Aspect-aware sentiment analysis of user reviews supporting AI-based decision making
Rating prediction systems and methods include extracting aspect-sentiment pairs from an input text. An attention-property-aware rating is estimated for the input text using the extracted aspect-sentiment pairs with a neural network that captures implicit and explicit features of the text. A response to the input text is performed based on the estimated rating.
1. A computer-implemented rating prediction method, comprising:
extracting aspect-sentiment pairs from an input text;
estimating an attention-property-aware rating for the input text using the extracted aspect-sentiment pairs with a neural network that captures implicit and explicit features of the text, including encoding item reviews and user reviews into respective explicit and implicit features and generating the rating as
ŝ u,t =b u +b t + ex ([ v u ;v t ])+<γ, im ([ G u ;G t ])>
where b u is a user bias, b t is an item bias, (⋅) is an explicit channel model, (⋅) is an implicit channel model, v u and v t are respective user- and review-based token-level features, G u is a matrix of a user attention representation of an aspect, and G t is a matrix of an item attention representation of the aspect; and
performing a response to the input text based on the estimated rating.
2. The method of claim 1 , wherein extracting aspect-sentiment pairs includes determining a set of sentiment terms using a plurality of different sources.
3. The method of claim 2 , wherein the plurality of different sources include a sentiment lexicon, a pointwise mutual information measurement of statistical word associations within a text corpus, and a coextraction model.
4. The method of claim 1 , wherein extracting aspect-sentiment pairs includes identifying grammatical dependency relationships between words in the input text to generate candidate aspect-sentiment pairs.
5. The method of claim 4 , wherein extracting aspect-sentiment pairs further includes filtering the candidate aspect-sentiment pairs to exclude pairs that do not include an aspect from a predetermined set of sentiment terms.
6. The method of claim 1 , wherein explicit features characterize mentions of sentiment words and wherein implicit features characterize sentiment expressions that do not use sentiment words.
7. The method of claim 1 , further comprising training the neural network using a training dataset that includes a set of written reviews and corresponding numerical ratings.
8. The method of claim 1 , wherein performing the response includes identifying a product that a user would rate highly, based on the attention-property-aware rating, and recommending the product to the user.
9. A rating prediction system, comprising:
a hardware processor; and
a memory that stores a computer program, which, when executed by the hardware processor, causes the hardware processor to:
extract aspect-sentiment pairs from an input text;
estimate an attention-property-aware rating for the input text using the extracted aspect-sentiment pairs with a neural network that captures implicit and explicit features of the text, including encoding item reviews and user reviews into respective explicit and implicit features and generating the rating as
ŝ u,t =b u +b t + ex ([ v u ;v t ])+<γ, im ([ G u ;G t ])>
where b u is a user bias, b t is an item bias, (⋅) is an explicit channel model, (⋅) is an implicit channel model, v u and v r are respective user- and review-based token-level features, G u is a matrix of a user attention representation of an aspect, and G t is a matrix of an item attention representation of the aspect; and
perform a response to the input text based on the estimated rating.
10. The system of claim 9 , wherein the computer program further causes the hardware processor to determine a set of sentiment terms using a plurality of different sources.
11. The system of claim 10 , wherein the plurality of different sources include a sentiment lexicon, a pointwise mutual information measurement of statistical word associations within a text corpus, and a coextraction model.
12. The system of claim 9 , wherein the computer program further causes the hardware processor to identify grammatical dependency relationships between words in the input text to generate candidate aspect-sentiment pairs.
13. The system of claim 12 , wherein the computer program further causes the hardware processor to filter the candidate aspect-sentiment pairs to exclude pairs that do not include an aspect from a predetermined set of sentiment terms.
14. The system of claim 9 , wherein explicit features characterize mentions of sentiment words and wherein implicit features characterize sentiment expressions that do not use sentiment words.
15. The system of claim 9 , wherein the computer program further causes the hardware processor to train the neural network using a training dataset that includes a set of written reviews and corresponding numerical ratings.
16. The system of claim 9 , wherein the computer program further causes the hardware processor to identify a product that a user would rate highly, based on the attention-property-aware rating, and recommending the product to the user.