我也喜欢吃臭豆腐 User: 我来自上海 I come from Shanghai. Machine: 真巧, 我来自海南, 老乡啊 So cool, I am from Hainan, the same province! 5

6 Bad Examples 语义理解的问题 Semantic Understanding User: 你家多大 How large is your house? 小冰: 这么快就要爆年龄了 So soon to reveal my age? User: 你是素食主义者还是吃肉啊 Are you vegan or meatatarian? 小冰: 怎么减肥 How can I lose weight? 上下文理解的问题 Context issue User: 早饭不好吃小冰: 那明早一起吃 User: 行你推荐的应该不错小冰: 是吗 User: 哈哈我相信你小冰: 呃呃 6 微软小冰个性身份一致性问题 Inconsistency in personality User: 你几岁了小冰: 我啊反正是九零后 User: 那你是几几年出生的小冰: 错 75年阳历一月出生 User: 你是九零后怎么会75年出生小冰: 生在九零后在深圳只能被当做八零后了

7 Bad Examples (AI Ethics) 7 Picture from Prof. FrankRudzicz, University of Toronto

8 Challenges in Chatting Machines Semantics Logics Consistency Interactiveness Content Quality Personality Identity Language Style Emotion & Sentiment Dialogue Strategy Open-domain, open-topic conversational systems 8

9 Challenges in Chatting Machines One-to-many: one input, many many possible responses Knowledge & Reasoning: real understanding requires various knowledge, world facts, or backgrounds Situational Context u Who are you talking with? Stranger, or friend? u His mood and emotion? u Shared backgrounds that are only accessible by two acquaintances 9

10 Open-domain Chatting Machines Perceiving and Expressing Emotions (AAAI 2018) Behaving More Consistently with Personality Assignment (IJCAI- ECAI 2018) Better Understanding and Generation Using Commonsense Knowledge (IJCAI-ECAI 2018) Behaving More Proactively by Asking Good Questions (ACL 2018) References: 1 Emotional Chatting Machine: Emotional Conversation Generation with Internal and External Memory. AAAI Assigning personality/identity to a chatting machine for coherent conversation generation. IJCAI-ECAI Commonsense Knowledge Aware Conversation Generation with Graph Attention. IJCAI-ECAI Learning to Ask Questions in Open-domain Conversational Systems with Typed 10 Decoders. ACL Generating Informative Responses with Controlled Sentence Function. ACL 2018.

11 Emotion Perception and Expression 11

Emotional Chatting Machine Emotion intelligence is a key human behavior for intelligence (Salovey and Mayer, 1990; Picard and Picard, 1997) Understanding emotion and affect is

12 Emotional Chatting Machine Emotion intelligence is a key human behavior for intelligence (Salovey and Mayer, 1990; Picard and Picard, 1997) Understanding emotion and affect is important for dialogue and conversation u Enhance user performance u Improve user satisfaction u Less breakdowns Rule-based emotion adaptation is widely seen in early dialogue systems 12

Emotional Chatting Machine Perceiving and Expressing emotion by machine Closer to

Classifier Emotion Tagged data Emotional Chatting Machine Post Response Post Response

emotional undercurrent of a conversation and respond with something akin to empathy.

13 Emotional Chatting Machine Perceiving and Expressing emotion by machine Closer to human-level intelligence Social Interaction Data Post Response Post Response Emotion Classifier Emotion Tagged data Emotional Chatting Machine Post Response Post Response Happy Angry Sad Our work was reported by MIT Technology Review, the Guardian, Cankao News, Xinhua News Agency etc. Prof Björn Schuller: an important step towards personal assistants that could read the emotional undercurrent of a conversation and respond with something akin to empathy. 13 Hao Zhou, Minlie Huang, Xiaoyan Zhu, Bing Liu. Emotional Chatting Machine: Emotional Conversation Generation with Internal and External Memory. AAAI 2018.

14 Emotional Chatting Machine Emotion category embedding: High level abstraction of emotions Emotion internal state: Capturing the change of emotion state during decoding Emotion external memory: Treating emotion/generic words differentially 14 Hao Zhou, Minlie Huang, Xiaoyan Zhu, Bing Liu. Emotional Chatting Machine: Emotional Conversation Generation with Internal and External Memory. AAAI 2018.

15 Emotional Chatting Machine 15 Hao Zhou, Minlie Huang, Xiaoyan Zhu, Bing Liu. Emotional Chatting Machine: Emotional Conversation Generation with Internal and External Memory. AAAI 2018.

