知識橋接的文本生成算法.pdf

1、知識橋接的文本生成算法李丕績計算機科學與技術學院/人工智能學院南京航空航天大學2023-3-183/18/2023Piji Li,NSC&Dialogue2ChatGPT Mar.14,20233/18/2023Piji Li,Knowledge&Symbolic3GPT-4知識橋接的對話生成符號知識的控制生成ChatGPT后，做什么?3/18/2023Piji Li,NSC&Dialogue4Outline知識橋接的對話生成 共情對話、個性化對話符號知識的控制生成ChatGPT后，做什么?3/18/2023Piji Li,NSC&Dialogue5Outline3/18/2023Piji L

2、i,NSC&Dialogue6Empathetic Dialogue System Empathy is a crucial step towards a more humanized human-machine conversation.Empathetic dialogue generation aims to recognize feelings in the conversation partner and reply accordingly.1.A commonsense knowledge graphConceptNet2.An emotional lexicon NRC_VADC

3、hallengesHumans usually rely on experience and external knowledge to acknowledge and express implicit emotions.Lacking external knowledge makes it difficult to perceive implicit emotions from limited dialogue history.valence(positivenessnegativeness/pleasure displeasure)arousal(activepassive)dominan

4、ce(dominantsubmissive)3/18/2023Piji Li,NSC&Dialogue7Empathetic Dialogue System1.This phenomenon demonstrates that humans need to infer more knowledge to conduct empathetic dialogues.2.External knowledge is essential in acquiring useful emotional knowledge and improving the performance of empathetic

5、dialogue generation.3/18/2023Piji Li,NSC&Dialogue8Empathetic Dialogue SystemModelling emotional dependencies between interlocutors is crucial to enhance the accuracy of external knowledge representation in empathetic dialogues.323/18/2023Piji Li,NSC&Dialogue9Knowledge-aware Empathetic Dialogue Gener

6、ation-KEMPA framework KEMP The early attempt to leverage external knowledge to enhance empathetic dialogue generation.An emotional context encoder and an emotion-dependency decoder Learn the emotional dependencies between the dialogue history and target response with bunches of external emotional co

7、ncepts.Conducted on a benchmark dataset EMPATHETICDIALOGUES(Rashkin et al.,2019),experimental results confirm the effectiveness of KEMP.3/18/2023Piji Li,NSC&Dialogue10Knowledge-aware Empathetic Dialogue Generation-KEMPPreliminaries ConceptNet A large-scale knowledge graph that describes general huma

8、n knowledge in natural language.It comprises 5.9M tuples,3.1M concepts,and 38 relations.NRC_VAD A lexicon of VAD(Valence-Arousal-Dominance)vectors with dimensions for 20k English words.Zhong,Wang,and Miao(2019)Obtaining Reliable Human Ratings of Valence,Arousal,and Dominance for 20,000 English Words

9、.Saif M.Mohammad.ACL 2018.3/18/2023Piji Li,NSC&Dialogue11Knowledge-aware Empathetic Dialogue Generation-KEMPInput:1.Multi-turn Dialogue History2.ConceptNet3.NRC_VADTask DefinitionOutput(two subtasks):1.Predict the emotion expressed in the dialogue context.2.Generate an empathetic response.3/18/2023P

10、iji Li,NSC&Dialogue12Knowledge-aware Empathetic Dialogue Generation-KEMP3/18/2023Piji Li,NSC&Dialogue13ExperimentsDataset EMPATHETICDIALOGUES(Rashkin et al.,2019)Automatic Metrics Emotion Accuracy Perplexity Distinct-1 and Distinct-2Human Metrics Empathy Relevance Fluency3/18/2023Piji Li,NSC&Dialogu

11、e14ExperimentsOur model KEMP outperforms state-of-the-art baselines by a large margin in terms of all automatic metrics.3/18/2023Piji Li,NSC&Dialogue15ExperimentsKEMP obtains the best performance on both Empathy and Relevance scores.There is no obvious difference among models in terms of Fluency.3/1

12、8/2023Piji Li,NSC&Dialogue16Experiments3/18/2023Piji Li,NSC&Dialogue17Experiments3/18/2023Piji Li,NSC&Dialogue18Experiments3/18/2023Piji Li,NSC&Dialogue19SummaryA Knowledge-aware EMPathetic dialogue generation method leverages external knowledge to enhance empathetic dialogue generation.KEMP enhance

