# US 2020/0401852 A1 · Method and apparatus for generating information assessment model

> Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt

- Number: US 2020/0401852 A1
- Document ID: US20200401852A1
- Status: Published application
- Kind: A1
- Application number: 16/700,593
- Filed: 2019-12-02
- Published: 2020-12-24
- Earliest claimed: 2019-06-20
- Assignees: Baidu Online Network Technology (Beijing) Co., Ltd.
- Applicants: Baidu Online Network Technology (Beijing) Co., Ltd.
- Inventors: Miao Fan; Sen Ye; Chao Feng; Mingming Sun; Ping Li; Haifeng Wang
- Web page: https://patentgrep.com/patent/US20200401852A1

## CPC

- G06N 3/08: Learning methods
- G06F 18/211: Selection of the most significant subset of features
- G06F 18/2155: {characterised by the incorporation of unlabelled data, e.g. multiple instance learning [MIL], semi-supervised techniques using expectation-maximisation [EM] or naïve labelling}
- G06N 3/044: Recurrent networks, e.g. Hopfield networks
- G06N 3/0442: characterised by memory or gating, e.g. long short-term memory [LSTM] or gated recurrent units [GRU]
- G06N 3/045: Combinations of networks
- G06N 3/047: Probabilistic or stochastic networks
- G06N 3/0475: Generative networks
- G06N 3/084: Backpropagation, e.g. using gradient descent
- G06N 3/0895: Weakly supervised learning, e.g. semi-supervised or self-supervised learning
- G06N 3/09: Supervised learning
- G06N 3/092: Reinforcement learning
- G06N 3/094: Adversarial learning
- G06N 20/10: using kernel methods, e.g. support vector machines [SVM]
- G06N 20/20: Ensemble learning
- Y02P 90/30: Computing systems specially adapted for manufacturing
- G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06F: ELECTRIC DIGITAL DATA PROCESSING (computer systems based on specific computational models G06N)
- Y02P: CLIMATE CHANGE MITIGATION TECHNOLOGIES IN THE PRODUCTION OR PROCESSING OF GOODS

## Abstract

Embodiments of the present disclosure provide a method for generating an information assessment model, a method for determining the usefulness of comment information, apparatus, electronic device, and computer-readable medium. The method may include: acquiring training samples, the training samples including first sample comment information with a usefulness label and second sample comment information without a usefulness label; acquiring a predictor model and a discriminator model respectively constructed based on a generative network and a discrimination network in a generative adversarial network, and pre-training the predictor model using the first sample comment information, the predictor model being used to predict a usefulness label of a piece of comment information, the discriminator model being used to discriminate authenticity of a usefulness label; and training the predictor model and the discriminator model by iteratively performing a plurality of times of training operations, using the trained predictor model as an information assessment model.

## Claims

### Claim 1

A method for generating an information assessment model, comprising:
acquiring training samples, the training samples comprising first sample comment information with a usefulness label and second sample comment information without a usefulness label;
acquiring a predictor model and a discriminator model respectively constructed based on a generative network and a discrimination network in a generative adversarial network, and pre-training the predictor model using the first sample comment information, the predictor model being used to predict a usefulness label of a piece of comment information, the discriminator model being used to discriminate authenticity of a usefulness label; and
training the predictor model and the discriminator model by iteratively performing a plurality times of training operations, using the trained predictor model as an information assessment model;
wherein a training operation comprises:
predicting a usefulness label of the second sample comment information using the predictor model; using the usefulness label of the first sample comment information as a real label, using the usefulness label of the second sample comment information as a false label, discriminating authenticity of the usefulness label of the first sample comment information and authenticity of the usefulness label of the second sample comment information by using the discriminator model; iteratively updating, based on an error of the discrimination result of the discriminator model, a parameter of the discriminator model and a reward function, the reward function being associated with the parameter of the discriminator model and a parameter of the predictor model; determining a desired reward of the predictor model based on the reward function and the predictor model, and iteratively updating the parameter of the predictor model based on an error of the desired reward.

### Claim 2 (depends on 1)

The method according to claim 1, wherein the discriminating authenticity of the usefulness label of the first sample comment information and authenticity of the usefulness label of the second sample comment information by using the discriminator model comprises:
selecting usefulness labels of an equal number of first sample comment information and second sample comment information, and discriminating authenticity of the selected usefulness labels by using the discriminator model.

### Claim 3 (depends on 1)

The method according to claim 1, wherein the error of the discrimination result of the discriminator model comprises a sum of an opposite number of a first cross entropy and an opposite number of a second cross entropy;
the first cross entropy comprises: a cross entropy between a first probability distribution of that the usefulness label of the first sample comment information is a real label and a second probability distribution of that the discriminator model discriminates the usefulness label of the first sample comment information as a real label;
the second cross entropy comprises: a cross entropy between a third probability distribution of that the usefulness label of the second sample comment information predicted by the predictor model is a real label and a fourth probability distribution of that the discriminator model discriminates the usefulness label of the second sample comment information as a real label.