16 More Examples 情绪 Post: 早上的天气不错啊! 哈哈哈 emotion 哈哈哈 Other 今天是个好天气喜欢 Like 今天的天气不错啊! 悲伤 Sad 天气好, 心情不好厌恶 Disgust 这是哪里啊? 求解释愤怒 Angry 怎么了? 什么情况? 快乐 Happy 今天的天气真好啊! 情绪 Post: 为了解决抄作业的不好习惯, emotion 高数老师这回要出绝招了! Other 这老师是要逆天吗? 喜欢 Like 这老师太有才了! 悲伤 Sad 这样的老师伤不起啊厌恶 Disgust 这老师也太狠了点吧愤怒 Angry 这老师是要逆天啊! 快乐 Happy 哈哈, 这老师太有才了情绪 emotion Post: 送一个这样的蛋糕, 给你美丽的闺蜜吧! Other 这是蛋糕蛋糕吗? 喜欢 Like 好漂亮的蛋糕啊! 悲伤 Sad 为什么我觉得很恶心厌恶 Disgust 这样的蛋糕你忍心吗? 愤怒 Angry 有没有想开网店的? 快乐 Happy 哈哈, 好可爱的蛋糕! 情绪 emotion post: 承诺, 就是一个骗子说给一个傻子听的 Other 承诺就是一个承诺喜欢 Like 承诺是一个专一的傻子悲伤 Sad 我就是一个悲伤的傻子厌恶 Disgust 一个骗子说给一个傻子听愤怒 Angry 不管你信不信, 反正我信了! 快乐 Happy 我就是一个开心的傻子 16 Hao Zhou, Minlie Huang, Xiaoyan Zhu, Bing Liu. Emotional Chatting Machine: Emotional Conversation Generation with Internal and External Memory. AAAI 2018.

17 Behaving More Consistently with Personality 17

18 Personality of Chatting Machine 18 Personality is important for game, custom service, etc. User: 你几岁了小冰: 我啊反正是九零后 User: 那你是几几年出生的小冰: 错 75年阳历一月出生 User: 你是九零后怎么会75年出生小冰: 生在九零后在深圳只能被当做八零后了

19 Personality of Chatting Machine Passing the Turning Test? u Deep semantic understanding u Existing chatting machine lacks identity or personality Personality is a well-defined concept in psychology(norman, 1963; Gosling et al., 2003) Extremely subtle, implicit in language expression: u Age, gender, language, speaking style, level of knowledge, areas of expertise Existing works 19 u Implicit personalization: learn implicit conversation style (Li et al., 2016; Al-Rfou et al., 2016) u Require dialogue data from different users with user attributes tagged

Personality of Chatting Machine Deliver coherent conversations w.r.t. identity/personality Generic DialogueDatafor Training UserA: how old are you?

Personality-coherent Chatbot Generated Dialogues User: how old are you? Machine: I am three years old.

20 Personality of Chatting Machine Deliver coherent conversations w.r.t. identity/personality Generic DialogueDatafor Training UserA: how old are you? UserB: I am six. Pre-specified ChatbotProfile UserA: do you like to play piano? UserB: I play violin. Personality-coherent Chatbot Generated Dialogues User: how old are you? Machine: I am three years old. 20 User: do you like to play piano? Machine: Yes, I play piano.

21 Personality of Chatting Machine Encoder Profile detector Position detector Decoder 21 Qiao Qian, Minlie Huang, Haizhou Zhao, Jingfang Xu, Xiaoyan Zhu. Assigning personality/identity to a chatting machine for coherent conversation generation. IJCAI-ECAI 2018.

Personality of Chatting Machine Post-level evaluation Generated sample responses that exhibit session-level consistency Session-level evaluation Qiao Qian,

22 Personality of Chatting Machine Post-level evaluation Generated sample responses that exhibit session-level consistency Session-level evaluation Qiao Qian, Minlie Huang, Haizhou Zhao, Jingfang Xu, Xiaoyan Zhu. Assigning personality/identity to a chatting machine for coherent conversation generation. IJCAI-ECAI

23 Better Understanding and Generation with Commonsense Knowledge 23

24 Commonsense Knowledge Commonsense knowledge consists of facts about the everyday world, that all humans are expected to know. (Wikipedia) u Lemons are sour u Tree has leafs u Dog has four legs Commonsense Reasoning ~ Winograd Schema Challenge: u The trophy would not fit in the brown suitcase because it was too big. What was too big? u The trophy would not fit in the brown suitcase because it was too small. What was too small? 24

25 Commonsense Knowledge in Chatbots lung disease respiratory disease IsA IsA Prevented_by asthma Caused_by air pollution avoiding triggers Caused_by chest tightness 25

26 Commonsense Knowledge in Chatbots lung disease respiratory disease IsA IsA Prevented_by asthma Caused_by air pollution avoiding triggers Caused_by chest tightness 26

27 Commonsense Knowledge in Chatbots I have an asthma since three years old. Triples in knowledge graph: (lung disease, IsA, asthma ) (asthma, Caused-by, chest tightness) lung disease respiratory disease IsA IsA Prevented_by asthma Caused_by air pollution avoiding triggers Caused_by From ConceptNet 27 chest tightness

28 Commonsense Knowledge in Chatbots Conversation is all about semantic understanding Commonsense knowledge is important for language understanding and generation 28

29 Commonsense Knowledge in Chatbots Output: Because I m a brittle man. Decoding words by attending to knowledge graphs and then to triples Encoding the retrieved knowledgegraphsfor each word 29 Input: why are you so breakable?

