Open domain targeted sentiment classification using semisupervised dynamic generation of feature attributes

A method for determining sentiment in a dataset of micro-blogs, comprising:

• word embedding, with a computing system having circuitry configured for modelling and processing, the dataset of micro-blogs;
• normalizing the word embedded dataset of micro-blogs;
• formatting the normalized word embedded dataset of micro-blogs to generate a formatted dataset;
• classifying the formatted dataset by applying the applying the formatted dataset to a trained SVM HMM;
• and
• outputting a sentiment analysis of the dataset of micro-blogs.

A non-transitory computer readable medium having instructions stored therein that, when executed by one or more processors, causes the one or more processors to perform a semi-supervised learning method for generating feature attributes in open domain targeted sentiment classification of a dataset of microblogs, comprising:

• training, with a computing system having circuitry configured for modelling and processing, a hidden Markov model support vector machine (SVM HMM) by the steps of:
• dynamically generating feature attributes by
• combining pre-trained word embeddings from a plurality of different sources to form a first training dataset;
• normalizing the first training dataset to generate a normalized training dataset;
• concatenating, with a summer, discrete features with the normalized training set to form a second training dataset;
• formatting the second training dataset;
• initializing a C parameter;
• applying the second training dataset and the C parameter to a SVM HMM;
• training the SVM HMM to classify the second training dataset set by word frequency and sentiment;
• generating a development dataset by selecting a subset of the second dataset;
• classifying the development dataset with the SVM HMM;
• calculating a first zero/one error of the classified development dataset;
• comparing the first zero/one error to an initial zero/one error value to determine whether the first zero/one error is less than the threshold;
• if the first zero/one error equals the initial zero/one error value, the stop training the SVM HMM;
• if the first zero/one error is less than the initial zero/one error value,
• increasing the C parameter;
• reclassifying the development dataset;
• calculating a second zero/one error using the reclassified development dataset;
• comparing the second zero/one error to the first zero/one error value to determine whether the second zero/one error is less than the first zero/one error;
• if the second zero/one error equals the first zero/one error value, stop training the SVM HMM;
• if the second zero/one error is less than the first zero/one error value,
• repeating the steps of increasing the value of the C parameter, reclassifying the development set, and calculating the zero/one error until a current zero/one error equals a previous zero/one error, then
• stop training the SVM HMM;
• identifying, by the computing system, sentiment polarities within a dataset of micro-blogs, by
• word embeddings extracted from the dataset of micro-blogs;
• normalizing the word embedded dataset of micro-blogs;
• formatting the normalized word embedded dataset of micro-blogs to form a formatted dataset;
• classifying the formatted dataset by applying the formatted dataset to the trained SVM HMM;
• and
• outputting a sentiment analysis of the dataset of micro-blogs.