13、s the emotion perception and dependencies between dialogue history and empathetic response with bunches of emotion-related concepts.As for the future work,we plan to explore the emotional knowledge embedded in the parameters of large pre-trained language model for empathetic dialogue generation.Qint

14、ong Li,Piji Li,Zhumin Chen,Pengjie Ren and Zhaochun Ren.Knowledge Bridging for Empathetic Dialogue Generation.AAAI 2022.3/18/2023Piji Li,NSC&Dialogue20Knowledge-aware Personalized Dialogue Generation3/18/2023Piji Li,NSC&Dialogue21Knowledge-aware Personalized Dialogue GenerationMinghong Xu,Piji Li,Ha

15、oran Yang,Pengjie Ren,Zhaochun Ren,Zhumin Chen and Jun Ma.A Neural Topical Expansion Framework for Unstructured Persona-oriented Dialogue Generation.ECAI 2020.3/18/2023Piji Li,NSC&Dialogue22Personalized Dialogue GenerationConsistentConsistentI just got back from Disney world.Do you like it?OutputPer

16、sona-Chat Dataset(Zhang et al.2018)Input I love to go to Disney world every year.I love to sing songs from the movie frozen.1.persona2.contextHey buddy,how are you doing?3/18/2023Piji Li,NSC&Dialogue23Personalized Dialogue Generation-ProblemsConsistency Consistency They like to play video games and

17、sing songs from the movie frozen.LogicLogic What is your family like?They are okay,but I like to sing in the park.SOTA 1SOTA 2I love to sing songs from the movie frozen.Persona:EgocentrismEgocentrismConsistencyConsistency1）Show self-persona eagerly while2）Show less interests about the partners.3/18/

18、2023Piji Li,NSC&Dialogue24Personalized Dialogue Generation-ProblemsI love to sing songs from the movie frozen.Great!I like music too and thats why I play guitar!User Experience User Experience Model InteractivityModel InteractivityI have a friend who plays guitar.SOTA 2SOTA 1I love to sing songs fro

19、m the movie frozen.Persona:How old were you when you learned to play?Do you play in band?EgocentrismEgocentrism1）Show self-persona eagerly while2）Show less interests about the partners.ConsistencyConsistency3/18/2023Piji Li,NSC&Dialogue25MotivationPersonalizedSelf/PartnerPersona Expression Egocentri

20、cSelf/PartnerPersona ExpressionThis workPersonalization or Egocentrism?The key difference between personalization and egocentrism lies in:whether the self-persona expression sacrifices its partners.3/18/2023Piji Li,NSC&Dialogue26MethodologyPersonalizedSelf/PartnerPersona expression EgocentricSelf/Pa

21、rtnerPersona expressionThis work1）Balance“answering”and“asking”:Keeping curiosity to your partner.partnermodelteaches2）Balance“speaking”and“listening”:Finding the common ground.3/18/2023Piji Li,NSC&Dialogue27Methodology1.Balance“answering”and“asking”Reinforcement learning by the self-playMutual Bene

22、fit Reward3/18/2023Piji Li,NSC&Dialogue28MethodologyHow to deal with the persona sparsity problem?Concept Set Framework1111111111111111111111111111 1norm&dotdot&norm11(d)Set Intersection(e)Set Distance(a)Concept Set(b)Set Expansion(c)Set UnionVector-Concept Set over a concept vocabularyMatrix-Concep

23、t Similarity from knowledge graphVector-Matrix Calculation-Concept Set Operations3/18/2023Piji Li,NSC&Dialogue29Methodology2.Balance“speaking”and“listening”Concept Copy Mechanism(How)Lead responses around mutual personasCommon Ground Reward(Which)Finding the common ground3/18/2023Piji Li,NSC&Dialogu

24、e30Methodology2.Balance“speaking”and“listening”Concept Copy Mechanism(How)Common Ground Reward(Which)Lead responses around mutual personasFinding the common groundCommon Ground Modeling Geometric ModelingCommon Ground Modeling Geometric ModelingWhere is the optimal location for F Fin?Three points co

25、linear.PartnerPersonaSel fPersonaFuture3/18/2023Piji Li,NSC&Dialogue31ExperimentsChen Xu,Piji Li,Wei Wang,Haoran Yang,Siyun Wang,Chuangbai Xiao.COSPLAY:Concept Set Guided Personalized Dialogue Generation Across Both Party Personas.SIGIR 2022.3/18/2023Piji Li,NSC&Dialogue32Experiments3/18/2023Piji Li

