Systems and methods for mapping a predicted entity to a product based on an online query

1. A system comprising:

one or more processors; and

one or more non-transitory storage devices storing computing instructions configured to run on the one or more processors and perform:

receiving an online query from an electronic device of a user, the online query comprising natural language of the user;

obtaining a first set of rules (1) comprising a neural network having a plurality of hidden layers, (2) defining an intent of the online query as at least one of a product-related intent or a non-product-related intent, and (3) operating as a function of at least:

(i) a word vector;

(ii) a total number of the plurality of hidden layers;

(iii) a respective hidden dimension vector for each respective hidden layer of the plurality of hidden layers; and

(iv) one or more training parameter matrices inserted into a sigmoid function;

determining the intent of the online query based on the first set of rules;

in response to determining that the intent is the product-related intent:

obtaining a second set of rules that define an entity prediction from the online query;

predicting an entity from the online query based on the second set of rules, the entity comprising one or more of a product name, a product attribute, a product price range, an average customer review, or a product brand;

mapping the entity, as predicted from the online query, to product metadata associated with one or more products; and

coordinating displaying product information of the one or more products on the electronic device of the user; and

in response to determining that the intent is the non-product-related intent:

determining, using the first set of rules, that the non-product-related intent comprises a request to file a complaint; and

coordinating displaying a message on the electronic device of the user informing the user that a user support assistant has been notified of the request to file the complaint, wherein the first set of rules further comprises:

converting the online query into a sequence of word vectors {x t } T , where (a) x is the word vector, (b) t is an index of the plurality of hidden layers, (c) T is the total number of the plurality of hidden layers, and (d) each respective hidden layer of the plurality of hidden layers comprises:

 (i) a respective hidden dimension vector {h t−1 } T input for the respective hidden dimension vector;

 (ii) at least one respective corresponding word vector {x t } T input; and

 (iii) a respective output {h t } T for a next respective hidden layer of the plurality of hidden layers, where h is a hidden dimension; and

applying a formula h t =σ(h t−1 H+x t W), where H and W are respective training parameter matrices of the one or more training parameter matrices used to learn from training data, and σ is the sigmoid function.

2. The system of claim 1 , wherein:

the one or more non-transitory storage devices storing the computing instructions are further configured to run on the one or more processors and perform, after determining that the intent is the product-related intent:

determining a physical store that is closest to the electronic device of the user; and

determining whether the one or more products are available for purchase at the physical store that is closest to the electronic device of the user; and

the product information comprises at least one of a price for the one or more products at the physical store or a location of the one or more products at the physical store.

3. The system of claim 1 , wherein:

the one or more non-transitory storage devices storing the computing instructions are further configured to run on the one or more processors and perform, after determining that the intent is the product-related intent:

obtaining a third set of rules that define a personalization of the one or more products for display on the electronic device of the user based on at least one of an online order history of the user, an in-store purchase history of the user, or an online browsing history of the user; and

mapping the entity, as predicted from the online query, to the product metadata associated with the one or more products comprises mapping the entity, as predicted from the online query, to the product metadata associated with the one or more products based on the third set of rules.

4. The system of claim 1 , wherein:

the one or more non-transitory storage devices storing the computing instructions are further configured to run on the one or more processors and perform, after determining the intent is the product-related intent:

determining a physical store that is closest to the electronic device of the user;

determining whether the one or more products are available for purchase in the physical store that is closest to the electronic device of the user; and

obtaining a third set of rules that define a personalization of the one or more products for display on the electronic device of the user based on at least one of an online order history of the user, an in-store purchase history of the user, or an online browsing history of the user;

mapping the entity, as predicted from the online query, to the product metadata associated with the one or more products comprises mapping the entity, as predicted from the online query, to the product metadata associated with the one or more products based on the third set of rules; and

the product information comprises at least one of a price for the one or more products at the physical store or a location of the one or more products at the physical store.

5. The system of claim 1 , wherein predicting the entity from the online query based on the second set of rules comprises:

predicting a plurality of entities and a respective confidence level for each respective entity of the plurality of entities from the online query based on the second set of rules.

