Method and device for evaluating comment quality, and computer readable storage medium

Patent metadata

Field	Value
Publication number	US-11756094-B2
Application number	US-202016803943-A
Country	US
Kind code	B2
Filing date	Feb 27, 2020
Priority date	Mar 1, 2019
Publication date	Sep 12, 2023
Grant date	Sep 12, 2023

Abstract

Embodiments of the present disclosure provide a method and a device for evaluating a comment quality, an electronic device, and a computer readable storage medium. The method includes: selecting a metadata key associated with a comment of an object from metadata of the object, the metadata including a plurality of key-value pairs; determining a value corresponding to the metadata key based on the metadata; and evaluating the comment quality based on the comment and the value corresponding to the metadata key.

First claim

What is claimed is:

1. A method for evaluating an object comment, performed by an electronic device, comprising:
receiving an object comment in the form of audio, video, or picture, and converting the object comment into a text;
selecting by a selector model, a target metadata key associated with the text from metadata of the object, wherein the metadata of the object includes a plurality of key-value pairs and metadata keys in the plurality of key-value pairs are hardware parameters of the object;
determining by the selector model, a target value corresponding to the target metadata key by querying the metadata;
determining by a predictor model, a score of the text based on the text and the target value corresponding to the target metadata key; and
displaying on top an object comment corresponding to a text with the maximum score in a comment display interface of an e-commerce platform;
wherein the selector model and the predictor model are simultaneously trained with a training set that comprises a plurality of texts and annotated data indicating a score of each text, by acts of:
initializing parameters of each of the selector model and the predictor model;
inputting a first text in the training set and first metadata corresponding to the first text into the selector model to select a first metadata key;
obtaining a bi-linear relationship between the first metadata key and a context of the first text, and selecting a first value corresponding to the first metadata key based on a policy of p=Softmax(Reduce_max(B,axis=1), where B indicates the bi-linear relationship;
inputting the first text and the first value corresponding to the first metadata key into the predictor model to generate a first prediction result;
optimizing the predictor model by a stochastic gradient descent method based on a predicted loss between the first prediction result and first annotated data of the first text;
determining whether to give the selector model a reward based on a performance of the predictor model, wherein the reward is within a range of (0.0, 1.0] based on a loss function that predicts a true score by taking the context as features; and
optimizing the selector model by a policy gradient method using the reward in response to determining to give the selector model the reward, wherein the policy gradient method is configured to update policy parameters by continuously calculating a gradient of the expected total reward regarding the policy parameters and converge to an optimal policy.

2. The method according to claim 1 , wherein said selecting the metadata key comprises:
generating a first vector of the text by converting each word or phrase in the text into a word vector;
generating a second vector of the metadata by converting each metadata key in the metadata to a predetermined dimension vector;
determining a relevance of each metadata key in the metadata to the text based on the first vector and the second vector; and
selecting a metadata key having a maximum relevance to the text from the metadata.

3. The method according to claim 2 , wherein said determining the score of the text comprises:
generating a third vector of the value corresponding to the metadata key; and
determining the score of the text based on the third vector and the first vector.

4. The method according to claim 3 , wherein said determining the score of the text comprises:
determining that the text of the object comment has a high score in response to determining that the score is greater than a first threshold; and
determining that the text of the object comment has a low score in response to determining that the score is less than a second threshold, wherein the first threshold is greater than or equal to the second threshold.

5. The method according to claim 1 , wherein the annotated data comprises voting data of the object comment fetched from network, and the voting data is provided by a plurality of network users.

6. A device for evaluating an object comment, comprising:
one or more processors; and
a storage device, configured to store one or more programs,
wherein when the one or more programs are executed by the one or more processors, the one or more processors are configured to:
receive an object comment in the form of audio, video, or picture, and converting the object comment into a text;
select a target metadata key associated with the text from metadata of the object, wherein the metadata of the object includes a plurality of key-value pairs and metadata keys in the plurality of key-value pairs are hardware parameters of the object;
determine a target value corresponding to the target metadata key by querying the metadata; and
determine a score of the text based on the text and the target value corresponding to the target metadata key; and
display on top an object comment corresponding to a text with the maximum score in a comment display interface of an e-commerce platform;
wherein the selector model and the predictor model are simultaneously trained with a training set that comprises a plurality of texts and annotated data indicating a quality of each text, by acts of:
initializing parameters of each of the selector model and the predictor model;
inputting a first text in the training set and first metadata corresponding to the first text into the selector model to select a first metadata key;
obtaining a bi-linear relationship between the first metadata key and a context of the first text, and selecting a first value corresponding to the first metadata key based on a policy of p=Softmax(Reduce_max(B,axis=1), where B indicates the bi-linear relationship;
inputting the first text and the first value corresponding to the first metadata key into the predictor model to generate a first prediction result;
optimizing the predictor model by a stochastic gradient descent method based on a predicted loss between the first prediction result and first annotated data of the first text;
determining whether to give the selector model a reward based on a performance of the predictor model, wherein the reward is within a range of (0.0, 1.0] based on a loss function that predicts a true score by taking the context as features; and
optimizing the selector model by a policy gradient method using the reward in response to determining to give the selector model the reward, wherein the policy gradient method is configured to update policy parameters by continuously calculating a gradient of the expected total reward regarding the policy parameters and converge to an optimal policy.

7. The device according to claim 6 , wherein the one or more processors are further configured to:
generate a first vector of the text by converting each word or phrase in the text into a word vector;
generate a second vector of the metadata by converting each metadata key in the metadata to a predetermined dimension vector;
determine a relevance of each metadata key in the metadata to the text based on the first vector and the second vector; and
select a metadata key having a maximum relevance to the text from the metadata.

8. The device according to claim 7 , wherein the one or more processors are further configured to:
generate a third vector of the value corresponding to the metadata key; and
determine the score of the text based on the third vector and the first vector.

9. The device according to claim 8 , wherein the one or more processors are further configured to:
determine that the text of the object comment has a high score in response to determining that the score is greater than a first threshold; and
determine that the text of the object comment has a low score in response to determining that the score is less than a second thresh

Assignees

• Baidu online network technology beijing co ltd

Inventors

• Fan Miao
• Sun Mingming
• Li Ping
• Wang Haifeng

Classifications

• G06N3/092

  Reinforcement learning · CPC title

• G06N3/0442

  characterised by memory or gating, e.g. long short-term memory [LSTM] or gated recurrent units [GRU] · CPC title

• G06N3/09

  Supervised learning · CPC title

• G06Q30/0625 Primary

  by formulating product or service queries, e.g. using keywords or predefined options · CPC title

• G06F40/30

  Semantic analysis · CPC title