Share
Embed
Transcript
25 slides
-
Towards Universal Dialogue State Tracking Overview • “Towards Universal Dialogue
State Tracking” • Liliang Ren, Kaige Xie, Lu Chen and Kai Yu (Shanghai Jiao Tong University) • EMNLP 2018 Oral • Contributions • Dialogue State Trackingਸ ਤೠ “StateNet” ઁউ • DSTC2, WoZ datasetীࢲ state-of-the-art ࢿמ -
Dialogue System 1. Introduction Natural Language Understanding (NLU) Dialogue Management
(DM) Natural Language Generation (NLG) Response Query -
Dialogue System 1. Introduction Natural Language Understanding (NLU) Dialogue Management
(DM) Natural Language Generation (NLG) Response Query Dialogue Management (DM) -
Dialogue Management 1. Introduction • DM ѱ 2о ࣌ਸ ыҊ
1. Dialogue State Tracking (DST) • അ ചо যڌѱ ൗ۞оҊ ח ঈ • Әө ചܳ ా೧ ࠁܳ ӝ߈ਵ۽ ࢚కܳ ҳೞҊ ਬ 2. Decision Making • അ ࢚కী ঌݏ ঘ࣌ਸ Policyী ٮۄ Ѿ -
Dialogue State Tracking 1. Introduction • ച ࢚కܳ ୶ೞח NLP,
Dialogue ࠙ঠ ੋ ޙઁ • Task-oriented Dialogueܳ ۽ ೣ • ݾ = User Goalী ب׳ೞҊ Requestܳ ೧Ѿ೧ח Ѫ • Ex. ۨझషی ୶ୌਸ ਤೠ ച दझమ • ݾ = Userо ਗೞח ۨझషی ઑѤਸ ইղח Ѫ • DSTח ۠ Goalਸ ੜ ইоח ݒ ఢ݃ ഛੋೞח Ѫ -
Dialogue State Tracking Challenge 2 (DSTC2) 1. Introduction • Target
Label = (act, slot, value) • act = {inform, request} • slot of inform = {food, pricerange, name, area} • food = {korean, chinese, japanese, french, halal, …} # 91ѐ • pricerange = {cheap, moderate, expensive} • name = {оѱ ܴ} # 113ѐ • area = {north, south, east, west, centre} • slot of request = slot of inform + {addr, phone, postcode, signature} • Evaluation Metric = (Joint) Accuracy -
Limitations of Previous Works 1. Introduction • Slot Valuesо زਵ۽
߸ೞח ജ҃ীࢲ ੜ زೞ ঋ • ҙҟ ࠁ दझమী ࢜۽ Value(ۨझషی, ഐభ ١) ୶оؼ ࣻ • Slot ѐࣻী ࠺۹೧ࢲ ݽ؛ ۄఠ ӝо ழ • Slot݃ ݽ؛ਸ ݅ٞ • Hand-crafted Lexicon Featureܳ ӝ߈ਵ۽ ೣ • ী۞ , ࣘب ೞ -
StateNet: A Universal Dialogue State Tracker 2. Method Utterance Machine
Act Slot Values for Slot Probability Distribution of Values -
StateNet: A Universal Dialogue State Tracker 2. Method “ೠध જਸ
Ѫ эই” request(food) food {ೠध, ध, …} [ೠध: 0.8, ध: 0.1 …] -
Utterance Representation 2. Method • Multi-scale Receptors Layerܳ ా೧ n-gram
ӝ߈ repr ࢤࢿ = 1-gram, 2-gram, …, n-gram reprܳ Sumೠ Ѫ • K-gram Representation • kѐ ױਤ seq of wordsܳ ҳࢿೞח word vector • ਤ word vector ਸ ݽف concat • Concat ӝ߈ӝ ٸޙী kী ٮۄࢲ ߭ఠ ӡо ׳ۄઉࢲ Linear కਕࢲ ӡ ݏ -
Machine Act Representation 2. Method • Machine Actח दझమ
ߊച ӝ߈ غח Dialogue Act • e.g., welcomemsg, request, canthelp, inform, offer ١ • ਤ ױয(pre-defined)ী ೧ࢲ Bag-of-words۽ അ • റ োਸ ਤ೧ Linear కਕࢲ ରਗਸ Utterance Representationҗ ݏ -
Slot Information Decoding 2. Method • Slotী ೠ ࠁܳ ೣԋ
֍যષ • Slot ܴ ী ೠ pre-trained word embeddingਸ ഝਊ • द ରਗਸ ݏӝ ਤ೧ Linear క -
Turn-level Feature Vector 2. Method • খࢲ ҳೠ Utterance, Machine
Act, Slotਸ ઙೠ Vector • Utterance৬ Machine Actח concatೞҊ Ӓ Ѿҗী Slotਸ point-wise multiplicationೣ • ୭ઙਵ۽ ҳೠ Turn-level Feature Vectorח LSTM Input ؽ -
2-Norm Distance 2. Method • LSTM output, Context৬ Slotਸ
Ҋ۰ೠ reprਸ ҳೣ • ী ೧ࢲ оמೠ ݽٚ Valueٜҗ ਬࢎب(L2)ܳ ҳೣ • Ѿҗী ೧ࢲ Softmax ஂ೧ࢲ ୭ઙ Slotী ೠ Value Prediction • Slotী ೠ Valueо (ই) হਸ ࣻ ӝী “None”ਸ Value Setী ୶о • Әө җਸ п ఢ݃ ݽٚ Slot(inform)ী ೧ ೯ [ೠध: 0.8, ध: 0.1 …] {ೠध, ध, …} food Context -
Performance: Joint Goal Accuracy 3. Experiments • Parameter Sharing among
the Different Slots • োҳח Slot݃ ݽ؛ਸ ٜ݅חؘ ܳ ӓࠂ • Initialization with a Pre-trained Model • ೞա Slotী ೧ࢲ Pre-trainingਸ ೞҊ, Ѿҗ Weight۽ ݽٚ Slotী ೠ ݽ؛ਸ Initialization -
Initialization with a Pre-trained Model 3. Experiments • Slot
foodী ೧ࢲ pre-trainingਸ ೮ਸ ٸ о જও • ܲ Slotী ࠺೧ foodо о য۵ӝ ٸޙী pre-trainingী ೠ ٙ о ѱ ইקө ୶ஏ • ੌઙ Weak Slotী ೠ Boosting ѐ֛ਵ۽ ࢤп೧ࠅ ࣻ (Ӓր ੜ ޅ ݏ୶ח Ѣ(Weakness)ী ೧ णਸ ט۰ࢲ ࢿמਸ ֫ੋ ࣅ = Boosting) -
Contributions 4. Conclusion • ݽ؛ٜ ೠ҅ܳ ӓࠂೞח StateNet ઁউ
• زਵ۽ Slot Valueо ߸ೡ ٸ ੜ زೞ ঋ • Slotী ٮܲ ѐ߹ ݽ؛۽ ੋೠ ۄఠ ন ૐо • Hand-crafted Lexicon Feature ࢎਊ • DSTC2, WoZ datasetী ೧ࢲ state-of-the-art ࢿמਸ ࠁ -
хࢎפ✌ ୶о ޙ ژח ҾӘೠ ݶ ઁٚ ইې োۅ۽
োۅ ࣁਃ! ഘ (ML Engineer, Pingpong) Email. [email protected] Facebook. @roomylee Linked in. @roomylee