Text categorization using natural language processing

A device, comprising:

• one or more memories;
• and
• one or more processors, communicatively coupled to the one or more memories, to:
• identify a plurality of samples of textual content, the plurality of samples of textual content including user-generated reviews of products;
• perform tokenization of the plurality of samples of textual content to generate a respective plurality of tokenized samples using a domain-specific corpus, the domain-specific corpus to identify a mapping of one or more terms from the plurality of samples of textual content to the respective plurality of tokenized samples;
• perform embedding of the respective plurality of tokenized samples using the domain-specific corpus to generate a sample matrix;
• determine groupings of attributes of the sample matrix using a convolutional neural network, the groupings of attributes of the sample matrix being passed to a bidirectional long short-term memory (LSTM) layer;
• determine context relationships between the groupings of attributes of the sample matrix using a bidirectional LSTM technique associated with the bidirectional LSTM layer;
• use a nonparametric paired comparison of respective scores associated with a plurality of predicted labels to determine that a first predicted label, of the plurality of predicted labels, is significant in comparison to a second predicted label of the plurality of predicted labels, wherein the first predicted label and the second predicted label being associated with a sample of textual content, of the plurality of samples of textual content, wherein at least one of the first predicted label or the second predicted label relates to one or more of:
• a sentiment of the sample of textual content,
• a product, of the products, associated with the sample of textual content,
• a service associated with the sample of textual content, or
• an issue associated with the sample of textual content, and
• wherein the respective scores indicate a respective confidence level of a model with regard to a label and are determined based on a Monte Carlo simulation using a dropout technique;
• select the first predicted label using the model and based on the nonparametric paired comparison of the respective scores associated with the plurality of predicted labels;
• and
• provide information identifying the first predicted label to a user for classification.

A non-transitory computer-readable medium storing instructions, the instructions comprising:

• one or more instructions that, when executed by one or more processors, cause the one or more processors to:
• receive a plurality of samples of textual content, the plurality of samples of textual content including user-generated reviews of products;
• perform tokenization of the plurality of samples of textual content to generate a respective plurality of tokenized samples using a domain-specific corpus, the domain-specific corpus to identify a mapping of one or more terms from the plurality of samples of textual content to the respective plurality of tokenized samples;
• perform embedding of the respective plurality of tokenized samples using the domain-specific corpus to generate a sample matrix;
• determine groupings of attributes of the sample matrix using a convolutional neural network, the groupings of attributes of the sample matrix being passed to a bidirectional long short-term memory (LSTM) layer;
• determine context relationships between the groupings of attributes of the sample matrix using a bidirectional LSTM technique associated with the bidirectional LSTM layer;
• use a nonparametric paired comparison of respective scores associated with a plurality of predicted labels to determine that a first predicted label, of the plurality of predicted labels, is significant in comparison to a second predicted label of the plurality of predicted labels, wherein the first predicted label and the second predicted label being associated with a sample of textual content, of the plurality of samples of textual content, wherein at least one of the first predicted label or the second predicted label relates to a sentiment of the sample of textual content, and wherein the respective scores indicate a respective confidence level of a model with regard to a label and are determined based on a Monte Carlo simulation using a dropout technique;
• select the first predicted label using the model and based on the nonparametric paired comparison of the respective scores associated with the plurality of predicted labels;
• and
• provide information identifying the first predicted label.