Successfully reported this slideshow.
We use your LinkedIn profile and activity data to personalize ads and to show you more relevant ads. You can change your ad preferences anytime.

Clovaと自然言語処理

865 views

Published on

LINE DEVELOPER DAY 2017 C-7

Published in: Technology
  • Be the first to comment

Clovaと自然言語処理

  1. 1. Clovaと自然言語処理(NLP) Toshinori SATO ( overlast ) _ Data Labs
  2. 2. - Clova is one of the forefront of an application of NLP Summary
  3. 3. - Clova is one of the forefront of an application of NLP - Our task is very interesting Summary
  4. 4. - Clova is one of the forefront of an application of NLP - Our task is very interesting - AI platform developing still in “the early stage” Summary
  5. 5. - Clova is one of the forefront of an application of NLP - Our task is very interesting - AI platform developing still in “the early stage” - We want co-workers Summary
  6. 6. - What is Clova - Natural Language Understanding(NLU) - Overview of our NLU task - Challenges of our NLU task - OSS: NEologd - OSS: Resembla - About us Agenda
  7. 7. Natural Language Understanding(NLU) - A subtopic of Natural Language Processing - Extract an intention from unknown input texts - Generate output texts according to the intention - To making a “very wide” and “very deep” system - Exceeds the current state of the art - Necessary for Natural Language User Interface
  8. 8. Overview of our NLU task Same as Natural Language User Interface task
  9. 9. Speech Recognition Overview of our NLU task Same as Natural Language User Interface task
  10. 10. Speech Recognition NLP IR Machine Learning Overview of our NLU task Same as Natural Language User Interface task
  11. 11. Speech Recognition NLP IR Machine Learning Speech Synthesize Overview of our NLU task Same as Natural Language User Interface task
  12. 12. Speech Recognition NLP IR Machine Learning Speech Synthesize Overview of our NLU task Same as Natural Language User Interface task The area spanning three areas is particularly interesting !!
  13. 13. @overlast is - Software Engineer of Data Labs - Japanese Natural Language Processing professional - About 9.5 years of experience - Main developer of Japanese NLU System of Clova - Main developer of NEologd
  14. 14. Input: natural sentence including misrecognition text Multi-domain / multi-turn spoken dialogue systems Word Sense Disambiguation for real-world tasks Challenges of our NLU task NLP between Speech Recognition and Speech Synthesize Output: natural sentence which progress user’s life
  15. 15. Input: natural sentence including misrecognition text - Support recognition results on NEs and unusual context from Natural Language Processing side - Think various situations and various user’s posture - At living room, kitchen, car, entrance and store - When both hands are occupied with child rearing
  16. 16. Multi-domain / multi-turn spoken dialogue systems - Multi-domain and domains will gradually increase - Yesterday, Clova activated the news domain !! - We want to Maintain domain classification performance - Multi-turn dialogue is the basis of conversation - Learn, listen again, clarify details, resolve ambiguity - Difficult to realize normal conversation for humans
  17. 17. Word Sense Disambiguation for real-world tasks - Various commands to realize a command - E.g. In English: Play / In Japanese: 再生, かけて, 流して - How should Clova handle command “掃除機をかけて”
  18. 18. Word Sense Disambiguation for real-world tasks - Various commands to realize a command - E.g. In English: Play / In Japanese: 再生, かけて, 流して - How should Clova handle command “掃除機をかけて” - Activate vacuum cleaner with sound remote control
  19. 19. Word Sense Disambiguation for real-world tasks - Various commands to realize a command - E.g. In English: Play / In Japanese: 再生, かけて, 流して - How should Clova handle command “掃除機をかけて” - Activate vacuum cleaner with sound remote control - Search LINE Music and play “掃除機の音”
  20. 20. Word Sense Disambiguation for real-world tasks - Various commands to realize a command - E.g. In English: Play / In Japanese: 再生, かけて, 流して - How should Clova handle command “掃除機をかけて” - Activate vacuum cleaner with sound remote control - Search LINE Music and play “掃除機の音” - Interpret it as chat, say “私も手が欲しいです”
  21. 21. Output: natural sentence which progress customer’s everyday life - Users unconsciously seek a human friendly response - This fact becomes clear by a voice response - What is a human friendly response? - Output the result desired by the customer - Maintain consistent character image - Accuracy of speech, intonation and accent
  22. 22. Make platform available for business - Realize "natural relationship between people and computers” - in various devices and daily life scenes - Ability to respond to unknown input - Controllable output for each customer - Very fast response speed
  23. 23. Clova also uses NEologd - NEologd is a neologism dictionary generate system - https://github.com/neologd/neologd - Now supported by Data Labs - NEologd provides some OSS language resources - mecab-(ipa|uni)dic-NEologd (updating twice weekly) - neologd-solr-elasticsearch-synonyms
  24. 24. Resembla: Word-based Japanese similar sentence search library - Developed by Takashi Uemura, et al on Data Labs - https://github.com/tuem/resembla (リセンブラ) - Basic feature - (SimString → DP) with multiple setting → Reranking - Utilize the results of Morphological Analysis
  25. 25. Example of an usage of Resembla
  26. 26. 京都/府 東京/都
  27. 27. 京都/府 東京/都
  28. 28. About us
  29. 29. We are Hiring 💖
  30. 30. - Clova is one of the forefront of an application of NLP - Our task is very interesting - AI platform developing still in “the early stage” - We want co-workers Summary

×