End-to-end deep learning

47. Introducing end-to-end learning

Imagine that you want to create a system that analyzes product reviews on the Internet and automatically evaluates whether the product was liked by the reviewer or not. For example, the review “This is a great mop!” must be recognized by such a system as very positive. And the review “This mop of low quality - I bought it in vain” is extremely negative.

The task of recognizing positive and negative opinions is called a “sentiment classification”.
Such a system may look like a “conveyor” consisting of two components:

Parser (parser): a system that annotates text with information describing significant words. For example, you can use the analyzer to denote all adjectives and nouns. The result is an annotated text:
“This is a great _adjective mop _noun !”
Attitude classifier: a trained algorithm that accepts annotated text as input and predicts its emotional coloring. The annotation of the parser significantly contributes to the learning of the algorithm: for example, by attaching more weight to adjectives, the algorithm quickly adjusts to meaningful words, such as “great”, and ignores words that do not carry meaning, for example, the word “this”.

Author's note: The parser provides a much richer text annotation, but this simplified description will be enough to explain the in-depth deep learning.

You can depict the resulting two-component "conveyor" as follows: