Beyond Intents! NLU for Conversational UIs Dr. Rebecca Jonsson Head of Research MetaForum 2017, Brussels
Virtual Assistants Personal Assistants Virtual Agents Dialogue systems Chatbots Digital Employees Digital Assistants Natural Language Interaction Solutions Conversational UIs
Siri the personal assistant Apple bought Siri Inc, April 2010 Launched Oct, 2011 (Iphone 4s) Huge media focus Tech Giants Conversational Race started
Battle of the smart home assistants
The battle of the home assistants Copyright Artificial Solutions 2016
Pre-Siri There was Life before Siri! Yellostrom Holly Ikea Anna Elbot Mia Ikea Anna Sandra Ida SAS Eva
Who are we? An end-to-end platform for enterprises to build conversational UIs in multiple languages and for multiple channels and devices. Teneo Platform 100+ conversational systems (VAs, PAs) Make technology understand people in their own language since 2001 Offices Stockholm Barcelona Hamburg Newbury Milano Utrecht Chicago Mountain View 100+ employees 30+ languages Multimodal
Chatbot revolution!
B o t f r a m e w o r k s Wit.ai Api.ai/DialogFlow Luis.ai Watson Conversation Amazon Lex
How does the NLU work? 1. Create an intent 2. Add example inputs for each intent 3. Create an entity type 4. Add values for the entity 5. Intent classifier is trained 6. Entity detector(s) is trained 7. Intents are predicted by model 8. Entities are annotated
How well does intent classification work (with a few examples)? It works surprisingly well! With just 10 examples per intent and with 50 (or even up to 150) intents we have on several data sets been able to achieve quite high accuracy. Intents (classes) the fewer the better Intents - the more distinct the better Examples per intent the more the better
Things intent classification is good at Ill-formed inputs hello in French translate how to setting your voice? Lyra can you tell can you call me Hmmm playsomeof your music can you wake up me Lyra woke me up after 40 minutes get up me at 7 Tale me a joke Robust to errors Horrorscop tomorows horos ope Vam you sebe a texst to 0604470798 pleass anorger joke turn on the flas light Pony farts can you call me remainder me take food at 8clock on everyday Different word order or wordings how do you translate mom in Canada yes you tell me where I am camera and click me a photo capture my photo in front camera Lyra can you please take a snap on me would shorts be good for today raise the volume up have you any music I can hear Superfluous words could you transfer translate good morning in French for me ready go just translate in Spanish you are my heart yes you have something story or jokes or whatever for me can you stop that please read the message last message for me can you send will you call me back Artificial Solutions
Why go beyond intents? Details Matters 'Where can I get a taxi' in Hindi vs Where can I get a Taxi? I want to order a beer vs I want to order a beer in German can you turn ON the alarm vs can you turn OFF the alarm Structure matters can you call me vs can you make a call for me vs can you call me anna Can you call me vs Who called me I want to fly somewhere nice vs I want to fly to Nice Play a song vs What song is playing? Relate words and extract info I want to set the alarm at 7 am and order a taxi for 8 am edit the alarm at 2 p.m. to make it 2 a.m. please translate if I could be in Bangalore could have suffering from fever in Hindi Complex or long structures how many hours until my alarm clock goes off tomorrow morning But if i set off the internet connection, will the alarm be still on? what is taking so long why are you not calling Melissa for me Can you show me who called me in the last hour Multiple intents or coordination Delete the alarm at 9 a.m. And add an alarm at 8 a.m. not send text message but message application please Please switch off alarm and schedule power off from settings Artificial Solutions
NLU - Data driven (ML) vs Rulebased Broad NLU Broader Narrow NLU Broad but shallow Shallow NLU Data-driven approaches Deeper Fine-grained but brittle Rule-based approaches Deep NLU 01 02 Broader Capture diversity, robust to errors and differences, large vocabulary to achieve higher recall. 03 Deeper Precise, fine-grained understanding, subtle of nuances in language to achieve high precision.
NLU - Data driven (ML) vs Rulebased Broad NLU Broader Narrow NLU Broad but shallow Shallow NLU Data-driven approaches Deeper Deep and broad Fine-grained but brittle Rule-based approaches Deep NLU The goal 01 There is no NLU approach today that is both deep and broad. 02 Broader Capture diversity, robust to errors and differences, large vocabulary to achieve higher recall. 03 Deeper Precise, fine-grained understanding, subtle of nuances in language to achieve high precision.
NLU - Data driven (ML) vs Rulebased Broad NLU Broader Narrow NLU Broad but shallow Shallow NLU Data-driven approaches Deeper Deep and broad Hybrid NLU Fine-grained but brittle Rule-based approaches Deep NLU The goal 01 There is no NLU approach today that is both deep and broad. 02 Broader Capture diversity, robust to errors and differences, large vocabulary to achieve higher recall. 03 Deeper Precise, fine-grained understanding, subtle of nuances in language to achieve high precision.
Uses the Teneo Language Resources a z Uses machine learning algorithms Looks for synonyms, lists, phrases i_want.phr flight.syn how do i boo ook a fli lig igh ght Looks for features Needs few example inputs LINGUISTIC H Y B R I D STATISTICAL Needs training data Editable, improved by a human Very precise and predictable 0100101 0 1001010 1 0010011 0 0110010 0 Improves model by retraining Can pick up subtle patterns an analyst might miss Copyright Artificial Solutions 2016
Contextually aware NLU!! Artificial Solutions
Artificial Solutions