6. The system of claim 1 , wherein:

the sigmoid function comprises

σ

⁡

(

z

)

=

1

1

+

e

-

z

,

where z is a dummy variable and e is an exponential function; and

the first set of rules further comprises:

inserting the output from a last hidden layer h T into a softmax function to obtain a probability over a plurality of intents as ŷ=s(h T U+b), where ŷ is a final output after softmax transformation, s is the softmax function

s

⁡

(

z

i

)

=

e

z

i

∑

j

=

1

n

⁢

∑

j

⁢

e

z

j

,

n is a total number of the plurality of intents, i and j are respective indexes for summation, and matrices W∈R d×h , H∈R h×h , U∈R h×n and b∈R 1×h are model parameters that are determined by model training, where R is defined as a real number set;

manually labeling each respective training example with a respective intent of the plurality of intents;

predicting a probability for each position in y to be 1 with the softmax function, where y is a one-hop vector that represents a true label of the online query, only a position corresponding to the respective intent, as manually labeled, takes value 1, and a remainder of positions in the one-hop vector take value 0;

optimizing the model parameters by minimizing cross-entropy loss using an equation comprising:

-

1

N

⁢

∑

k

=

1

N

⁢

∑

i

=

1

n

⁢

y

i

(

k

)

⁢

log

⁡

(

y

^

i

(

k

)

)

where N is a number of training examples that effectively minimizes an error of model output ŷ compared with ground-truth label y;

using a stochastic gradient descent method for batch parameter optimization; and

calculating ŷ using optimized parameters and a highest probability intent of the plurality of intents, the highest probability intent having a highest probability given by ŷ as a final predicted intent to make a prediction after training.

7. The system of claim 1 , wherein the second set of rules comprises:

determining context of a word in the online query by representing the context as a window comprising the word concatenated with additional words neighboring the word on both sides as x (t) =[x t−k , . . . , x t , . . . , x t+ ], where x is the word vector, t is an index of the plurality of hidden layers, {x i } i=t−k t+k is a word vector representation of word x t and the additional words, and k is a word window size;

predicting the entity of the word x t ;

calculating a hidden layer h=tan h (x (t) W+b 1 ) to get a respective probability of each respective entity of a plurality of entities by using an equation comprising ŷ=s(hU+b 2 ), where tan h is a hyperbolic tangent function, W is one or more first matrix variables for training, b 1 and b 2 are respective intercept variables for training, ŷ is an output from a logistic function, h is an output from a tan h function, s is a softmax function, and a ground-truth label of the entity for the word is represented as a one-hop vector y; and

optimizing parameters W∈R (2k+1)×h , b 1 ∈R h , U∈R h×n , b 2 ∈R n by minimizing a cross-entropy loss using a stochastic gradient descent method and a batch optimization, where U is one or more second matrix variables for training, R is defined as a real number set, and h and n are dimensions of a matrix.

8. The system of claim 1 , wherein:

the one or more non-transitory storage devices storing the computing instructions are further configured to run on the one or more processors and perform, after determining the intent is the product-related intent:

determining a physical store that is closest to the electronic device of the user;

determining whether the one or more products are available for purchase at the physical store that is closest to the electronic device of the user; and

obtaining a third set of rules that define a personalization of the one or more products for display on the electronic device of the user based on at least one of an online order history of the user, an in-store purchase history of the user, or an online browsing history of the user;

mapping the entity, as predicted from the online query, to the product metadata associated with the one or more products comprises mapping the entity, as predicted from the online query, to the product metadata associated with the one or more products based on the third set of rules;

the product information comprises at least one of a price for the one or more products at the physical store or a location of the one or more products at the physical store based on the third set of rules;

predicting the entity from the online query based on the second set of rules comprises predicting a plurality of entities and a respective confidence level for each respective entity of the plurality of entities from the online query based on the second set of rules;

the sigmoid function comprises

σ

⁡

(

z

)

