Deep Learning for Dialog Nate Kushman Researcher Microsoft Research Labs
Microsoft Research Labs Cambridge, UK Basic Computer Science Research Long-term goal: real-world impact
https://www.microsoft.com/en-us/research/blog/microsoft-unveils-project-brainwave/
Example problem from Gaunt et al. 2017 STRUCTURE PROGRAM LEARNING MACHINE synthesising source code to guide a traditional program descent to synthesize human- Can we use neural networks Machine learning for How can we use gradient interpretable programs? synthesizer?
Representation Learning: Generative Models 𝑑 𝐻 𝑡 𝑗 𝛸 𝑑 𝛸 𝑡 𝜄 𝑦 𝑗 𝑗 ∈ 𝐻 𝐻 ∈ G Multi-Level VAE
observation action Project Malmo:
Agent Applications Services Infrastructure Optimal medication outcomes require a concert of patient, practitioner and health-system insights and actions, in more timely and targeted ways: higher definition healthcare
ech Services? Seamless integration between human and AI agents Example: Customer Care Intelligence Frictionless human-like conversations Why is Dialog Relevant for Fin/T
2016 2015 2014 2013 Speech 2012 2011 2010 2009
2016 2015 2014 2013 Speech 2012 2011 2010 2009
2016 2015 2014 Vision 2013 Speech 2012 2011 2010 2009
2016 2015 Natural Language 2014 2013 2012 Vision 2011 Speech 2010 2009
• Manual engineering for Large amounts of data Many Possible Combinations each new domain Need Either: • “The button is above the menu” “The menu is below the button” Followed by: “Now Click it” Subtlety Matters Context is challenging for three reasons: Three main Challenges: Solution: “Upgrade VS240 driver” Long Distance “I’m printing in powerpoint ” Relationships ” “I run Windows 10” …. “
Struggle with long distance relationships ” Neural Context Representations 9.1 In practice used mostly for chit-chat Requires large amounts of data 8.9 9.0 8.3 2.7 6.2 Excel at Subtlety Symbolic Context Representations Great for long distance relationships Dominant approach in Real-World Systems Requires Engineering per domain Struggle with subtly of meaning Intent: Projector Setup Domain: Technical Support Device: Epson VS240 “
0 0 0 1 T 1 0 0 0 A 0 1 0 0 G 0 0 1 0 G C 0 1 0 0 1 0 0 0 A task: Classify junk or gene? 0 0 0 1 data: DNA sequences T 1 0 0 0 A 0 1 0 0 G Example 0 0 1 0 G C 0 1 0 0
0 0 0 1 T 1 0 0 0 A H t r t H t+1 0 1 0 0 G 0 0 1 0 H t – “an h - dimensional ‘compressed summary’ of t tokens” G C 0 1 0 0 t + 2 NN 1 0 0 0 H t+1 A t + 1 0 0 0 1 T t 1 0 0 0 A t - 1 0 1 0 0 G H t h 0 0 1 0 G C 0 1 0 0
0 0 0 1 T 1 0 0 0 A 0 1 0 0 G 0 0 1 0 G C NN 0 1 0 0 t + 2 H t+2 1 0 0 0 A t + 1 0 0 0 1 T t 1 0 0 0 A t - 1 0 1 0 0 G 0 0 1 0 G C 0 1 0 0
0 0 0 1 T 1 0 0 0 A 0 1 0 0 G 0 0 1 0 G C NN 0 1 0 0 t + 2 H t+2 NN 1 0 0 0 A t + 1 0 0 0 1 T t 1 0 0 0 A t - 1 0 1 0 0 G H t 0 0 1 0 G C 0 1 0 0
0 0 0 1 T 1 0 0 0 A 0 1 0 0 G 0 0 1 0 G C NN 0 1 0 0 t + 2 H t+2 NN 1 0 0 0 A t + 1 NN 0 0 0 1 T t NN 1 0 0 0 A t - 1 NN 0 1 0 0 G NN 0 0 1 0 G C NN 0 1 0 0
0 0 0 0 0 1 … … … [10 2 – 10 3 ] [10 4 – 10 5 ] 1 0 0 0 Aardwolf 0 1 0 0 0 0 0 0 … Aardvark Zymurgy 0 0 ... ... x t 0 1 0 0 embedding 1 0 0 0 0 1 0 0 0 0 1 0 0 0 0 1 G A C T
great great <end> That’s am happy <start> That’s 3. Sequence-to-sequence I 2. Next-token prediction <end> am happy am happy I 1. Sequence classification am happy I
<end> Output: Next Utterance Decoder That’s great <start> Context Neural Input: Preceding dialog context am happy Encoder I
Supervised Learning Faster learning but requires a human to perform the task to provide the “correct” response <end> Output: Next Utterance Decoder That’s great <start> Context Neural Input: Preceding dialog context am happy Encoder I
Reinforcement Learning +10 10 -10 Hard to learn from, but only requires users to provide a rating at the end of a dialog <end> Output: Next Utterance Decoder You’re mean <start> Context Neural Input: Preceding dialog context am happy Encoder I
