Exploiting document knowledge for aspect-level sentiment classification
Methods, systems, and computer-readable storage media for receiving a set of document-level training data including a plurality of documents, each document having a sentiment label associated therewith, receiving a set of aspect-level training data including a plurality of aspects, each aspect having a sentiment label associated therewith, training the aspect-level sentiment classifier including a long short-term memory (LSTM) network, and an output layer using one or more of pretraining, and multi-task learning based on the document-level training data and the aspect-level training data, pretraining including initializing parameters based on pretrained weights that are fine-tuned during training, and multi-task learning including simultaneous training of document-level classification and aspect-level classification, and providing the aspect-level sentiment classifier for classifying one or more aspects in one or more sentences of one or more input documents based on sentiment classes.
1. A computer-implemented method for providing an aspect-level sentiment classifier, the method being executed by one or more processors and comprising:
receiving a set of document-level training data comprising a plurality of documents, each document having a sentiment label associated therewith;
receiving a set of aspect-level training data comprising a plurality of aspects, each aspect having a sentiment label associated therewith;
training the aspect-level sentiment classifier comprising a long short-term memory (LSTM) network, and an output layer using one or more of pretraining, and multi-task learning based on the document-level training data and the aspect-level training data, pretraining comprising initializing parameters based on pretrained weights that are fine-tuned during training, and multi-task learning comprising simultaneous training of document-level classification and aspect-level classification; and
providing the aspect-level sentiment classifier for classifying one or more aspects in one or more sentences of one or more input documents based on sentiment classes.
2. The method of claim 1 , wherein training the aspect-level sentiment classifier using pretraining further comprises, for each document, using a last hidden vector provided from the LSTM network as a document representation that is processed by the output layer.
3. The method of claim 1 , wherein training the aspect-level sentiment classifier using pretraining further comprises, for each document, using a mean hidden vector of a set of hidden vectors provided from the LSTM network as a document representation that is processed by the output layer.
4. The method of claim 1 , wherein training the aspect-level sentiment classifier comprises using pretraining and multi-task learning in combination, where, for each document, a mean hidden vector of a set of hidden vectors provided from the LSTM network being used as a document representation that is processed by the output layer.
5. The method of claim 1 , wherein a training objective for training the aspect-level sentiment classifier comprises cross-entropy minimization.
6. The method of claim 1 , wherein the sentiment classes comprise positive, and negative.
7. The method of claim 1 , wherein training of the aspect-level sentiment classifier comprises initializing parameters using parameters of a trained document-level classifier.
8. A non-transitory computer-readable storage medium coupled to one or more processors and having instructions stored thereon which, when executed by the one or more processors, cause the one or more processors to perform operations for providing an aspect-level sentiment classifier, the operations comprising:
receiving a set of document-level training data comprising a plurality of documents, each document having a sentiment label associated therewith;
receiving a set of aspect-level training data comprising a plurality of aspects, each aspect having a sentiment label associated therewith;
training the aspect-level sentiment classifier comprising a long short-term memory (LSTM) network, and an output layer using one or more of pretraining, and multi-task learning based on the document-level training data and the aspect-level training data, pretraining comprising initializing parameters based on pretrained weights that are fine-tuned during training, and multi-task learning comprising simultaneous training of document-level classification and aspect-level classification; and
providing the aspect-level sentiment classifier for classifying one or more aspects in one or more sentences of one or more input documents based on sentiment classes.
9. The computer-readable storage medium of claim 8 , wherein training the aspect-level sentiment classifier using pretraining further comprises, for each document, using a last hidden vector provided from the LSTM network as a document representation that is processed by the output layer.
10. The computer-readable storage medium of claim 8 , wherein training the aspect-level sentiment classifier using pretraining further comprises, for each document, using a mean hidden vector of a set of hidden vectors provided from the LSTM network as a document representation that is processed by the output layer.
11. The computer-readable storage medium of claim 8 , wherein training the aspect-level sentiment classifier comprises using pretraining and multi-task learning in combination, where, for each document, a mean hidden vector of a set of hidden vectors provided from the LSTM network being used as a document representation that is processed by the output layer.
12. The computer-readable storage medium of claim 8 , wherein a training objective for training the aspect-level sentiment classifier comprises cross-entropy minimization.
13. The computer-readable storage medium of claim 8 , wherein the sentiment classes comprise positive, and negative.
14. The computer-readable storage medium of claim 8 , wherein training of the aspect-level sentiment classifier comprises initializing parameters using parameters of a trained document-level classifier.
15. A system, comprising:
a computing device; and
a computer-readable storage device coupled to the computing device and having instructions stored thereon which, when executed by the computing device, cause the computing device to perform operations for providing an aspect-level sentiment classifier, the operations comprising:
receiving a set of document-level training data comprising a plurality of documents, each document having a sentiment label associated therewith;
receiving a set of aspect-level training data comprising a plurality of aspects, each aspect having a sentiment label associated therewith;
training the aspect-level sentiment classifier comprising a long short-term memory (LSTM) network, and an output layer using one or more of pretraining, and multi-task learning based on the document-level training data and the aspect-level training data, pretraining comprising initializing parameters based on pretrained weights that are fine-tuned during training, and multi-task learning comprising simultaneous training of document-level classification and aspect-level classification; and
providing the aspect-level sentiment classifier for classifying one or more aspects in one or more sentences of one or more input documents based on sentiment classes.
16. The system of claim 15 , wherein training the aspect-level sentiment classifier using pretraining further comprises, for each document, using a last hidden vector provided from the LSTM network as a document representation that is processed by the output layer.
17. The system of claim 15 , wherein training the aspect-level sentiment classifier using pretraining further comprises, for each document, using a mean hidden vector of a set of hidden vectors provided from the LSTM network as a document representation that is processed by the output layer.
18. The system of claim 15 , wherein training the aspect-level sentiment classifier comprises using pretraining and multi-task learning in combination, where, for each document, a mean hidden vector of a set of hidden vectors provided from the LSTM network being used as a document representation that is processed by the output layer.
19. The system of claim 15 , wherein a training objective for training the aspect-level sentiment classifier comprises cross-entropy minimization.
20. The system of claim 15 , wherein the sentiment classes comprise positive, and negative.