=

1

1

+

e

-

z

,

where z is a dummy variable and e is an exponential function;

the first set of rules further comprises:

inserting the output from a last hidden layer h T into a softmax function to obtain a probability over a plurality of intents as ŷ=s(h T U+b), where ŷ is a final output after softmax transformation, s is the softmax function

s

⁡

(

z

i

)

=

e

z

i

∑

j

=

1

n

⁢

∑

j

⁢

e

z

j

,

n is a total number of the plurality of intents, i and j are respective indexes for summation, and matrices W∈R d×h , H∈R h×h , U∈R h×n and b∈R 1×h are model parameters that are determined by model training, where R is defined as a real number set;

manually labeling each respective training example with a respective intent of the plurality of intents;

predicting a probability for each position in y to be 1 with the softmax function, where y is a one-hop vector that represents a true label of the online query, only a position corresponding to the respective intent, as manually labeled, takes value 1, and a remainder of positions in the one-hop vector take value 0;

optimizing the model parameters by minimizing cross-entropy loss using an equation comprising:

-

1

N

⁢

∑

k

=

1

N

⁢

∑

i

=

1

n

⁢

y

i

(

k

)

⁢

log

⁡

(

y

^

i

(

k

)

)

where N is a number of training examples that effectively minimizes an error of model output ŷ compared with ground-truth label y;

using a stochastic gradient descent method for batch parameter optimization; and

calculating ŷ using optimized parameters and a highest probability intent of the plurality of intents, the highest probability intent having a highest probability given by ŷ as a final predicted intent to make a prediction after training; and

the second set of rules comprises:

determining context of a word in the online query by representing the context as a window comprising the word concatenated with additional words neighboring the word on both sides as x (t) =[x t−k , . . . , x t , . . . , x t+k ], where x is the word vector, t is an index of the plurality of hidden layers, {x i } i=t−k t+k is a word vector representation of word x t and the additional words, and k is a word window size;

predicting the entity of the word x t ;

calculating a hidden layer h=tan h(x (t) W+b 1 ) to get a respective probability of each respective entity of a plurality of entities by using an equation comprising ŷ=s(hU+b 2 ), where tan h is a hyperbolic tangent function, W is one or more first matrix variables for training, b 1 and b 2 are respective intercept variables for training, ŷ is an output from a logistic function, h is an output from a tan h function, s is the softmax function, and a ground-truth label of the entity for the word is represented as a one-hop vector y; and

optimizing parameters W∈R (2k+1)×h , b 1 ∈R h , U∈R h×n , b 2 ∈R n by minimizing a cross-entropy loss using a stochastic gradient descent method and a batch optimization, where U is one or more second matrix variables for training, R is defined as a real number set, and h and n are dimensions of a matrix.

9. A method comprising:

receiving an online query from an electronic device of a user, the online query comprising natural language of the user;

obtaining a first set of rules (1) comprising a neural network having a plurality of hidden layers, (2) defining an intent of the online query as at least one of a product-related intent or a non-product-related intent, and (3) operating as a function of at least:

(i) a word vector;

(ii) a total number of the plurality of hidden layers;

(iii) a respective hidden dimension vector for each respective hidden layer of the plurality of hidden layers; and

(iv) one or more training parameter matrices inserted into a sigmoid function;

determining the intent of the online query based on the first set of rules;

in response to determining that the intent is the product-related intent:

obtaining a second set of rules that define an entity prediction from the online query;

predicting an entity from the online query based on the second set of rules, the entity comprising one or more of a product name, a product attribute, a product price range, an average customer review, or a product brand;

mapping the entity, as predicted from the online query, to product metadata associated with one or more products; and

coordinating displaying product information of the one or more products on the electronic device of the user; and

in response to determining that the intent is the non-product-related intent:

determining, using the first set of rules, that the non-product-related intent comprises a request to file a complaint; and

coordinating displaying a message on the electronic device of the user informing the user that a user support assistant has been notified of the request to file the complaint, wherein the first set of rules further comprises:

converting the online query into a sequence of word vectors {x t } T , where (a) x is the word vector, (b) t is an index of the plurality of hidden layers, (c) T is the total number of the plurality of hidden layers, and (d) each respective hidden layer of the plurality of hidden layers comprises:

(i) a respective hidden dimension vector {h t−1 } T input for the respective hidden dimension vector;

(ii) at least one respective corresponding word vector {x t } T input; and

(iii) a respective output {h t } T for a next respective hidden layer of the plurality of hidden layers, where h is a hidden dimension; and

applying a formula h t =σ(h t−1 H+x t W), where H and W are respective training parameter matrices of the one or more training parameter matrices used to learn from training data, and σ is the sigmoid function.

