Skip to content

AI

Draft

Customer Feedback Tagging with NLP

Overview

Figure 1 Lanaguage Model in AI

Project

With the data pipeline below to collect, pre-process, feature-engineer, NLP alogorithm applied to provide useful dashboard for analysis and further actions.

Steps

Data Collection:

Data mining or ETL (extract-transform-load) process to collect a corpus of unstructured data.

Data Preprocessing:

  • Tokenization: Segmentation of running text into words.

  • Lemmatization: Removal of inflectional endings to return the base form.

  • Parts-of-speech tagging: Identification of words as nouns, verbs, adjectives etc.

  • Lanaguage detection: Identification of the lanauges from single or several sentensces even a short one.

Feature Engineering (NLP visualization):

  • Word Embeddings: Transforming text into a meaningful vector or array of numbers.

  • N-grams: An unigram is a set of individual words within a document; bi-gram is a set of 2 adjacent words within a document.

  • TF-IDF values: Term-Frequency-Inverse-Document-Frequency is a numerical statistic representing how important a word is to a document within a collection of documents.

Application of NLP Algorithms:

  • Latent Dirichlet Allocation: Topic modeling algorithm for detecting abstract themes from a collection of documents.

  • Support Vector Machine: Classification algorithm for detection of underlying consumer sentiment.

  • Long Short-Term Memory Network: Type of recurrent neural networks for machine translation used in Google Translate.

Scope

  • Topic Modeling: How to automatically categorize customer complaints or intent classification?

  • Sentiment analysis
    • How to detect sentiment from customer feedback, a complaint or a positive feedback?
    • How to detect urgency?

Implementation: WinkNLP

Customize tagging keyword

Implementation: spaCy

Customize tagging keyword

A high level view of generic model and the refine model in the whole process.

The detailed NLP refinement model is as below to improve the models of NER Tagging in spaCy model on user feedback.

Another better idea is Active learning as below

and the whole data pipeline diagram for user feedback tagging is as below

References

spaCy NER

NLP Kits

Label Annotation

ML Backend

Tensorflow.js POC 13: Avatar Generator with Face-API.js

Overview

Figure 1 Computer Vision in AI

Avatar Generator with Face-API.js

This POC, a consequential POC of face-api.js, regonize the face from camera and find the nearest avatar from thousands avatars generated from avatar generators.

References

U2Net

SOD

SOD (Salient Object Detection) is a topics in deep learning that by given a image, SOD can automatically segmentize the most interested objects of the image without any hints. SOD learns how human see the interested objects by detecting the denisity of feature points and segmentize the most dense parts. So far, U2Net provide a state of art performance.

First results of U2Net

These are the first results of the U2Net on target benchmark images. For the full results can be checked in Chimay-SOD1 and asubset Chimay-SOD2 can be found.

{% include ideal-image-slider/slider.html selector="slider1" %}

Image sliders

In this page, image slider for jekyll and its js code is used for image slider. Also a Jekyll Ideal Image Slider Include Demo shows the possiblity of Ideal Image Slider.

References

Recommendation

Overview

Figure 1 DataForecast in AI

pinreset and its pin alogirthm

Amplitude Based Recommendation

amplitude user cohort lists

Here gives a demo for amplitude cohort download and query JSON-Server for Amplitude User Cohorts

References

  • in AI
  • 3 min read

Pytorch POC 2: OpenTTS

Git Repo Status Progress Comments
OpenTTS status progress Pytorch POC #2
mozillatts status progress Pytorch POC #3
MaryTTS status progress Pytorch POC #4

Based on last time keyword spotting topics on Chimay, I even mention items about TTS (text-to-speech) and showed POCs. Here I adopt Opentts to create a API server for speech and later ultrasound generation from Web.

Opentts

In live opentts demo site, you can check the conventional (non-deep learning) speech synthesis (marytts, nanotts) and deep-learning ones (Mozillatts with Tacotron and Tacotron2). Deep-learing ones provide a beeter speech quality. A public MOS test results as below also show similar conclusions.

MOS

Demo wave file as

Demo wave

Swagger API also includes the following:

Opentts swagger

The following diagram is from mozzila project. It shows the whole picture of nature lanaugege iteration with end users. But, of course, it will be a long way to go.

References