US 8,554,701 B1Granted

Determining sentiment of sentences from customer reviews

AssigneeAmazon Technologies, Inc.

InventorsLogan L. Dillard, Eric B. Fox, Russell A. Dicker

Life of this patent

Estimating the term…

Loading patent…

Abstract

Technologies are described herein for classifying sentences or phrases as expressing positive or negative sentiment based on machine learning from training data comprising sentences manually labeled as to sentiment. A list of terms is generated from the manually labeled sentences and sentiment scores are determined for the terms in the list of terms based on the manually labeled sentences. A collection of sentences or phrases may then be classified as to sentiment utilizing one or more logistic regression classifiers trained on the sentiment scores determined for the terms in the list of terms. The classified collection of sentences may be further analyzed to determine an overall majority sentiment regarding a topic discussed in the sentences and/or to extract specific sentences or phrases expressing a particular sentiment for display to a customer.

Claims28

Dependent claims

A computer-implemented method of classifying sentences as to sentiment, the method comprising executing instructions in a computer system to perform the operations of:

acquiring training data comprising a plurality of sentences labeled as to sentiment;
pre-processing the plurality of labeled sentences from the training data;
generating a list of terms from the plurality of labeled sentences;
determining sentiment scores for the terms in the list of terms utilizing a machine
learning technique based on the plurality of labeled sentences;
receiving a collection of sentences to be classified;
pre-processing the collection of sentences;
classifying each sentence from the collection of sentences as having a neutral sentiment or a non-neutral sentiment utilizing a first logistic regression classifier trained on the sentiment scores determined for the terms in the list of terms;
and
classifying each sentence having a non-neutral sentiment from the collection of sentences as having a positive sentiment or a negative sentiment by applying a second logistic regression classifier trained on the sentiment scores determined for the terms in the list of terms to determine a positive score for the sentence, applying a third logistic regression classifier trained on the sentiment scores determined for the terms in the list of terms to determine a negative score for the sentence, and combining the positive score and the negative score for the sentence.

7 more elements

Dependencies are inferred from the claim text.