10. The method of claim 9 , further comprising, after determining that the intent is the product-related intent:

determining a physical store that is closest to the electronic device of the user; and

determining whether the one or more products are available for purchase at the physical store that is closest to the electronic device of the user, wherein the product information comprises at least one of a price for the one or more products at the physical store or a location of the one or more products at the physical store.

11. The method of claim 9 , wherein:

the method further comprises, after determining that the intent is the product-related intent:

obtaining a third set of rules that define a personalization of the one or more products for display on the electronic device of the user based on at least one of an online order history of the user, an in-store purchase history of the user, or an online browsing history of the user; and

mapping the entity, as predicted from the online query, to the product metadata associated with the one or more products comprises mapping the entity, as predicted from the online query, to the product metadata associated with the one or more products based on the third set of rules.

12. The method of claim 9 , wherein:

the method further comprises, after determining that the intent is the product-related intent:

determining a physical store that is closest to the electronic device of the user when the online query is received from the electronic device of the user;

determining whether the one or more products are available for purchase in the physical store that is closest to the electronic device of the user; and

obtaining a third set of rules that define a personalization of the one or more products for display on the electronic device of the user based on at least one of an online order history of the user, an in-store purchase history of the user, or an online browsing history of the user;

mapping the entity, as predicted from the online query, to the product metadata associated with the one or more products comprises mapping the entity, as predicted from the online query, to the product metadata associated with the one or more products based on the third set of rules; and

the product information comprises at least one of a price for the one or more products at the physical store or a location of the one or more products at the physical store.

13. The method of claim 9 , wherein predicting the entity from the online query based on the second set of rules comprises:

predicting a plurality of entities and a respective confidence level for each respective entity of the plurality of entities from the online query based on the second set of rules.

14. The method of claim 9 , wherein:

the sigmoid function comprises

σ

⁡

(

z

)

=

1

1

+

e

-

z

,

where z is a dummy variable and e is an exponential function; and

the first set of rules further comprises:

inserting the output from a last hidden layer h T into a softmax function to obtain a probability over a plurality of intents as ŷ=s(h T U+b), where ŷ is a final output after softmax transformation, s is the softmax function

s

⁡

(

z

i

)

=

e

z

i

∑

j

=

1

n

⁢

∑

j

⁢

e

z

j

,

n is a total number of the plurality of intents, i and j are respective indexes for summation, and matrices W∈R d×h , H∈R h×n , U∈R h×n and b∈R 1×h are model parameters that are determined by model training, where R is defined as a real number set;

manually labeling each respective training example with a respective intent of the plurality of intents;

predicting a probability for each position in y to be 1 with the softmax function, where y is a one-hop vector that represents a true label of the online query, only a position corresponding to the respective intent, as manually labeled, takes value 1, and a remainder of positions in the one-hop vector take value 0;

optimizing the model parameters by minimizing cross-entropy loss using an equation comprising:

-

1

N

⁢

∑

k

=

1

N

⁢

∑

i

=

1

n

⁢

y

i

(

k

)

⁢

log

⁡

(

y

^

i

(

k

)

)

where N is a number of training examples that effectively minimizes an error of model output ŷ compared with ground-truth label y;

using a stochastic gradient descent method for batch parameter optimization; and

calculating ŷ using optimized parameters and a highest probability intent of the plurality of intents, the highest probability intent having a highest probability given by ŷ as a final predicted intent to make a prediction after training.

