Method and system for inferring answers from knowledge graphs
Automated response generation systems and methods are disclosed. The systems can include a deep learning model specially configured to apply inferencing techniques to redesign natural language querying systems for use over knowledge graphs. The disclosed systems and methods provide a model for inferencing referred to as a Hierarchical Recurrent Path Encoder (HRPE). An entity extraction and linking module as well as a data conversion and generation module process the content of a given query. The output is processed by the proposed model to generate inferred answers.
1. A computer implemented method of using a Hierarchical Recurrent Path Encoder (HRPE) model to control a virtual agent by generating responses to queries received by the virtual agent, the method comprising:
receiving a natural language query;
generating hypothesis based on the received query;
receiving, at an attention component of the HRPE model, the hypothesis;
encoding, by the HRPE model, a first path in a knowledge graph to generate a first encoding, the first path connecting to a first possible answer to the query;
collectively encoding, by the HRPE model, a first plurality of paths in the knowledge graph to generate a second encoding, the first plurality of paths including the first path;
generating, by the HRPE model, attention weights for each of the first encoding and the second encoding to produce weighted encodings;
selecting, by the attention component and based on the hypothesis, an encoding from the weighted encodings based on which encoding is associated with a greater attention weight, the selected encoding corresponding to a first vector;
determining, via a feed forward network layer of the HRPE model, whether the first vector follows the hypothesis or not;
classifying the first vector based on output of the feed forward network layer, the classification indicating whether the selected encoding entails or contradicts the hypothesis; and
generating, via an inference-based response generation module associated with the HRPE model, a natural language response to the query based on the classification.
2. The computer implemented method of claim 1 , further comprising encoding, by the HRPE model, a second path in the knowledge graph to generate a third encoding, the second path connecting to a second possible answer to the query, wherein the first plurality of paths includes the second path.
3. The computer implemented method of claim 1 , wherein the query is converted into a hypothesis in part by removing interrogatory-type words from the query.
4. The computer implemented method of claim 1 , further comprising:
encoding the hypothesis using a bidirectional-long short-term memory neural network (bi-LSTM) to produce a second vector; and
using the second vector to generate the attention weights.
5. The computer implemented method of claim 4 , further comprising:
encoding the first encoding and the second encoding by a long short-term memory (LSTM) neural network to produce a third vector; and
concatenating the first vector with both the second vector and the third vector to produce a concatenated vector.
6. The computer implemented method of claim 5 , further comprising processing the concatenated vector through the feed forward network layer in order to determine whether the first vector follows the hypothesis or not.
7. The computer implemented method of claim 6 , further comprising processing the output of the feed forward network layer through a SoftMax classifier in order to perform the classification.
8. A non-transitory computer-readable medium storing software comprising instructions executable by one or more computers which, upon such execution, cause the one or more computers to use a Hierarchical Recurrent Path Encoder (HRPE) model to control a virtual agent by generating responses to queries received by the virtual agent by performing the following operations:
receive a natural language query;
generate hypothesis based on the received query;
receive, at an attention component of the HRPE model, the hypothesis;
encode, by the HRPE model, a first path in a knowledge graph to generate a first encoding, the first path connecting to a first possible answer to the query;
collectively encode, by the HRPE model, a first plurality of paths in the knowledge graph to generate a second encoding, the first plurality of paths including the first path;
generate, by the HRPE model, attention weights for each of the first encoding and the second encoding to produce weighted encodings;
select, by the attention component and based on the hypothesis, an encoding from the weighted encodings based on which encoding is associated with a greater attention weight, the selected encoding corresponding to a first vector;
determine, via a feed forward network layer of the HRPE model, whether the first vector follows the hypothesis or not;
classify the first vector based on output of the feed forward network layer, the classification indicating whether the selected encoding entails or contradicts the hypothesis; and
generate, via an inference-based response generation module associated with the HRPE model, a natural language response to the query based on the classification.
9. The non-transitory computer-readable medium storing software of claim 8 , wherein the instructions further cause the one or more computers to encode, by the HRPE model, a second path in the knowledge graph to generate a third encoding, the second path connecting to a second possible answer to the query, wherein the first plurality of paths includes the second path.
10. The non-transitory computer-readable medium storing software of claim 8 , wherein the query is converted into a hypothesis in part by removing interrogatory-type words from the query.
11. The non-transitory computer-readable medium storing software of claim 8 , wherein the instructions further cause the one or more computers to:
encode the hypothesis using a bidirectional-long short-term memory neural network (bi-LSTM) to produce a second vector; and
use the second vector to generate the attention weights.
12. The non-transitory computer-readable medium storing software of claim 11 , wherein the instructions further cause the one or more computers to:
encode the first encoding and the second encoding by a long short-term memory (LSTM) neural network to produce a third vector; and
concatenate the first vector with both the second vector and the third vector to produce a concatenated vector.
13. The non-transitory computer-readable medium storing software of claim 12 , wherein the instructions further cause the one or more computers to process the concatenated vector through the feed forward network layer in order to determine whether the first vector follows the hypothesis or not.
14. The non-transitory computer-readable medium storing software of claim 13 , wherein the instructions further cause the one or more computers to process the output of the feed forward network layer through a SoftMax classifier in order to perform the classification.
15. A system for using a Hierarchical Recurrent Path Encoder (HRPE) model to control a virtual agent by generating responses to queries received by the virtual agent, the system comprising:
one or more computers and one or more storage devices storing instructions that are operable, when executed by the one or more computers, to cause the one or more computers to:
receive a natural language query;
generate hypothesis based on the received query;
receive, at an attention component of the HRPE model, the hypothesis;
encode, by the HRPE model, a first path in a knowledge graph to generate a first encoding, the first path connecting to a first possible answer to the query;
collectively encode, by the HRPE model, a first plurality of paths in the knowledge graph to generate a second encoding, the first plurality of paths including the first path;
generate, by the HRPE model, attention weights for each of the first encoding and the second encoding to produce weighted encodings;
select, by the attention component and based on the hypothesis, an encoding from the weighted encodings based on which encoding is associated with a greater attention weight, the selected encoding corresponding to a first vector;
determine, via a feed forward network layer of the HRPE model, whether the first vector follows the hypothesis or not;
classify the first vector based on output of the feed forward network layer, the classification indicating whether the selected encoding entails or contradicts the hypothesis; and
generate, via an inference-based response generation module associated with the HRPE model, a natural language response to the query based on the classification.
16. The system of claim 15 , wherein the instructions further cause the one or more computers to encode, by the HRPE model, a second path in the knowledge graph to generate a third encoding, the second path connecting to a second possible answer to the query, wherein the first plurality of paths includes the second path.
17. The system of claim 15 , wherein the query is converted into a hypothesis in part by removing interrogatory-type words from the query.
18. The system of claim 15 , wherein the instructions further cause the one or more computers to:
encode the hypothesis using a bidirectional-long short-term memory neural network (bi-LSTM) to produce a second vector; and
use the second vector to generate the attention weights.
19. The system of claim 18 , wherein the instructions further cause the one or more computers to:
encode the first encoding and the second encoding by a long short-term memory (LSTM) neural network to produce a third vector; and
concatenate the first vector with both the second vector and the third vector to produce a concatenated vector.
20. The system of claim 19 , wherein the instructions further cause the one or more computers to process the concatenated vector through the feed forward network layer in order to determine whether the first vector follows the hypothesis or not.