Identifying search results using deep query understanding
An improved search engine is disclosed. The search engine receives search queries from client devices and inputs these queries into a first neural network (an action understanding model) that includes an action embedding layer. The action embedding layer can be a word embedding layer constructed using action terms. The action understanding model outputs a filter match associated with a type of filter and, in some scenarios, an action-condition pair. The action-condition pair includes an action associated with the type of filter and a condition comprising an adaptive value associated with the action. Based on the filter and, if present, action-condition pair(s), the embodiments generate a structured query and issue the structured query to a data repository (e.g., database). The search engine then returns a search results page responsive to the search query that includes the results returned by the data repository in response to the structured query.
1. A method comprising:
inputting a search query from a client into an action understanding model, the action understanding model including an action embedding layer, the action embedding layer comprising a word embedding layer constructed using a plurality of action terms;
receiving, from the action understanding model, a filter match, the filter match including a type of filter and an action-condition pair, the action-condition pair comprising an action associated with the type of filter and a condition comprising an adaptive value associated with the action;
generating a structured query based on the filter;
issuing the structured query to a data repository; and
receiving search results from the data repository in response to the structured query.
2. The method of claim 1 , wherein the action understanding model includes a plurality of layers, wherein the word embedding layer comprises an initial layer of the plurality of layers.
3. The method of claim 2 , wherein the plurality of layers includes one or more hidden layers, the hidden layers comprising bidirectional long short-term memory (LSTM) layers.
4. The method of claim 3 , wherein at least one of the hidden layers comprises a dense layer having a corresponding dropout hyperparameter.
5. The method of claim 1 , further comprising: generating a label for the search query using a first neural network; generating a named entity in the search query using a second neural network; and including outputs of the first neural network and second neural network in the structured query.
6. The method of claim 5 , wherein the first neural network includes an ELMo embedding layer and the second neural network includes a Global Vectors for Word Representation (GloVe) embedding layer.
7. A non-transitory computer-readable storage medium for tangibly storing computer program instructions capable of being executed by a computer processor, the computer program instructions defining steps of:
inputting a search query from a client into an action understanding model, the action understanding model including an action embedding layer, the action embedding layer comprising word embedding layer constructed using a plurality of action terms;
receiving, from the action understanding model, a filter match, the filter match including a type of filter and an action-condition pair, the action-condition pair comprising an action associated with the type of filter and a condition comprising an adaptive value associated with the action;
generating a structured query based on the filter;
issuing the structured query to a data repository; and
receiving search results from the data repository in response to the structured query.
8. The non-transitory computer-readable storage medium of claim 7 , wherein the action understanding model includes a plurality of layers, wherein the word embedding layer comprises an initial layer of the plurality of layers.
9. The non-transitory computer-readable storage medium of claim 8 , wherein the plurality of layers includes one or more hidden layers, the hidden layers comprising bidirectional long short-term memory (LSTM) layers.
10. The non-transitory computer-readable storage medium of claim 9 , wherein at least one of the hidden layers comprises a dense layer having a corresponding dropout hyperparameter.
11. The non-transitory computer-readable storage medium of claim 7 , the steps further comprising: generating a label for the search query using a first neural network; generating a named entity in the search query using a second neural network; and including outputs of the first neural network and second neural network in the structured query.
12. The non-transitory computer-readable storage medium of claim 11 , wherein the first neural network includes an ELMo embedding layer and the second neural network includes a Global Vectors for Word Representation (GloVe) embedding layer.
13. A device comprising:
a processor configured to:
input a search query from a client into an action understanding model, the action understanding model including an action embedding layer, the action embedding layer comprising word embedding layer constructed using a plurality of action terms;
receive, from the action understanding model, a filter match, the filter match including a type of filter and an action-condition pair, the action-condition pair comprising an action associated with the type of filter and a condition comprising an adaptive value associated with the action;
generate a structured query based on the filter;
issue the structured query to a data repository; and
receive search results from the data repository in response to the structured query.
14. The device of claim 13 , wherein the action understanding model includes a plurality of layers, wherein the word embedding layer comprises an initial layer of the plurality of layers.
15. The device of claim 14 , wherein the plurality of layers includes one or more hidden layers, the hidden layers comprising bidirectional long short-term memory (LSTM) layers.
16. The device of claim 15 , wherein at least one of the hidden layers comprises a dense layer having a corresponding dropout hyperparameter.
17. The device of claim 13 , the processor further configured to: generate a label for the search query using a first neural network; generate a named entity in the search query using a second neural network; and include outputs of the first neural network and second neural network in the structured query.