Searching method and system based on multi-round inputs, and terminal

Patent metadata

Field	Value
Publication number	US-10664755-B2
Application number	US-201715854481-A
Country	US
Kind code	B2
Filing date	Dec 26, 2017
Priority date	Nov 16, 2017
Publication date	May 26, 2020
Grant date	May 26, 2020

Abstract

A searching method and system based on multi-round inputs and a terminal are provided. The method comprises: acquiring search conditions input by a user in multiple searches; determining a multi-round property between at least two searches of the multiple searches; determining a search purpose of one of the search conditions, and determining that the search purpose of the one of the search conditions is a multi-round search purpose; generating search results based on the multi-round search purpose and search conditions input by the user; and ranking the generated search results, and determining and outputting an optimal search result. According to the searching method provided by the present application, a machine can understand a user's purpose under a continuous multi-round interactions by understanding the context, so that the use initiative of the user is improved.

First claim

What is claimed is:

1. A searching method based on multi-round inputs, comprising:
acquiring search conditions input by a user in multiple searches, wherein the multiple searches comprise at least two searches;
determining a multi-round property between at least two searches of the multiple searches; wherein the multi-round property represents a correlation of the search conditions of the at least two searches;
determining a search purpose of one of the search conditions, and determining that the search purpose of the one of the search conditions is a multi-round search purpose, wherein the multi-round search purpose indicates that the one of the search conditions is associated with the other search conditions input by the user;
generating search results based on the multi-round search purpose and search conditions input by the user, wherein the search results is generated by combining the search conditions; and
ranking the generated search results, and determining and presenting an optimal search result,
wherein the determining a search purpose of one of the search conditions and determining that the search purpose of the one of the search conditions is a multi-round search purpose comprises:
calculating a probability that the one of the search condition is a combined search condition based on a structured analysis model or a deep learning model; and
determining that the search purpose of the one of the search is the multi-round search purpose if the calculated probability is greater than a preset threshold.

2. The searching method according to claim 1 , wherein the determining a multi-round property between at least two searches of the multiple searches comprises:
smoothing the search conditions according to resources, requirements and a prior distribution;
calculating a bi-gram-based language model for the search condition based on the smoothed search conditions; and
determining the multi-round property between the at least two searches of the multiple searches according to the bi-gram-based language model.

3. The searching method according to claim 1 , wherein the deep learning model is an Long Short Term Memory (LSTM) model, and the probability that the one of the search conditions is the combined search condition is obtained by performing a classification training with the LSTM model; or
the probability that the search condition of this search is a combined search condition is calculated in the structured analysis model by a formula:

p
⁡

(
x
)

=

∑

i
=
1

n

⁢

⁢

(

x
i

⁢

w
i

⁢

λ
i

|
domain

,
ϕ

)

∑

i
=
1

n

⁢

⁢

x
i

⁢

w
i

,

wherein ϕ is a set of structured features, w is importance of a term in the one of the search condition, λ is a weight of a positive or negative feature, domain is a search purpose of the last search, and x is a set of terms in the one of the search condition; wherein a term is a semantic unit.

4. The searching method according to claim 1 , wherein ranking the generated search results, and determining and presenting an optimal search result comprises:
pruning the generated search results; and
ranking the search results retained after pruning, and determining and presenting the optimal search result.

5. The searching method according to claim 4 , wherein the pruning is performed based on following conditions that:
the one of the search conditions is only associated to a search condition of the last search;
the generated search results comprises newly added semantic data in this search; and
the generated search results comprises reference semantic expressions in this search and data corresponding to the reference semantic expressions.

6. The searching method according to claim 1 , wherein the ranking the generated search results comprises:
calculating occurrence probabilities of the respective generated search results based on a generative model or a discriminant model; and
ranking the generated search results according to the calculated occurrence probabilities.

7. The searching method according to claim 6 , wherein the discriminative model is a Gradient Boosted Decision Tree (GBDT) model and the occurrence probability of a search result is calculated by performing a discriminative training with the GBDT model; or
the occurrence probabilities of the respective generated search results are calculated with the generative model according to a probability calculation formula as follows:
P (candidate n )=η· P (slots n |o t+1 ,h t+1 ,a t ),

wherein in a case that the search conditions are independent of each other, a following calculation formula is obtained:

P
⁡

(

candidate
n

)

=

η
·

∏

i
=
0

⁢

P
⁡

(

Assignees

• Baidu online network technology beijing co ltd

Inventors

• Lu Guang
• Ju Qiang
• Luo Xiajun

Classifications

• G06F16/3346

  using probabilistic model · CPC title

• G06F16/951

  Indexing; Web crawling techniques · CPC title

• G06F16/3344

  using natural language analysis · CPC title

• G06F16/90332

  Natural language query formulation or dialogue systems · CPC title

• G06N20/00

  Machine learning · CPC title