15. The method of claim 9 , wherein the second set of rules comprises:

determining context of a word in the online query by representing the context as a window comprising the word concatenated with additional words neighboring the word on both sides as x (t) =[x t−k , . . . , x t , . . . , x t+k ], where x is the word vector, t is an index of the plurality of hidden layers, {x i } i=t−k t+k is a word vector representation of word x t and the additional words, and k is a word window size;

predicting the entity of the word x t ;

calculating a hidden layer h=tan h (x (t) W+b 1 ) to get a respective probability of each respective entity of a plurality of entities by using an equation comprising ŷ=s(hU+b 2 ), where tan h is a hyperbolic tangent function, W is one or more first matrix variables for training, b 1 and b 2 are respective intercept variables for training, ŷ is an output from a logistic function, h is an output from a tan h function, s is a softmax function, and a ground-truth label of the entity for the word is represented as a one-hop vector y; and

optimizing parameters W∈R (2k+1)×h , b 1 ∈R h , U∈R h×n , b 2 ∈R n by minimizing a cross-entropy loss using a stochastic gradient descent method and a batch optimization, where U is one or more second matrix variables for training, R is defined as a real number set, and h and n are dimensions of a matrix.

16. The method of claim 9 , wherein:

the method further comprises:

determining a physical store that is closest to the electronic device of the user;

determining whether the one or more products are available for purchase at the physical store that is closest to the electronic device of the user; and

obtaining a third set of rules that define a personalization of the one or more products for display on the electronic device of the user based on at least one of an online order history of the user, an in-store purchase history of the user, or an online browsing history of the user;

mapping the entity, as predicted from the online query, to the product metadata associated with the one or more products comprises mapping the entity, as predicted from the online query, to the product metadata associated with the one or more products based on the third set of rules;

the product information comprises at least one of a price for the one or more products at the physical store or a location of the one or more products at the physical store based on the third set of rules;

predicting the entity from the online query based on the second set of rules comprises predicting a plurality of entities and a respective confidence level for each respective entity of the plurality of entities from the online query based on the second set of rules;

the sigmoid function comprises

σ

⁡

(

z

)

=

1

1

+

e

-

z

,

where z is a dummy variable and e is an exponential function;

the first set of rules further comprises:

inserting the output from a last hidden layer h T into a softmax function to obtain a probability over a plurality of intents as ŷ=s(h T U+b), where ŷ is a final output after softmax transformation, s is the softmax function

s

⁡

(

z

i

)

=

e

z

i

∑

j

=

1

n

⁢

∑

j

⁢

e

z

j

,

n is a total number of the plurality of intents, i and j are respective indexes for summation, and matrices W∈R d×h , H∈R h×h , U∈R h×n and b∈R 1×h are model parameters that are determined by model training, where R is defined as a real number set;

manually labeling each respective training example with a respective intent of the plurality of intents;

predicting a probability for each position in y to be 1 with the softmax function, where y is a one-hop vector that represents a true label of the online query, only a position corresponding to the respective intent, as manually labeled, takes value 1, and a remainder of positions in the one-hop vector take value 0;

optimizing the model parameters by minimizing cross-entropy loss using an equation comprising:

-

1

N

⁢

∑

k

=

1

N

⁢

∑

i

=

1

n

⁢

y

i

(

k

)

⁢

log

⁡

(

y

^

i

(

k

)

)

where N is a number of training examples that effectively minimizes an error of model output ŷ compared with ground-truth label y;

using a stochastic gradient descent method for batch parameter optimization; and

calculating ŷ using optimized parameters and a highest probability intent of the plurality of intents, the highest probability intent having a highest probability given by ŷ as a final predicted intent to make a prediction after training; and

the second set of rules comprises:

determining context of a word in the online query by representing the context as a window comprising the word concatenated with additional words neighboring the word on both sides as x (t) =[x t−k , . . . , x t , . . . , x t+k ], where x is the word vector, t is an index of the plurality of hidden layers, {x i } i=t−k t+k is a word vector representation of word x t and the additional words, and k is a word window size;

predicting the entity of the word x t ;