26、,NSC&Dialogue33Character AIhttps:/beta.character.ai/Glow app清華黃民烈聆心智能3/18/2023Piji Li,NSC&Dialogue34Challenge:Long-range Coherence3/18/2023Piji Li,NSC&Dialogue35Challenge:Long-range CoherenceTo produce a coherent story continuation which often involves multiple events,given limited preceding context

27、,a language models(e.g.,GPT-2)need the ability of modeling long-range coherence.Context:Jennifer has a big exam tomorrow.Story:She got so stressed,she pulled an all-nighter.She went into class the next day,weary as can be.Her teacher stated that the test is postponed for next week.Jennifer felt bitt

28、ersweet about itMostafazadeh et al.A Corpus and Evaluation Framework for Deeper Understanding of Commonsense Stories.NAACL 2016.3/18/2023Piji Li,NSC&Dialogue36Model Additional Help?Given story context:1.Extract corresponding event transition path.2.Develop potential ensuing event transition paths.3.

29、The planned paths accordingly guide the text generation model.3/18/2023Piji Li,NSC&Dialogue37Resources for Event Planning1.Commonsense atlas about inferential event description.2.Parameters of pre-trained language model.3.Downstream text generation datasets.1 Radford et al.Language Models are Unsupe

30、rvised Multitask Learners.OpenAI Blog.2 Sap et al.ATOMIC:An Atlas of Machine Commonsense for If-then Reasoning.AAAI 2019.3/18/2023Piji Li,NSC&Dialogue38Two-stage Method3/18/2023Piji Li,NSC&Dialogue39How to Generate High-quality Event Transition Path?1.We prefix-tune a GPT-2 on a large amount of even

31、t paths extracted from commonsense graphs ATOMIC of Planner.2.Then we prefix-tune on training set of the specific task of Planner.Why?Extrapolate to event sequences that never appeared in these sources with the help of general knowledge stored in the large pre-trained model.Li and Liang.Prefix-tunin

32、g:Optimizing continuous prompts for generation.ACL 2021.3/18/2023Piji Li,NSC&Dialogue40How to Use the Planned Event Path for Text Generation?1.Another GPT-2 is fine-tuned on specific downstream dataset.Transformer parameters of Generator2.Work effectively under the supervision of the even transition

33、 path.Event query layer of GeneratorWhy?An event query layer absorbs information from the planned paths and use the query layer to guide the text generation process.3/18/2023Piji Li,NSC&Dialogue41ExperimentDatasets ROCStories EmpatheticDialoguesRQ1:How to develop a better event transition planner?RQ

34、2:Whether the integration of event transition paths enhances the open-ended text generation?RQ3:How do the event transition paths benefit text generation？1 Mostafazadeh et al.A Corpus and Evaluation Framework for Deeper Understanding of Commonsense Stories.NAACL 2016.2 Rashkin et al.Towards Empathet

35、ic Open-domain Conversation Models:a New Benchmark and Dataset.ACL 2019.3/18/2023Piji Li,NSC&Dialogue42Event Transition Planning(RQ1)3/18/2023Piji Li,NSC&Dialogue43Event-path-aware Text Generation(RQ2)3/18/2023Piji Li,NSC&Dialogue44Case3/18/2023Piji Li,NSC&Dialogue45SummaryWe design a coarse-to-fine

36、 framework:-a special-trained event transition planner to explicitly arrange the ensuing events;-an event-path-aware text generator to exploit the event transition guidance for language generation.We investigate two open-ended text generation tasks,i.e.,story completion and dialogue generation.Expli

37、cit arrangement of event transition path facilitates models to generate more coherent and diverse text in open-ended scenery.Our method could be extended to any other language models and open-ended generation tasks.Qintong Li,Piji Li,Wei Bi,Zhaochun Ren,Yuxuan Lai,Lingpeng Kong.Event Transition Plan

38、ning for Open-ended Text Generation.Findings of ACL 2022.3/18/2023Piji Li,NSC&Dialogue46Sentence Semantic Regression for Text Generation Wei Wang,Piji Li and Hai-Tao Zheng.Sentence Semantic Regression for Text Generation.arXiv:2108.02984.Aug.2021.3/18/2023Piji Li,NSC&Dialogue47Sentence Semantic Regr

