Response retrieval system and method
A method of obtaining a response to a query inputted by a user, the method comprising: receiving a user inputted query; encoding said query to produce a context vector; retrieving responses with associated response vectors; scoring response vectors in the database against the context vector wherein the scoring is a measure of the similarity between the context vector and a response vector; and outputting the responses with the closest response vectors, wherein encoding said query to produce a context vector comprises using a pre-trained model, wherein said pre-trained model has been trained using corresponding queries and responses such that an encoding is used that maximises the similarity between the response vector and context vector for a corresponding query and response.
1. A method of obtaining a response to a query inputted by a user, the method comprising:
obtaining a training dataset comprising queries and corresponding responses;
generating, using the training data set, a pre-trained model that comprises, for respective ones of the queries and corresponding responses of the dataset, a context vector for the query and a response vector for the response, wherein the response vector corresponds to the context vector for the query;
receiving a user inputted query;
encoding said query to produce a context vector;
identifying, using the pre-trained model, responses having response vectors associated with the context vector;
scoring the response vectors of the responses against the context vector to generate similarity scores for the response vectors, wherein the similarity score for a response vector and a context vector is a measure of a similarity between the context vector and the response vector;
identifying, based on the similarity scores for the response vectors, responses that correspond to the user inputted query;
outputting the responses that correspond to the user inputted query;
receiving a further query;
encoding said further query to produce a further context vector;
producing a weighted context vector from a weighted sum of the further context vector from the further query and the context vector from a previous query, wherein the further context vector of the further query has a higher weighting than the context vector of the previous query;
identifying, using the pre-trained model, further response vectors associated with the weighted context vector;
scoring the further response vectors against the weighted context vector, wherein the scoring of a further response vector against the weighted context vector is a measure of a similarity between the weighted context vector and the further response vector; and
identifying, based on the scoring of the further response vectors against the weighted context vector, further responses that correspond to the further query;
outputting the further responses that correspond to the further query.
2. A method according to claim 1 , wherein the responses are selected from phrases and images.
3. A method according to claim 1 , wherein the scoring of the response vectors comprises a scaled cosine similarity scoring.
4. A method according to claim 1 , wherein the outputted responses comprise at least one image and at least one phrase.
5. A method according to claim 1 , further comprising passing the context vector through a classifier, wherein the responses are identified based on the output of the classifier.
6. A method according to claim 5 , wherein the classifier is responsive to a request to re-start a query.
7. A method according to claim 6 , wherein the classifier is responsive to a user requiring a service.
8. A method according to claim 1 , wherein said user query relates to information about a plurality of entities and said responses relate to said plurality of entities, wherein the method is configured to output the responses from the highest ranked M entities, and wherein M is an integer of 2 or more.
9. A method according to claim 8 , further comprising identifying the highest ranked M entities by:
taking the closest N responses, where N is an integer of at least 3; and
determining a total score for each entity by determining a value representing the score of each response and adding these values for each entity.
10. A method according to claim 9 , wherein the method is adapted to receive a further query, said method further comprising:
encoding said further query to produce a further context vector;
identifying, using the pre-trained model, further response vectors associated with the further context vector, wherein the further response vectors are response vectors relating to the highest ranked M entities of a previous query; and
identifying further responses associated with the further response vectors; and
outputting the further responses.
11. A method according to claim 1 , further comprising:
subdividing said response vectors into clusters in a response vector space; and
determining a cluster of the response vectors associated with the context vector,
wherein the responses that correspond to the user inputted query are associated with the response vectors of the cluster.
12. A method according to claim 1 , wherein the user input is a speech input and said speech input is converted to text.
13. A method according to claim 1 , wherein the responses are output using speech synthesis.
14. A method according to claim 1 , wherein the training dataset comprises comment and response pairs obtained by way of a social media platform.
15. A system for obtaining a response to a query inputted by a user, the system comprising an input interface, a processor and an output, the system being adapted to:
obtain a training dataset comprising queries and corresponding responses;
generate, using the training data set, a pre-trained model that comprises, for respective ones of the queries and responses of the dataset, a context vector for the query and a response vector for the response, wherein the response vector corresponds to the context vector for the query;
receive a user inputted query using the input interface;
encode said query to produce a context vector;
identify, using the pre-trained model, responses having response vectors associated with the context vectors;
score the response vectors of the responses against the context vector to generate similarity scores for the response vectors, wherein the similarity score for a response vector and a context vector is a measure of a similarity between the context vector and the response vector;
identify, based on the similarity scores for the response vectors, responses that correspond to the user inputted query; output the responses that correspond to the user inputted query;
receive a further query;
encode said further query to produce a further context vector;
produce a weighted context vector from a weighted sum of the further context vector from the further query and the context vector from a previous query, wherein the further context vector of the further query has a higher weighting than the context vector of the previous query;
identify, using the pre-trained model, further response vectors associated with the weighted context vector;
score the further response vectors against the weighted context vector, wherein the scoring of a further response vector against the weighted context vector is a measure of a similarity between the weighted context vector and the further response vector; and
identify, based on the scoring of the further response vectors against the weighted context vector, further responses that correspond to the further query;
output the further responses that correspond to the further query.
16. A non-transitory carrier medium comprising computer readable code configured to cause a computer to perform the following:
obtaining a training dataset comprising queries and corresponding responses;
generating, using the training data set, a pre-trained model that comprises, for respective ones of the queries and corresponding responses of the dataset, a context vector for the query and a response vector for the response, wherein the response vector corresponds to the context vector for the query;
receiving a user inputted query;
encoding said query to produce a context vector;
identifying, using the pre-trained model, responses having response vectors associated with the context vector;
scoring the response vectors of the responses against the context vector to generate similarity scores for the response vectors, wherein the similarity score for a response vector and a context vector is a measure of a similarity between the context vector and the response vector;
identifying, based on the similarity scores for the response vectors, responses that correspond to the user inputted query;
outputting the responses that correspond to the user inputted query;
receiving a further query;
encoding said further query to produce a further context vector;
producing a weighted context vector from a weighted sum of the further context vector from the further query and the context vector from a previous query, wherein the further context vector of the further query has a higher weighting than the context vector of the previous query;
identifying, using the pre-trained model, further response vectors associated with the weighted context vector;
scoring the further response vectors against the weighted context vector, wherein the scoring of a further response vector against the weighted context vector is a measure of a similarity between the weighted context vector and the further response vector; and
identifying, based on the scoring of the further response vectors against the weighted context vector, further responses that correspond to the further query;
outputting the further responses that correspond to the further query.