calculating a hidden layer h=tan h (x (t) W+b 1 ) to get a respective probability of each respective entity of a plurality of entities by using an equation comprising ŷ=s(hU+b 2 ), where tan h is a hyperbolic tangent function, W is one or more first matrix variables for training, b 1 and b 2 are respective intercept variables for training, ŷ is an output from a logistic function, h is an output from a tan h function, s is the softmax function, and a ground-truth label of the entity for the word is represented as a one-hop vector y; and

optimizing parameters W∈R (2k+1)×h , b 1 ∈R h , U∈R h×n , b 2 ∈R n by minimizing a cross-entropy loss using a stochastic gradient descent method and a batch optimization, where U is one or more second matrix variables for training, R is defined as a real number set, and h and n are dimensions of a matrix.

17. A system comprising:

one or more processors; and

one or more non-transitory storage devices storing computing instructions configured to run on the one or more processors and perform:

receiving an online query from an electronic device of a user, the online query comprising natural language of the user;

determining an intent of the online query as at least one of a product-related intent or a non-product-related intent using (1) a neural network having a plurality of hidden layers and (2) at least:

(i) a word vector;

(ii) a total number of the plurality of hidden layers;

(iii) a respective hidden dimension vector for each respective hidden layer of the plurality of hidden layers; and

(iv) one or more training parameter matrices inserted into a sigmoid function;

in response to determining the intent is the product-related intent:

predicting an entity from the online query, the entity comprising one or more of a product name, a product attribute, a product price range, an average customer review, or a product brand;

mapping the entity, as predicted from the online query, to product metadata associated with one or more products; and

coordinating displaying product information of the one or more products on the electronic device of the user;

in response to determining that the intent is the non-product-related intent:

determining, using a first set of rules, that the non-product-related intent comprises a request to file a complaint; and

coordinating displaying a message on the electronic device of the user informing the user that a user support assistant has been notified of the request to file the complaint, wherein the first set of rules further comprises:

converting the online query into a sequence of word vectors {x t } T , where (a) x is the word vector, (b) t is an index of the plurality of hidden layers, (c) T is the total number of the plurality of hidden layers, and (d) each respective hidden layer of the plurality of hidden layers comprises:

 (i) a respective hidden dimension vector {h t−1 } T input for the respective hidden dimension vector;

 (ii) at least one respective corresponding word vector {x t } T input; and

 (iii) a respective output {h t } T for a next respective hidden layer of the plurality of hidden layers, where h is a hidden dimension; and

applying a formula h t =σ(h t−1 H+x t W), where H and W are respective training parameter matrices of the one or more training parameter matrices used to learn from training data, and σ is the sigmoid function.

18. The system of claim 17 , wherein:

the one or more non-transitory storage devices storing the computing instructions are further configured to run on the one or more processors and perform, after determining that the intent is the product-related intent:

determining a physical store that is closest to the electronic device of the user; and

determining whether the one or more products are available for purchase at the physical store that is closest to the electronic device of the user; and

the product information comprises at least one of a price for the one or more products at the physical store or a location of the one or more products at the physical store.

19. The system of claim 17 , wherein mapping the entity, as predicted from the online query, to the product metadata associated with the one or more products comprises:

mapping the entity, as predicted from the online query, to the product metadata associated with the one or more products based on at least one of an online order history of the user, an in-store purchase history of the user, or a browsing history of the user.

20. The system of claim 17 , wherein:

the one or more non-transitory storage devices storing the computing instructions are further configured to run on the one or more processors and perform, after determining that the intent is the product-related intent:

determining a physical store that is closest to the electronic device of the user;

determining whether the one or more products are available for purchase in the physical store that is closest to the electronic device of the user; and

determining a personalization of the one or more products to the user based on at least one of an online order history of the user, an in-store purchase history of the user, or a browsing history of the user;

mapping the entity, as predicted from the online query, to the product metadata associated with the one or more products comprises mapping the entity, as predicted from the online query, to the product metadata associated with the one or more products based on the personalization of the one or more products; and

the product information comprises at least one of a price for the one or more products at the physical store or a location of the one or more products at the physical store based on the personalization of the one or more products.