30 Commonsense Knowledge in Chatbots Automatic evaluation Manual evaluation (Sign-test, p-value<0.005) 30 Hao Zhou, Tom Yang, Minlie Huang, Haizhou Zhao, Jingfang Xu, Xiaoyan Zhu. Commonsense Knowledge Aware Conversation Generation with Graph Attention. IJCAI-ECAI 2018, Stockholm, Sweden.

31 Generation Examples 1 Post: He proposed March 5th. We will be married October 10th. So 7 months Knowledge: (engaged, RelatedTo, married), (wedding, RelatedTo, married), (miss, DistinctFrom, married), (March, RelatedTo, months) Seq2Seq: I'm not sure how to answer that. I'm not sure how to make a date. MemNet: I'm assuming he's married to a gay marriage? CopyNet: I'm sorry, I'm not sure what you mean by that. CCM: I'm pretty sure he's going to be engaged in the next 6 months. 31

32 Behaving More Proactively by Asking Good Questions 32

33 Asking Questions in Chatbots Asking good questions requires scene understanding Scene: Dining at a restaurant 我昨天晚上去聚餐了 I went to dinner yesterday night. 33 Yansen Wang, Chenyi Liu, Minlie Huang, Liqiang Nie. Learning to ask questions in open-domain conversation systems. ACL 2018.

34 Asking Questions in Chatbots Asking good questions requires scene understanding Scene: Dining at a restaurant 我昨天晚上去聚餐了 I went to dinner yesterday night. Friends? Place? Food? Persons? Bill? WHO WHERE HOW-ABOUT HOW-MANY WHO 34 Yansen Wang, Chenyi Liu, Minlie Huang, Liqiang Nie. Learning to ask questions in open-domain conversation systems. ACL 2018.

35 Asking Questions in Chatbots Responding + asking (Li et al., 2016) Proactive behaviors (Yu et al., 2016) Asking good questions are indication of machine understanding Key differences to traditional question generation: u Different goals: Information seeking vs. Enhancing interactiveness and persistence of human-machine interactions u Various patterns: YES-NO, WH-, HOW-ABOUT, etc. u Topic transition: from topics in post to topics in response 35 Yansen Wang, Chenyi Liu, Minlie Huang, Liqiang Nie. Learning to ask questions in open-domain conversation systems. ACL 2018.

36 Asking Questions in Chatbots A good question is a natural composition of u Interrogatives for using various questioning patterns u Topic words for addressing interesting yet novel topics u Ordinary words for playing grammar or syntactic roles 36

37 Asking Questions in Chatbots Typed decoders: soft typed decoder 37 Yansen Wang, Chenyi Liu, Minlie Huang, Liqiang Nie. Learning to ask questions in open-domain conversation systems. ACL 2018.

38 Asking Questions in Chatbots Typed decoders: hard typed decoder 38 Yansen Wang, Chenyi Liu, Minlie Huang, Liqiang Nie. Learning to ask questions in open-domain conversation systems. ACL 2018.

39 Asking Questions in Chatbots Type prediction at each decoding position

40 Asking Questions in Chatbots Metrics: Appropriateness, richness, willingness 40 Yansen Wang, Chenyi Liu, Minlie Huang, Liqiang Nie. Learning to ask questions in open-domain conversation systems. ACL 2018.

41 Asking Questions in Chatbots 41 Yansen Wang, Chenyi Liu, Minlie Huang, Liqiang Nie. Learning to ask questions in open-domain conversation systems. ACL 2018.

42 Summary Emotion, personality, and knowledge is important to Chatbots Open-domain chatting machine is one of the most challenging AI tasks u Requires the ability of understanding semantics, knowledge, and situational context u Ability of making reasoning u Exhibits consistent personality Still a long way to go: existing generation models are still far from the expectation of real-world applications 42

43 Future Research Problems Multi-modality emotion perception and expression (voice, vision, text) Personality, identity, styleà human-like robot u Introvert or extrovert u Personalized (style, or profile) Learning to learn u Grow up from interactions with human partners and environment 43

44 Thanks for Your Attention Minlie Huang, Tsinghua University Recruiting post-doctors and international PhDs! 44

Microsoft Word doc

Microsoft Word doc 中考英语科考试标准及试卷结构技术指标构想 1 王后雄童祥林 ( 华中师范大学考试研究院, 武汉,430079, 湖北 ) 提要 : 本文从结构模式内容要素能力要素题型要素难度要素分数要素时限要素等方面细致分析了中考英语科试卷结构的