Search-engine-augmented dialogue response generation with cheaply supervised query production
Method, apparatus, and non-transitory storage medium for neural network based dialogue generation, including receiving an input dialogue context, and generating queries based on the input dialogue context using a query generating neural network. The query generating neural network may be trained using a cheap noisy supervision function. The method may further include retrieving responses from a web-based search engine based on the generated queries, and generating dialogue based on the retrieved responses and the input dialogue context.
1. A method for neural network based dialogue generation, the method being performed by at least one processor, the method comprising:
receiving an input dialogue context;
determining at least one entity referenced by a pronoun in the input dialogue context;
generating content focused input dialogue context based on discarding function words from the input dialogue context;
generating queries based on the input dialogue context using a query generating neural network, wherein the query generating neural network is trained using a cheap noisy supervision function;
retrieving responses from a web-based search engine based on the generated queries; and
generating dialogue based on the retrieved responses and the input dialogue context.
2. The method of claim 1 , wherein the cheap noisy supervision function comprises:
comparing a plurality of articles associated with a query among one or more queries with a gold response associated with the query among the one or more queries; and
generating a score for the query among the one or more queries based on the comparing, wherein the score indicates a quality associated with the query.
3. The method of claim 2 , wherein generating the score for the one or more queries comprises:
generating an article score for each of the plurality of articles associated with the query among the one or more queries; and
determining the score for the one or more queries based on a maximum article score among the article score for each of the plurality of articles associated with the query among the one or more queries.
4. The method of claim 1 , wherein training the query generating neural network using the cheap noisy supervision function comprises:
determining a query among one or more queries with a highest score based on the cheap noisy supervision function;
training the query generating neural network based on the determined query; and
fine-tuning the query generating neural network based on a reward-based algorithm and a loss function.
5. The method of claim 4 , wherein the loss function is based on a minimum score and the highest score associated with the query among the one or more queries.
6. The method of claim 1 , wherein the generating the queries based on the input dialogue context comprises:
generating contextualized embeddings based on a pre-trained language model; and
determining a probability distribution associated with the input dialogue context based on the generated contextualized embeddings.
7. The method of claim 1 , wherein the generated queries comprise one or more text spans based on the input dialogue context.
8. An apparatus for neural network based dialogue generation, the apparatus comprising:
at least one memory configured to store program code; and
at least one processor configured to read the program code and operate as instructed by the program code, the program code comprising:
receiving code configured to cause the at least one processor to receive an input dialogue context;
first determining code configured to cause the at least one processor to determine at least one entity referenced by a pronoun in the input dialogue context;
first generating code configured to cause the at least one processor to generate content focused input dialogue context based on discarding function words from the input dialogue context;
second generating code configured to cause the at least one processor to generate queries based on the input dialogue context using a query generating neural network, wherein the query generating neural network is trained using a cheap noisy supervision function;
retrieving code configured to cause the at least one processor to retrieve responses from a web-based search engine based on the generated queries; and
third generating code configured to cause the at least one processor to generate dialogue based on the retrieved responses and the input dialogue context.
9. The apparatus of claim 8 , wherein the cheap noisy supervision function comprises:
comparing code configured to cause the at least one processor to compare a plurality of articles associated with a query among one or more queries with a gold response associated with the query among the one or more queries; and
fourth generating code configured to cause the at least one processor to generate a score for the query among the one or more queries based on the comparing, wherein the score indicates a quality associated with the query.
10. The apparatus of claim 9 , wherein the fourth generating code comprises:
fifth generating code configured to cause the at least one processor to generate an article score for each of the plurality of articles associated with the query among the one or more queries; and
second determining code configured to cause the at least one processor to determine the score for the one or more queries based on a maximum article score among the article score for each of the plurality of articles associated with the query among the one or more queries.
11. The apparatus of claim 8 , wherein training code is configured to cause the at least one processor to train the query generating neural network using the cheap noisy supervision function, and wherein the training code comprises:
fourth determining code configured to cause the at least one processor to determine a query among one or more queries with a highest score based on the cheap noisy supervision function;
query model training code configured to cause the at least one processor to train the query generating neural network based on the determined query; and
fine-tuning code configured to cause the at least one processor to fine-tune the query generating neural network based on a reward-based algorithm and a loss function.
12. The apparatus of claim 11 , wherein the loss function is based on a minimum score and the highest score associated with the query among the one or more queries.
13. The apparatus of claim 8 , wherein the second generating code comprises:
seventh generating code configured to cause the at least one processor to generate contextualized embeddings based on a pre-trained language model; and
fourth determining code configured to cause the at least one processor to determine a probability distribution associated with the input dialogue context based on the generated contextualized embeddings.
14. A non-transitory computer readable medium storing instructions that, when executed by at least one processor for neural network based dialogue generation, cause the at least one processor to:
receive an input dialogue context;
determine at least one entity referenced by a pronoun in the input dialogue context;
generate content focused input dialogue context based on discarding function words from the input dialogue context;
generate queries based on the input dialogue context using a query generating neural network, wherein the query generating neural network is trained using a cheap noisy supervision function;
retrieve responses from a web-based search engine based on the generated queries; and
generate dialogue based on the retrieved responses and the input dialogue context.
15. The non-transitory computer readable medium of claim 14 , wherein training the query generating neural network using the cheap noisy supervision function comprises:
determining a query among one or more queries with a highest score based on the cheap noisy supervision function;
training the query generating neural network based on the determined query; and
fine-tuning the query generating neural network based on a reward-based algorithm and a loss function.
16. The non-transitory computer readable medium of claim 15 , wherein the loss function is based on a minimum score and the highest score associated with the query among the one or more queries.
17. The non-transitory computer readable medium of claim 14 , wherein generating the queries based on the input dialogue context comprises:
generating contextualized embeddings based on a pre-trained language model; and
determining a probability distribution associated with the input dialogue context based on the generated contextualized embeddings.
18. The non-transitory computer readable medium of claim 14 , wherein the generated queries comprise one or more text spans based on the input dialogue context.