39、ession for Text Generation Wei Wang,Piji Li and Hai-Tao Zheng.Sentence Semantic Regression for Text Generation.arXiv:2108.02984.Aug.2021.3/18/2023Piji Li,NSC&Dialogue48Surface RealizationDiffusion LM.Xiang Li 2022.知識橋接的對話生成符號知識的控制生成 歌詞創作ChatGPT后，做什么?3/18/2023Piji Li,NSC&Dialogue49Outline3/18/2023Pij

40、i Li,NSC&Dialogue50Symbolic Controlled Generation-SongNet歌詞、詩詞創作慶祝六一！騰訊AI艾靈與王俊凱領唱中國新兒歌點亮關鍵詞藏頭、蘊含的詩詞、對聯生成3/18/2023Piji Li,NSC&Dialogue51SongNet Background Challenges約束：嚴格的格式和模板格式正確、句子完整、押韻合理*關鍵詞埋入：5*5 Deploy王俊凱AI艾靈歌詞創作春節微視春聯紅包故宮騰訊音樂原歌詞：十年之前/我不認識你/你不屬于我/我們還是一樣/陪在一個陌生人左右/走過漸漸熟悉的街頭新配詞：夜深人靜/思念你模樣/多少次孤單/想

41、伴在你身旁/是什么讓我如此幻想/為何會對你那般癡狂3/18/2023Piji Li,NSC&Dialogue53BERT MASKing and Generation方案一：MASK MASK MASK MASK/MASK MASK MASK MASK MASK方案二：十年之前/我不MASK MASK MASK/MASK不屬于我/我們MASK MASK一樣/效果很差：方案一：-的的的的 的是的是的 是的的的的 的的的的的是 的的是是的的的是的 是的是是的是的的的-是是是是 是是是是是 是是是是是 是是是是的是 的是是是是是是是是 的是是是是的的是是方案二（超參：masking rate）：-十

42、年之后 我還認識你 你不等于我 我們還是這樣 在每一個陌生人左右走在漸漸熟悉的街道-百年之約 shall不可諼ta亦行至于yourself咱們就在這兒錯過一切人生活左向me向漸漸熟悉的街角-十年以后 me還得withyouyet不屬withya我們總在這樣 陪伴某個人生in左右想過漸漸熟悉的街巷原因分析：-Non-Autoregressive-Autoregressive原歌詞：十年之前/我不認識你/你不屬于我/我們還是一樣/陪在一個陌生人左右/走過漸漸熟悉的街頭3/18/2023Piji Li,NSC&Dialogue54BackgroundRigid Formats:#words,#sen

43、tences,rhyming rules,etc.Free Formats Generation-Machine Translation-Dialogue Generation-Summary Generation Rigid Formats Generation-Lyrics-SongCi-Sonnet 3/18/2023Piji Li,NSC&Dialogue55Task Definition Input:a rigid format ,:-denotes a place-holder symbol Output:a natural language sentence tally with

44、 C3/18/2023Piji Li,NSC&Dialogue56Task Definition Polishing:Since is arbitrary and flexible,based on the generated result Y,we can build a new format and generate new result,Task target:3/18/2023Piji Li,NSC&Dialogue57FrameworkSongNet3/18/2023Piji Li,NSC&Dialogue58SongNet-Symbols Format and Rhyme Symb

45、ols:-:general tokens-:punctuation characters-:rhyming tokens/positions3/18/2023Piji Li,NSC&Dialogue59SongNet-Symbols Intra-Position Symbols:-:local positions of tokens-:punctuation characters-:should be the ending words-Descending Order:The aim is to improve the sentence integrity by impelling the s

46、ymbols capture the sentence dynamic information,precisely,the sense to end a sequence.3/18/2023Piji Li,NSC&Dialogue60SongNet-Symbols Segment Symbols:-s is the symbol index for sentence Shakespeares Sonnet 116ABAB CDCD EFEF GGRhyme Scheme:3/18/2023Piji Li,NSC&Dialogue61SongNet Attention3/18/2023Piji

47、Li,NSC&Dialogue62SongNet Training Pre-training and Fine-tuning MLE:minimize the negative log-likelihood Polishing:3/18/2023Piji Li,NSC&Dialogue63SongNet Generation We can assign any format and rhyming symbols C.Given C,we obtain P and S automatically.SongNet can conduct generation starting from the

