Topic models with sentiment priors based on distributed representations

A device for topic modeling, the device comprising:

• a memory storing a topic model and instructions;
• and
• a processor which, by executing the instructions, is configured to:
• input a review into the topic model comprising sentiment priors;
• determine, by using the topic model, representations for each word in the review, wherein the representations are word vectors in a semantic space;
• predict user attributes by the topic model using the sentiment priors trained based on the representations distributed by using a regularizer, wherein the user attributes comprise at least one a location, a gender, and an age of a user;
• update the sentiment priors based on the predicted user attributes;
• estimate, by using the topic model, the representations, which are determined as the word vectors in the semantic space, using the sentiment priors which are included in the topic model and trained based on representations distributed using the regularizer which defines a same sentiment to words having similar word vectors;
• and
• determine a topic corresponding to the review based on the estimated representations,
• wherein each of the sentiment priors trained in the topic model is the same for the words having the similar word vectors, and
• wherein, for training the sentiment priors, the processor is further configured to:
• input a set of words into the topic model;
• and
• by using the topic model:
• determine a set of nearest neighbors for each of the words of the set;
• determine a representation for each of the words of the set, wherein the representation is a word vector in the semantic space;
• calculate a regularization coefficient of the regularizer to define the regularizer by estimating with a counter, for each of the words of the set, probabilities of occurring of each of the words of the set in sentiment by using the representation and the set of nearest neighbors of each of the words of the set, respectively;
• define the same sentiment to words having similar word vectors by using the regularizer;
• distribute the representations according to results of the defining;
• and calculate the sentiment priors using the distributed representations.