Datatbase Domain: Movie Are there any action movies Language Intent: Find Slot - Genre: “Action” Understanding to see this weekend? Slot - Date: “this weekend” Dialog Managment Natural Where would Language Query(LOCATION) you like to go? Generation Domain: Movie Intent: Find Slot - Genre: “Action” Slot - Date: “this weekend” Language Dialog How about the Capital Theater? Slot - Location: “Capital Theater” Understanding Management Domain: Movie Intent: Find Slot - Genre: “Action” Slot - Date: “this weekend” Slot – Location: “Capital”
Slot - Date: “this weekend” Slot - Type: “Taiwanese” Slot - Genre: “Action” Slot – Loc : “Oakland” Slot - Price: “cheap” Domain: Restaurant Intent: Find Domain: Movie Intent: Find Understanding Understanding Language Language Are there any action movies Find me a cheap Taiwanese Resturant in Oakland? to see this weekend? • Pipeline Decision • Ontology Based aurant nt in in Domain Domain Movie Restaura
2,5% RNN Error Rate Domain: Movie Traditional (ngram) 9,5% Domain: Movie weekend? Are there any action movies to see this weekend? Are there
Slot - Date: “this weekend” Slot - Genre: “Action” 3,4% RNN Error Rate Domain: Movie Intent: Find Traditional 4,3% (SVM) this weekend? B-Date I-Date see <> <> to B-Genre I-Genre action movies Are there any action movies Domain: Movie to see this weekend? any <> there <> Are <>
Joint RNN Slot – Date: “this weekend” 13,4% Slot – Genre: “Action” Error Rate Separate RNN Intent: Find Domain: Movie 13,7% Intent: Find Domain: Movie <EOS> Understanding this weekend? B-Date I-Date Language see <> <> to B-Genre I-Genre action movies Are there any action movies to see this weekend? any <> there <> Are <>
Domain: Movie Are there any action movies Language Intent: Find Slot - Genre: “Action” Understanding to see this weekend? Slot - Date: “this weekend” Dialog Managment Natural Where would Language Query(LOCATION) you like to go? Generation Domain: Movie Intent: Find Slot - Genre: “Action” Slot - Date: “this weekend” Language Dialog How about the Capital Theater? Slot - Location: “Capital Theater” Understanding Management Domain: Movie Intent: Find Slot - Genre: “Action” Slot - Date: “this weekend” Slot – Location: “Capital”
Slot - Date: “this weekend” Slot - Date: “this weekend” Slot – Location: “Capital” Slot - Genre: “Action” Slot - Genre: “Action” Management Dialog Domain: Movie Intent: Find Domain: Movie Intent: Find Slot - Location: “Capital Theater” Query(LOCATION)
Domain: Movie Are there any action movies Language Intent: Find Slot - Genre: “Action” Understanding to see this weekend? Slot - Date: “this weekend” Dialog Managment Natural Where would Language Query(LOCATION) you like to go? Generation Domain: Movie Intent: Find Slot - Genre: “Action” Slot - Date: “this weekend” Language Dialog How about the Capital Theater? Slot - Location: “Capital Theater” Understanding Management Domain: Movie Intent: Find Slot - Genre: “Action” Slot - Date: “this weekend” Slot – Location: “Capital”
Query(LOCATION) go? to like Generation Language Natural would you Where <S> Where would you like to go?
Management Domain: Movie Managment Domain: Movie Dialog Dialog … … Domain … Query… Slot… Understanding Understanding Generation Language Language Language Natural How about the Capital Theater? you like to go? Are there any action movies Where would to see this weekend?
Management Domain: Movie Managment Domain: Movie Dialog Dialog … … Domain … Query… Slot… Utterance Encoder Utterance Encoder Generation Language Natural How about the Capital Theater? you like to go? Are there any action movies Where would to see this weekend?
Domain: Movie Domain: Movie … … Domain … Query… Slot… Utterance Encoder Utterance Encoder Generation Language Natural How about the Capital Theater? you like to go? Are there any action movies Where would to see this weekend?
Domain: Movie Domain: Movie … … Domain … Query… Slot… Utterance Decoder Utterance Encoder Utterance Encoder How about the Capital Theater? you like to go? Are there any action movies Where would to see this weekend?
Dialog Encoder Context Context Neural Neural Context Context Neural Neural Utterance Decoder Utterance Encoder Utterance Encoder How about the Capital Theater? you like to go? Are there any action movies Where would to see this weekend?
Recommend
More recommend
