eGitty

Discover The Most Popular Algorithms

CNN + Sentiment Lexicon for Sentiment Analysis

Sentiment lexicons contain some useful sentiment words, they can be used to sentiment classification with neural networks.

In paper: Lexicon-Based Sentiment Convolutional Neural Networks for Online Review Analysis. It used cnn+sentiment lexicons for review sentiment analysis.

The structure looks like:

CNN + Sentiment Lexicon for Sentiment Analysis

In order to understand what is highway network, you can read this tutorial:

A Beginner Introduction to Highway Networks – Machine Learning Tutorial

However, as to this paper, we think sentiment lexicons are not useful to sentiment analysis in this structure. For example, if we remove sentences, cn the remaining positive and negative words classify review polarity correctly?

Sentiment Lexicon is not useful to Sentiment Analysis

If positive reviews contain more positive words and negative reviews contain more negative words, this structure may have a little improvement. However, if positive reviews contain much more negative words, this structure may worse than basic cnn networks.

Three lexicon-based attention mechanism

This paper proposed three lexicon-based attention mechanism, they are built on the distance matrix of postive words and negative words.

\[A[i,j]=|F_p[i,:]-F_n[j:]|\]

Here \(|…|\) is euclidean distance. \(A[i; j]\) is the euclidean distance of the \(i-th\) positive sentiment word and the \(j-th\) negative sentiment word.

Attention 1:

Sentiment Lexicon attention method 1

Attention 2:

Sentiment Lexicon attention method 2

Attention 3:

Sentiment Lexicon attention method 3

From the experiment, we can find attention 3 is much better.

Sentiment Lexicon + CNN experiment result

Leave a Reply

Your email address will not be published. Required fields are marked *