The document describes Smart Personal Shopping Assistant (SPA), a system that uses artificial intelligence to act as a personal shopper for online stores. SPA uses speech recognition, image analysis, and question answering to have natural conversations with customers about products. It draws from a knowledge database built using techniques like topic modeling and entity extraction. SPA can identify products from images and place orders based on conversations. The system is designed to scale using SpotDy's BigAITM platform to handle large volumes of requests.
2. Why we need?
A Personal Shopping Assistant is an occupation
where people help customers by giving advice and
making suggestions. They are employed by
departmental Stores.
But you have a Mobile/Web Ecommerce business.
How can you enable your mobile/web application turn
into a smart personal shopper for your customers?
Enter - Smart Personal Shopping Assistant.
Customer
Store
Assistant
Online Store
3. Why we need ?
I need skinny pants that
girls like. My size is 32
inch waist and 34 length.
Here you go. Let me
know If I should filter by
price, size or brand
I like it. My price range is
40-50 dollars.
I need skinny pants that
girls like. 32 inch waist,
34 length.
I like this pant. Let’s buy
it.
Ok, I placed the order.
You should receive your
order by tomorrow. Best
of luck.
4. Why we need ?
Can you place an order
of red skinny pants that I
ordered last year
Do you want the same
size?
Yes
I have placed the order.
You should receive your
pants by tomorrow.
5. Process Overview
ASR Image Q&A
Knowledge Graph/ Image DB
● Speech Recognition
● Image Matching
● Q & A Dialogue
ASR- Acoustic Speech Recognition
Q&A - Question and Answers Dialogue
6. SPA - System Call Flow
Q&A Dialogue and IR
ASR
Image Analysis
Engine
Pre-computed
KD
SpotDy BigAITM
Platform
Image
Text
Voice
IR - Information Retrieval
KD - Knowledge Graph DB
Dialogue/Action
Dialogue/Action
9. Image Matching
● Image Analysis
○ Extract Feature (SURF Feature Extraction)
■ Find keypoints
○ Grouping Descriptors (SURF Feature Descriptor)
■ Keypoints are grouped in descriptors
○ Match images in the precomputed descriptor database.
○ Post Processing
10. ASR
● Speech Recognition
○ Extract Feature vectors
○ Speech Decoder
■ Scoring (DNN)
■ Most Likely Text from Acoustic Model (HMM/Viterbi Algorithm)
○ Pass to Q & A system
12. Q&A Personalization
● Q & A results should be personalized and
aggregated based on:
○ Past user history
○ User Geo/Demo
○ Occasions such as Christmas, Thanksgiving etc ..
SpotDy BigAITM
Query Results
14. SURF (Speeded up Robust Features)
SURF is a feature detection process to examine an
image to extract features, that are unique to the objects
in the image. Based on SIFT but faster.
In our case, it help in retrieving similar products based
on images.
Process Involves :
○ Build Scale Space
○ LoG Approximation
○ Key Point Extraction
○ Generate Features
15. LoG Approximation
● The Laplacian is a 2-D isotropic measure of the 2nd spatial derivative of an
image.
● The Laplacian of an image highlights regions of rapid intensity change to
detect edges.
● Uses Gaussian smoothing filter in order to reduce its sensitivity to noise due
to second derivation
16. GMM/DNN-HMM
HMM is a generative probabilistic model that provides
a framework for modelling time-varying spectral vector
sequences. In our case, we use for speech recognition.
● GMM/DNN produce posterior probabilities for HMM States
● predicts likelihood of observation sequence being generated by
state sequence using Viterbi Algo
● Sub word HMMs concatenate to create larger word-based HMM
Observations (Feature vectors)
GMM/DNN
HMM States
(Senones)
Posterior Probabilities
17. NLP
Knowledge Database (KD) is the key for the query
processing and information retrieval
● NLP is extensively used to process unstructured data in building
KD.
Algorithms:
● Conditional Random Fields/Maxent for POS Tagging, Entity
Extraction, concept tagging etc.
● LDA for topic Analysis and Classification
Q&A Dialogue and IR
Indexed
KD
Product
Catalog
Product
metainfo
NLP Engine
18. Query Processing
Query
Indexed
KD
Annotators/Filters
Results
● User Query pass goes through
various annotators. Some of the
few annotators include :
○ Gazetteer, Lemmatization,
Stemming, POS Tagging,
Entity Extraction
● Query Rewrite
● Search - Similarity (IR). Basic
Algorithms include
○ Vector Space Modelling
○ BM25F
● Result Generation
20. SPA - HA Architecture
Significant computing resources are required while
scaling to millions of requests in real time.
21. BigAITM
BigAITM
is purpose built for the scalability
of applications such as SPA.
● Building KD (Knowledge Database)
● Image Repository Store
● Query Processing
● Scalable Machine Learning Models