Multilingual, end-to-end-aspect based sentiment analysis with opinion triplets predictions

1. A computer-implemented method for training a neural end-to-end aspect based sentiment analysis (ABSA) system, the computer-implemented method comprising:

inputting a batch of samples of a dataset into the neural end-to-end ABSA system,

wherein the neural end-to-end ABSA system includes:

a contextual language encoder configured to embed tokens with context;

a first self-attention network configured to, based on an output of the contextual language encoder, detect an aspect term and provide a category output corresponding to the aspect term; and

a second self-attention network configured to, based on the output of the contextual language encoder, detect the aspect term and provide a polarity output corresponding to the aspect term; and

based on the inputted batch of samples and a consistency loss function that is based on (a) a first probability of classifying a token as not being an aspect term by the first self-attention network and (b) a second probability of classifying the token as not being an aspect term by the second self-attention network, selectively adjusting weights of the neural end-to-end ABSA system based on consistent aspect term detection by the first self-attention network and the second self-attention network.

2. The computer-implemented method of claim 1 , wherein the dataset includes a multi-lingual dataset.

3. The computer-implemented method of claim 2 , further comprising:

training the end-to-end ABSA system based on a first language; and

selectively adjusting the weights of the end-to-end ABSA system based on a loss associated with samples of a second language that is different than the first language.

4. The computer-implemented method of claim 3 , further comprising selectively adjusting the weights of the end-to-end ABSA system further based on a weight value set based on a similarity between the first and second languages.

5. The computer-implemented method of claim 4 , wherein the weight value is set based on an F-score of a model trained on the first language and evaluated on the second language.

6. The computer-implemented method of claim 1 , wherein the selectively adjusting weights includes selectively adjusting the weights of the first self-attention network and the second self-attention network.

7. The computer-implemented method of claim 6 , wherein the selectively adjusting weights includes selectively adjusting the weights of the first self-attention network and the second self-attention network jointly.

8. The computer-implemented method of claim 1 , wherein the ABSA system is configured to, based on input text associated with one or more tokens, output a triplet including (1) the aspect term in the input text, (2) the category output corresponding to a category associated with the aspect term, and (3) the polarity output corresponding to a polarity associated with the aspect term.

9. The computer-implemented method of claim 1 , wherein the contextual language encoder has the transformer architecture.

10. The computer-implemented method of claim 1 , wherein selectively adjusting weights includes selectively adjusting the weights based on a first loss associated with the first self-attention network and a second loss associated with the second self-attention network.

11. A neural end-to-end aspect based sentiment analysis (ABSA) system comprising:

a contextual language encoder configured to embed tokens with context and generate encodings based on input text;

a first self-attention network configured to, based on an output of the contextual language encoder, detect an aspect term in the input text and to determine a category associated with the aspect term;

a second self-attention network configured to, based on the output of the contextual language encoder, detect the aspect term in the input text and to determine a polarity associated with the aspect term,

wherein the neural end-to-end ABSA system is configured to output a triplet including (1) the aspect term, (2) the category, and (3) the polarity and

one or more processors configured to, based on the input text and a consistency loss function that is based on (a) a first probability of classifying a token as not being an aspect term by the first self-attention network and (b) a second probability of classifying the token as not being an aspect term by the second self -attention network, selectively adjust weights of the neural end-to-end ABSA system based on consistent aspect term detection by the first self-attention network and the second self-attention network.

12. The neural end-to-end ABSA system of claim 11 wherein:

the contextual language encoder is configured to receive a sequence of tokens and to encode the tokens with context to produce representations;

the first self-attention network is configured to, based on the representations, detect the aspect term and determine the category; and

the second self-attention network is configured to, based on the representations, detect the aspect term and determine the polarity.

13. The neural end-to-end ABSA system of claim 11 , wherein the contextual language encoder is trained based on a multilingual dataset.

14. The neural end-to-end ABSA system of claim 11 , wherein the contextual language encoder is configured to receive input text in a first language and is trained using text in a second language that is different than the first language.

15. The neural end-to-end ABSA system of claim 14 , wherein the contextual language encoder is trained based on a weight value set based on a similarity between the first and second languages.

16. The neural end-to-end ABSA system of claim 15 , wherein the weight value is set based on an F-score of a model trained on the second language and evaluated on the first language.

17. The neural end-to-end ABSA system of claim 11 , wherein the contextual language encoder has the transformer architecture.

18. The neural end-to-end ABSA system of claim 11 , wherein the first self-attention network and the second self-attention network are trained jointly.

19. A computer-readable storage medium comprising instructions which, when executed by one or more processors, cause the one or more processors to perform functions including:

inputting a batch of samples of a dataset to a neural end-to-end ABSA system,

wherein the neural end-to-end ABSA system includes:

a contextual language encoder configured to embed tokens with context;

a first self-attention network configured to, based on an output of the contextual language encoder, detect an aspect term and provide a category output corresponding to the aspect term; and

a second self-attention network configured to, based on the output of the contextual language encoder, detect the aspect term and provide a polarity output corresponding to the aspect term; and

based on the inputted batch of samples and a consistency loss function that is based on (a) a first probability of classifying a token as not being an aspect term by the first self-attention network and (b) a second probability of classifying the token as not being an aspect term by the second self-attention network, selectively adjusting weights of the neural end-to-end ABSA system based on consistent aspect term detection by the first self-attention network and the second self-attention network.