- Seed words are chosen by hand. Either start with a large seed lexicon and depend on induction algorithm to fine tune to its specific domain or  use different seed words for different domains

- Computation of embeddings for the negative and positive words(pole words),either off the shelf word2vec embeddings or fine tuned embeddings.

- Create an embedding that represents each pole by taking the centroid(mean) of the embeddings of each of the seed words.

- The semantic axis defined by the poles is computed just by subtracting the pole centroid of positive seed word from the pole centroid of negative seed word.

- The semantic axis is a vector in the direction of positive sentiment.

- Compute ( cosine similarity) the angle between semantic axis and the direction of w’s embedding. A higher cosine means that w is more aligned with the set of embeddings for the positive seed words than embeddings for the negative seed words.


Computation of semantic axis Method

Seed words like good or bad are given.

For a given word, ‘w’ to be labeled ,a measurement of how much each word 'w' is similar to the good word and different from the bad word is calculated.


San Jose State University

This way of learning sentiment lexicons make use of two sets of seed words like good or bad which in turn will find out the similarity of a particular word to any of the two seed word sets.

There are two types of seed-based semi-supervised lexicon induction algorithms,

- Axis-based 

- Graph-based 

Semi-supervised Induction of Affect Lexicons

An on-going but a helpful book resource about NLP
https://web.stanford.edu/~jurafsky/slp3/

Speech and Language Processing (3rd ed. draft) 

Axis-based semi-supervised lexicon induction algorithms

Defining lexicons by propagating sentiment labels on graphs. SentProp (Sentiment Propagation) algorithm is developed on this idea.

Graph based Label Propagation

The core of semisupervised algorithms of induction of affect lexicons lies in the metric of measuring seed word similarities. Besides embedding cosine, there’re other methods.

Learn Before

Related

Learn After