Hackweek 2020 Q1 - Sentiment analysis

HW2020Q1 - Sentiment Analysis.md

👋 Hackweek 2020 Q1 - Sentiment analysis 👀

Goal

Plan

Tone analysis on text like Grammarly.

Result

Offline on-device sentiment analysis on texts.

Why tone matters?

Effective communication involves a variety of factors.

Tech background from Deep Learning's site

Plutchik's wheel of emotions Keywords: joy, trust, fear, surprise, sadness, disgust, anger, anticipation

Framing Sentiment Analysis as a Deep Learning Problem

1. Training a word vector generation model (such as Word2Vec) or loading pretrained word vectors
2. Creating an ID's matrix for our training set (We'll discuss this a bit later)
3. RNN (With LSTM units) graph creation
4. Training
5. Testing

*Word2Vec: words-to-vectors on a cosine distance scale *RNN: Recurrent Neural Networks *LSTM: Long Short Term Memory Units *CNN: Convolutional Neural Networks for Sentence Classification

Pretrained models

• https://tfhub.dev
• https://modelzoo.co/

Cool stuffs around the world

• IBM Watson Tone analyser (first 1000 API call/month is free)
  • Emotion
  • Language style
  • Social tendencies

What we have? (iOS12+)

Natural Language Processing • NaturalLanguage Framework - NLTagger Language Support Matrix

@mattt's findings tweet

• English, French, German, Italian, Portuguese, Spanish support all NLTagger schemes
• Russian and Turkish support all except sentiment
• All other languages support only language / script detection and tokenization

Design