Opinion snippet detection for aspect-based sentiment analysis

1. A computer-implemented method comprising:

receiving, by one or more processing units, at least one sentence and at least one aspect; determining, by one or more processing units, at least one opinion snippet of the at least one sentence based on the at least one sentence and the at least one aspect, through a trained aspect-sentence fusion model that outputs at least one word-aspect fusion result, wherein determining the at least one opinion snippet further comprises:

determining, by one or more processing units, a start position of the at least one opinion snippet in the at least one sentence based on the at least one word-aspect fusion result, and

determining, by one or more processing units, an end position of the at least one opinion snippet in the at least one sentence based on the determined start position by sampling the end position in a later part of the at least one sentence after the start position of the at least one sentence; and

calculating, by one or more processing units, a sentiment prediction of the at least one opinion snippet for the at least one aspect.

2. The method of claim 1 , wherein the at least one aspect is an aspect category or one or more aspect terms.

3. The method of claim 1 , wherein the aspect-sentence fusion model is bidirectional encoder representations from transformers (BERT).

4. The method of claim 1 , wherein determining the start position comprises:

computing, by one or more processing units, a multinomial distribution as a dot product between the at least one word-aspect fusion result and a trained start vector followed by a softmax over all the words of the at least one sentence; and

sampling, by one or more processing units, the start position based on the multinomial distribution.

5. The method of claim 1 , wherein calculating the sentiment prediction of the at least one opinion snippet comprises:

obtaining, by one or more processing units, an average representation of the at least one opinion snippet; and

calculating, by one or more processing units, the sentiment prediction based on the average representation.

6. The method of claim 4 , wherein the trained start vector is learned, by one or more processing units, through a reinforcement learning based on training data with sentences, aspects, and corresponding labeled sentiment classifications.

7. A computer system, comprising:

a processor;

a computer-readable memory coupled to the processor, the memory comprising instructions that when executed by the processor perform actions of:

receiving at least one sentence and at least one aspect;

determining at least one opinion snippet of the at least one sentence based on the at least one sentence and the at least one aspect, through a trained aspect-sentence fusion model that outputs at least one word-aspect fusion result, wherein determining the at least one opinion snippet comprises:

determining a start position of the at least one opinion snippet in the at least one sentence based on the at least one word-aspect fusion result, and

determining an end position of the at least one opinion snippet in the at least one sentence based on the determined start position by sampling the end position in a later part of the at least one sentence after the start position of the at least one sentence; and

calculating a sentiment prediction of the at least one opinion snippet for the at least one aspect.

8. The computer system of claim 7 , wherein the at least one aspect is an aspect category or one or more aspect terms.

9. The computer system of claim 7 , wherein the trained aspect-sentence fusion model is bidirectional encoder representations from transformers (BERT).

10. The computer system of claim 7 , wherein determining the start position comprises:

computing a multinomial distribution as a dot product between the at least one word-aspect fusion result and a trained start vector followed by a softmax over all the words of the at least one sentence; and

sampling the start position based on the multinomial distribution.

11. The computer system of claim 7 , wherein calculating the sentiment prediction of the at least one opinion snippet comprises:

obtaining an average representation of the at least one opinion snippet; and

calculating the sentiment prediction based on the average representation.

12. The computer system of claim 10 , wherein the trained start vector is learned through a reinforcement learning based on training data with sentences, aspects, and corresponding labeled sentiment classifications.

13. A computer program product for aspect-based sentiment analysis, comprising a computer readable storage medium having program instructions embodied therewith, the program instructions executable by a processor to cause the processor to:

receive at least one sentence and at least one aspect;

determine at least one opinion snippet of the at least one sentence based on the at least one sentence, the at least one aspect and a trained aspect-sentence fusion model that outputs at least one word-aspect fusion result, wherein determine the at least one opinion snippet comprises:

determine a start position of the at least one opinion snippet in the at least one sentence based on the at least one word-aspect fusion result, and

determine an end position of the at least one opinion snippet in the at least one sentence based on the determined start position by sampling the end position in a later part of the at least one sentence after the start position of the at least one sentence; and

calculate a sentiment prediction of the at least one opinion snippet for the at least one aspect.

14. The computer program product of claim 13 , wherein the trained aspect-sentence fusion model is bidirectional encoder representations from transformers (BERT).