14
AI applications Theory vs Practice

AI applications Theory vs Practice - FPT TechInsight · FPT.AI performance 5 Intent detection (English) FPT.AI Wit.ai IBM Watson MS LUIS F1 91.0 92.0 93.2 90.4 Nested entity extraction

  • Upload
    others

  • View
    5

  • Download
    0

Embed Size (px)

Citation preview

Page 1: AI applications Theory vs Practice - FPT TechInsight · FPT.AI performance 5 Intent detection (English) FPT.AI Wit.ai IBM Watson MS LUIS F1 91.0 92.0 93.2 90.4 Nested entity extraction

AI applications

Theory vs Practice

Page 2: AI applications Theory vs Practice - FPT TechInsight · FPT.AI performance 5 Intent detection (English) FPT.AI Wit.ai IBM Watson MS LUIS F1 91.0 92.0 93.2 90.4 Nested entity extraction

AI, Machine Learning & Data Science

2

Page 3: AI applications Theory vs Practice - FPT TechInsight · FPT.AI performance 5 Intent detection (English) FPT.AI Wit.ai IBM Watson MS LUIS F1 91.0 92.0 93.2 90.4 Nested entity extraction

The AI ecosystem

3

Page 4: AI applications Theory vs Practice - FPT TechInsight · FPT.AI performance 5 Intent detection (English) FPT.AI Wit.ai IBM Watson MS LUIS F1 91.0 92.0 93.2 90.4 Nested entity extraction

FPT.AI

4

Speech

Speech recognition

Speech synthesis

Natural language understanding

Language modeling

Entity extraction

Intent detection

Automated conversation

Dialogue management

Chatbot templates

Data connectors

Knowledge

Persistent knowledge bases

Natural query parser

Updaters

Vision

Document recognition

Face recognition

Page 5: AI applications Theory vs Practice - FPT TechInsight · FPT.AI performance 5 Intent detection (English) FPT.AI Wit.ai IBM Watson MS LUIS F1 91.0 92.0 93.2 90.4 Nested entity extraction

FPT.AI performance

5

Intent detection (English) FPT.AI Wit.ai IBM Watson MS LUIS

F1 91.0 92.0 93.2 90.4

Nested entity extraction (Genia) FPT.AI Neural Layered Model*

Layer-1 F1 74.2 71.3

Layer-2 F1 50.2 37.2

*A Neural Layered Model for Nested Named Entity Recognition. Ju, Miwa & Ananiadou. NAACL-HLT 2018.

Vietnamese Speech recognition WER Hybrid End-to-end NN Google

Noisy speech 6.8 10.8 14.8

Reading 7.4 11.3 12.8

ID card recognition Accuracy

ID Number 98.25%

Name 93.75%

Date of birth 98.75%

Address 92.75%

Page 6: AI applications Theory vs Practice - FPT TechInsight · FPT.AI performance 5 Intent detection (English) FPT.AI Wit.ai IBM Watson MS LUIS F1 91.0 92.0 93.2 90.4 Nested entity extraction

The limit of AI

Page 7: AI applications Theory vs Practice - FPT TechInsight · FPT.AI performance 5 Intent detection (English) FPT.AI Wit.ai IBM Watson MS LUIS F1 91.0 92.0 93.2 90.4 Nested entity extraction

The limit of Deep Learning

Page 8: AI applications Theory vs Practice - FPT TechInsight · FPT.AI performance 5 Intent detection (English) FPT.AI Wit.ai IBM Watson MS LUIS F1 91.0 92.0 93.2 90.4 Nested entity extraction

8

Peer group comparison

Page 9: AI applications Theory vs Practice - FPT TechInsight · FPT.AI performance 5 Intent detection (English) FPT.AI Wit.ai IBM Watson MS LUIS F1 91.0 92.0 93.2 90.4 Nested entity extraction

9

Anomaly detection

Page 10: AI applications Theory vs Practice - FPT TechInsight · FPT.AI performance 5 Intent detection (English) FPT.AI Wit.ai IBM Watson MS LUIS F1 91.0 92.0 93.2 90.4 Nested entity extraction

Data management platform

Page 11: AI applications Theory vs Practice - FPT TechInsight · FPT.AI performance 5 Intent detection (English) FPT.AI Wit.ai IBM Watson MS LUIS F1 91.0 92.0 93.2 90.4 Nested entity extraction

• Active user base: 30 million, 2 billion sessions per year.

• Extra service: ads optimization, automatic news tagging.

• Outcome: 13% increased page views for recommended articles, 7% increased time on site (user retention), 20% increased comments (user engagement).

11

Case study: Digital media publisher

Demographic analysis Interest discovery Lookalike discoveryCustomer lifetime value

prediction

Frequent pattern analysis

Personalized product offerings

Customer lifecycle analysis

Dynamic pricing

Page 12: AI applications Theory vs Practice - FPT TechInsight · FPT.AI performance 5 Intent detection (English) FPT.AI Wit.ai IBM Watson MS LUIS F1 91.0 92.0 93.2 90.4 Nested entity extraction

• Over 2000 SKUs of phones, tablets, laptops and 10000+ accessories, 368 ICT shops in Vietnam.

• Yearly revenue $500 million, e-commerce revenue $50 million.

• Extra service: web analytics & A/B testing.

• Outcome: 9% sale from recommendation, 10% increased orders from dynamic pricing, 30% increased accessory cross-sale revenue, 16% increased email campaign CTR.

12

Case study: ICT retailer

Demographic analysis Interest discovery Lookalike discoveryCustomer lifetime value

prediction

Frequent pattern analysisPersonalized product

offeringsCustomer lifecycle

analysisDynamic pricing

Page 13: AI applications Theory vs Practice - FPT TechInsight · FPT.AI performance 5 Intent detection (English) FPT.AI Wit.ai IBM Watson MS LUIS F1 91.0 92.0 93.2 90.4 Nested entity extraction

• C2C marketplace with ~10,000 merchants, 3 million orders per year.

• Gross merchandising value $50 million, approximately 50,000 SKUs.

• Extra service: purchase intent prediction.

• Outcome: 14% sale from recommended products, 17% increased cross sale, 20% increased email campaign CTR, 55% increase in repeat purchase revenue.

13

Case study: Online marketplace

Demographic analysis Interest discovery Lookalike discoveryCustomer lifetime value

prediction

Frequent pattern analysisPersonalized product

offeringsCustomer lifecycle

analysisDynamic pricing

Page 14: AI applications Theory vs Practice - FPT TechInsight · FPT.AI performance 5 Intent detection (English) FPT.AI Wit.ai IBM Watson MS LUIS F1 91.0 92.0 93.2 90.4 Nested entity extraction

14

Thank you