Predicting topic sentiment using a machine learning model trained with observations in which the topics are masked
A facility for determining sentiments expressed by a natural-language text string for each of one or more topics is described. In the natural-language text string, the facility identifies one or more topics. For each identified topic, the facility replaces the topic in the natural-language text string with a masking tag that occupies the same position in the natural-language text string as the topic. After the replacing, the facility applies a machine learning model to the natural-language text string to obtain a predicted sentiment for each of the identified topics.
1 . One or more memories collectively having contents configured to cause a computing system to perform a method, the one or more memories not constituting a propagating transitory data signal, the method comprising:
receiving, by a processor of the computing system, a plurality of training natural-language text strings;
identifying, by the processor of the computing system, one or more noun phrases in each of the plurality of training natural-language text strings;
determining, by the processor of the computing system, one or more sentiment-qualified topics based on the one or more noun phrases;
determining, by the processor of the computing system, an entity class for each of the one or more sentiment-qualified topics;
modifying, by the processor of the computing system, each training natural-language text string to preserve a location of each sentiment-qualified topic in each training natural-language text string while replacing all information about an identity of each sentiment-qualified topic;
using, by the processor of the computing system, the modified plurality of training natural-language strings and the entity class for each of the one or more sentiment-qualified topics to train a machine learning model to determine a sentiment of each sentiment-qualified topic by at least performing a largest-first comparison of substrings of each of the plurality of training natural-language text strings to identify longer, multi-word topics to an exclusion of included shorter topics, and comparing the longer, multi-word topics to a list of topics or named entities specified for a domain; and
storing, by the processor of the computing system, the sentiment with each training natural-language text string by restoring the identity and using the location associated with each sentiment-qualified topic in each training natural-language text string.
2 . The one or more memories of claim 1 , the method further comprising:
initializing the machine learning model as a Bidirectional Encoder Representations from Transformers model.
3 . The one or more memories of claim 1 , the method further comprising:
initializing the machine learning model as a Clinical Bidirectional Encoder Representations from Transformers model.
4 . A method in a computing system, the method comprising:
receiving, by a processor of the computing system, a plurality of training natural-language text strings;
identifying, by the processor of the computing system, one or more noun phrases in each of the plurality of training natural-language text strings;
determining, by the processor of the computing system, one or more sentiment-qualified topics based on the one or more noun phrases;
determining, by the processor of the computing system, an entity class for each of the one or more sentiment-qualified topics;
modifying, by the processor of the computing system, each training natural-language text string to preserve a location of each sentiment-qualified topic in each training natural-language text string while replacing all information about an identity of each sentiment-qualified topic;
using, by the processor of the computing system, the modified plurality of training natural-language strings and the entity class for each of the one or more sentiment-qualified topics to train a machine learning model to determine a sentiment of each sentiment-qualified topic by at least performing a largest-first comparison of substrings of each of the plurality of training natural-language text strings to identify longer, multi-word topics to an exclusion of included shorter topics, and comparing the longer, multi-word topics to a list of topics or named entities specified for a domain; and
storing, by the processor of the computing system, the sentiment with each training natural-language text string by restoring the identity and using the location associated with each sentiment-qualified topic in each training natural-language text string.
5 . The method of claim 4 , the method further comprising:
initializing the machine learning model as a Bidirectional Encoder Representations from Transformers model.
6 . The method of claim 4 , the method further comprising:
initializing the machine learning model as a Clinical Bidirectional Encoder Representations from Transformers model.
7 . The method of claim 4 , wherein the topic to which the training natural-language text strings is related is a distinguished domain of expression, the method further comprising:
receiving a subject input text string related to the distinguished domain of expression;
masking an intent expressed by the subject text string for each of one or more masked noun phrases and one or more entity classes determined for each of the masked noun phrases for which an intent is identified; and
applying the trained model to the masked subject input text string to predict a sentiment for each masked noun phrase of the subject input text string.
8 . A system comprising:
at least one processor; and
at least one memory, the at least one memory not constituting a transitory propagating data signal, wherein the at least one memory is configured to cause the at least one processor to:
receive a plurality of training natural-language text strings;
identify one or more noun phrases in each of the plurality of training natural-language text strings;
determine one or more sentiment-qualified topics based on the one or more noun phrases;
determine an entity class for each of the one or more sentiment-qualified topics:
modify each training natural-language text string to preserve a location of each sentiment-qualified topic in each training natural-language text string while replacing all information about an identity of each sentiment-qualified topic;
use the modified plurality of training natural-language strings and the entity class for each of the one or more sentiment-qualified topics to train a machine learning model to determine a sentiment of each sentiment-qualified topic by at least performing a largest-first comparison of substrings of each of the plurality of training natural-language text strings to identify longer, multi-word topics to an exclusion of included shorter topics, and comparing the longer, multi-word topics to a list of topics or named entities specified for a domain; and
store the sentiment with each training natural-language text string by restoring the identity and using the location associated with each sentiment-qualified topic in each training natural-language text string.
9 . The system of claim 8 , wherein the at least one processor is further configured to initialize the machine learning model as a Bidirectional Encoder Representations from Transformers model.
10 . The system of claim 8 , wherein the at least one processor is further configured to initialize the machine learning model as a Clinical Bidirectional Encoder Representations from Transformers model.
11 . The system of claim 8 , wherein the topic to which the training natural-language text strings is related is a distinguished domain of expression, wherein the at least one processor is further configured to:
receive a subject input text string related to the distinguished domain of expression;
mask an intent expressed by the subject text string for each of one or more masked noun phrases and one or more entity classes determined for each of the masked noun phrases for which an intent is identified; and
apply the trained model to the masked subject input text string to predict a sentiment for each masked noun phrase of the subject input text string.