48、special token iteratively until meet the ending marker.beam-search algorithm and truncated top-k sampling3/18/2023Piji Li,NSC&Dialogue64Experiment-Datasets Pre-training-Chinese:News(9200M Chars),Wikipedia(1700M Chars)-English:BooksCorpus(980M words),Wikipedia(2400M words)Fine-tuning-Chinese:SongCi-E

49、nglish:Shakespeares Sonnets3/18/2023Piji Li,NSC&Dialogue65Experiment Evaluation MetricsGeneral-PPL-DistinctDefined-Format:words match with C?-Rhyme:SongCi-rhyming group,Sonnet-“ABAB CDCD”-Sentence Integrity:3/18/2023Piji Li,NSC&Dialogue66Experiment Results3/18/2023Piji Li,NSC&Dialogue67Experiment Re

50、sults100 training samples3/18/2023Piji Li,NSC&Dialogue68Experiment Ablation Analysis3/18/2023Piji Li,NSC&Dialogue69Experiment Parameter Tuning-kTop-k sampling,k=323/18/2023Piji Li,NSC&Dialogue70Experiment Human EvaluationRelevance:+2:all the sentences are relevant to the same topic;+1:partial senten

51、ces are relevant;0:not relevant at all.Fluency:+2:fluent;+1:readable but with some grammar mistakes;0:unreadable.Style:+2:match with SongCi or Sonnet genres;+1:partially match;0:mismatch.3/18/2023Piji Li,NSC&Dialogue71Experiment Cases3/18/2023Piji Li,NSC&Dialogue72Experiment Cases-Polishing3/18/2023

52、Piji Li,NSC&Dialogue73Demo小船槳/桃花輕唱/婉約惹人懷鄉/湖畔旁蟬鳴鶯啼柳響/你在畫舫中央/微風吹亂著青紗帳/是誰輕聲吟唱/一曲婉約惹人懷想/古琴彈到遠方/楊柳搖蕩/荷塘也成雙/思念飛揚/讓記憶生長/只留歲月蒼茫/百轉柔腸/你說好夢何妨/別離還是憂傷/千年癡狂/萬水流觴/我聽得太絕望/卻不見她回望/心慌張/情惆悵/桃花盛開芬芳/落日余暉照的影彷徨/有話怎能藏/它仍舊會迷失瘋狂/笑問君歸向/注定依然愛滄桑/老街兩處散場/石板路旁/再找尋信仰/落葉夕陽/等待那一張/最美麗地模樣/十字街巷/相遇時很漫長/走過白晝荒涼/大雁南賞/繁華盡忘/往日曾經幻想/像晚霞般閃亮/騰訊音樂

53、2023/3/18Piji Li,NLG742023/3/18Piji Li,NLG75騰訊音樂https:/ Li,NSC&Dialogue76SummarySongNet(SongCi+Sonnet+Song)-Rigid formats,rhyming,sentence integrity.Need to be improved:-Relevance,coherence,logic,style,segment,etc.Piji Li,Haisong Zhang,Xiaojiang Liu,and Shuming Shi.Rigid Formats Controlled Text Gene

54、ration.ACL 2020.3/18/2023Piji Li,NSC&Dialogue77ChatGPT知識橋接的對話生成符號知識的控制生成ChatGPT后，做什么?3/18/2023Piji Li,NSC&Dialogue78Outline 推理機制？CoT原因？事實錯誤、邏輯錯誤？為什么會犯錯？Symbolic Knowledge+X？如何融入預訓練？如何約束解碼推理？依賴RLHF能解決一切問題么？不斷的標數據？3/18/2023Piji Li,Knowledge&Symbolic79知識和邏輯 Scaling laws？模型越大效果越好 為什么有的線性？有的任務是涌現？原因？樣本在向量空間有更好的Representation？能記住更多的樣本？小模型？任務相關？3/18/2023Piji Li,Knowledge&Symbolic80涌現能力 如何記憶？如何對model進行增刪查改？生成的內容如何溯源？Webgpt3/18/2023Piji Li,Knowledge&Symbolic81檢索溯源 百事似通 領域專家 如何蒸餾？3/18/2023Piji Li,Knowledge&Symbolic82領域突現3/18/2023Piji Li,Knowledge&Symbolic83THANKS!