Frequently asked questions and document retrieval using bidirectional encoder representations from transformers (BERT) model trained on generated paraphrases

1. A system, comprising a processor to:

receive a query;

retrieve ranked candidates from an index based on the query;

fine-tune a generative pretrained transformer trained on question-answer pairs of a concatenated frequently asked question (FAQ) dataset using randomly sampled sequences of the concatenated FAQ dataset;

automatically generate paraphrases for questions in the FAQ dataset via the fine- tuned generative pretrained transformer based on input answers from the FAQ dataset;

filter the automatically generated paraphrases to match a same FAQ as their generation questions using the index;

fine-tune a Bidirectional Encoder Representations from Transformers (BERT) query-question (Q-q) model using the filtered generated paraphrases to match queries to questions of the FAQ dataset;

re-rank the ranked candidates using the fine-tuned BERT Q-q model; and

return the re-ranked candidates in response to the query.

2. The system of claim 1 , wherein the Q-q BERT model is trained using triplets comprising a question, a positive paraphrase, and a negative paraphrase.

3. The system of claim 1 , wherein the processor is to re- rank the ranked candidates using a final re-ranking of the candidates by combining a plurality of re-rankers using an unsupervised late-fusion, wherein the plurality of re-rankers comprise the fine-tuned BERT Q-q model, a BERT query-answer (Q-a) model, and a passage-based re-ranker.

4. The system of claim 3 , wherein the unsupervised late- fusion comprises summing candidate scores assigned for each candidate by the fine-tuned BERT Q-q model, a BERT query-answer (Q-a) model, and a passage-based re-ranker.

5. The system of claim 4 , wherein the unsupervised late-fusion comprises applying an unsupervised query expansion step for re-ranking a candidate pool of the summed candidate scores.

6. A computer-implemented method, comprising:

receiving, via a processor, a query;

retrieving, via the processor, ranked candidates from an index based on the query;

fine-tuning, via the processor, a generative pretrained transformer trained on question-answer pairs of a concatenated frequently asked question (FAQ) dataset using randomly sampled sequences of the concatenated FAQ dataset;

automatically generating, via the processor, question paraphrases via the fine-tuned generative pretrained transformer based on input answers from the FAQ dataset;

filtering, via the processor, the automatically generated question paraphrases to match a same FAQ as their generation questions using the index by running the question paraphrases against the index of the FAQ dataset;

fine-tuning, via the processor, a Bidirectional Encoder Representations from Transformers (BERT) query-question (Q-q) model based on the filtered generated question paraphrases to match queries to questions of the FAQ dataset;

re-ranking, via the processor, the ranked candidates using the fine-tuned BERT Q- q model; and

returning, via the processor, the re-ranked candidates in response to the query.

7. The computer-implemented method of claim 6 , wherein fine-tuning the generative pretrained transformer using randomly sampled sequences of the concatenated FAQ dataset with special tokens.

8. The computer-implemented method of claim 6 , wherein generating the question paraphrases comprises using only the FAQ dataset as training input.

9. The computer-implemented method of claim 6 , wherein the question paraphrases comprise title paraphrases, the question-answer pairs comprise title-abstract pairs, and wherein the BERT Q-t model is fined-tuned based on filtered title paraphrases.

10. A computer program product for ranking query candidates, the computer program product comprising a computer-readable storage medium having program code embodied therewith, wherein the computer-readable storage medium is not a transitory signal per se, the program code executable by a processor to cause the processor to:

receive a query;

retrieve ranked candidates from an index based on the query;

fine-tune a generative pretrained transformer trained on question-answer pairs of a concatenated frequently asked question (FAQ) dataset using randomly sampled sequences of the concatenated FAQ dataset;

automatically generate question paraphrases for questions in the FAQ dataset via the fine-tuned generative transformer based on input answers from the FAQ dataset;

filter the automatically generated question paraphrases to match a same FAQ as their generation questions using the index by running the question paraphrases against the index;

fine-tune a Bidirectional Encoder Representations from Transformers (BERT) Query-question (Q-q) model based on the filtered generated paraphrases to match queries to questions of the FAQ dataset;

re-rank the ranked candidates using the fine-tuned BERT Q-q model; and

return the re-ranked candidates in response to the query.

11. The computer program product of claim 10 , further comprising program code executable by the processor to fine-tune the generative pretrained transformer using the randomly sampled sequences of the concatenated FAQ dataset with special tokens.

12. The computer program product of claim 10 , further comprising program code executable by the processor to generate the question paraphrases using only the FAQ dataset as training input.

13. The computer program product of claim 10 , wherein the question paraphrases comprise title paraphrases, the question-answer pairs comprise title-abstract pairs, and wherein fine-tuning the BERT Q-q model is based on filtered title paraphrases.

14. The system of claim 1 , wherein the processor is to further input a special token with each answer into the fine-tuned generative pretrained transformer, wherein the fine-tuned generative pretrained transformer generates tokens until the special token is reached and all the generated tokens are used as an automatically generated question paraphrase for the answer.