### Claim 4 (depends on 1)

The method according to claim 1, wherein when an input of the reward function is the second sample comment information, an output is a result of authenticity discrimination by the discriminator model for the usefulness label of the second sample comment information; and
the determining a desired reward of the predictor model based on the reward function and the predictor model, and iteratively updating the parameter of the predictor model based on an error of the desired reward comprises:
multiplying the reward function by a prediction result of the predictor model to calculate the desired reward, and iteratively updating, in response to determining that the desired reward does not reach a preset reward value condition, the parameter of the predictor model.

### Claim 5 (depends on 4)

The method according to claim 4, wherein
when the input of the reward function is the first sample comment information and the discriminator model discriminates the usefulness label of the first sample comment information as a real label, the value of the reward function is 1;
when the input of the reward function is the first sample comment information and the discriminator model discriminates the usefulness label of the first sample comment information as a false label, the value of the reward function is 0.

### Claim 6 (depends on 1)

The method according to claim 1, wherein the first sample comment information further has a usefulness index label, the usefulness index label is used to characterize a degree of usefulness, and the predictor model is further used to predict a usefulness index label of a piece of comment information;
the training operation further comprises:
predicting a usefulness index label of the second sample comment information using the predictor model; and
determining a prediction error of the predictor model for the usefulness index label; and
the iteratively updating the parameter of the predictor model based on an error of the desired reward comprises:
iteratively updating the parameter of the predictor model based on the error of the desired reward and the prediction error of the predictor model for the usefulness index label alternately.

### Claim 7 (depends on 1)

The method according to claim 1, wherein the method further comprises:
acquiring comment object information corresponding to the training samples; and
the predictor model comprises an information preprocessing network and a classification network;
wherein the information preprocessing network is used to convert a training sample and comment object information corresponding thereto into a mathematical representation, and the classification network predicts usefulness label of the training sample based on the mathematical representation.

### Claim 8 (depends on 7)

The method according to claim 7, wherein the information preprocessing network comprises a first bi-directional long short-term memory network and a second bi-directional long short-term memory network;
the first bi-directional long short-term memory network converts the training sample into a first vector representation, and the second bi-directional long short-term memory network converts the comment object information corresponding to the training sample into a second vector representation;
the information preprocessing network integrates, based on an attention mechanism, the first vector representation of the training sample and the second vector representation of the comment object information corresponding thereto into the mathematical representation corresponding to the training sample.

### Claim 9 (depends on 1)

A method for determining the usefulness of comment information, comprising:
acquiring a set of comment information; and
inputting the comment information in the set of comment information into the information assessment model generated by the method for generating an information assessment model according to claim 1, to obtain usefulness labels of the comment information.

### Claim 10 (depends on 9)

The method according to claim 9, wherein the method further comprises:
presenting, based on the usefulness labels of the comment information, the comment information in the set of comment information.

### Claim 11

An apparatus for generating an information assessment model, comprising:
at least one processor; and
a memory storing instructions, the instructions when executed by the at least one processor, cause the at least one processor to perform operations, the operations comprising:
acquiring training samples, the training samples comprising first sample comment information with a usefulness label and second sample comment information without a usefulness label;
acquiring a predictor model and a discriminator model respectively constructed based on a generation network and a discrimination network in a generative adversarial network, and pre-training the predictor model using the first sample comment information, the predictor model being used to predict a usefulness label of a piece of comment information, the discriminator model being used to discriminate authenticity of a usefulness label; and
training the predictor model and the discriminator model by iteratively performing a plurality times of training operations, using the trained predictor model as an information assessment model;
wherein a training operation comprises:
predicting a usefulness label of the second sample comment information using the predictor model; using the usefulness label of the first sample comment information as a real label, using the usefulness label of the second sample comment information as a false label, discriminating authenticity of the usefulness label of the first sample comment information and authenticity of the usefulness label of the second sample comment information by using the discriminator model; iteratively updating, based on an error of the discrimination result of the discriminator model, a parameter of the discriminator model and a reward function, the reward function being associated with the parameter of the discriminator model and a parameter of the predictor model; determining a desired reward of the predictor model based on the reward function and the predictor model, and iteratively updating the parameter of the predictor model based on an error of the desired reward.

### Claim 12 (depends on 11)

The apparatus according to claim 11, wherein the discriminating authenticity of the usefulness label of the first sample comment information and authenticity of the usefulness label of the second sample comment information by using the discriminator model comprises:
selecting usefulness labels of an equal number of first sample comment information and second sample comment information, and discriminating authenticity of the selected usefulness labels by using the discriminator model.

### Claim 13 (depends on 11)

The apparatus according to claim 11, wherein the error of the discrimination result of the discriminator model comprises a sum of an opposite number of a first cross entropy and an opposite number of a second cross entropy;
the first cross entropy comprises: a cross entropy between a first probability distribution of that the usefulness label of the first sample comment information is a real label and a second probability distribution of that the discriminator model discriminates the usefulness label of the first sample comment information as a real label;
the second cross entropy comprises: a cross entropy between a third probability distribution of that the usefulness label of the second sample comment information predicted by the predictor model is a real label and a fourth probability distribution of that the discriminator model discriminates the usefulness label of the second sample comment information as a real label.

### Claim 14 (depends on 11)

The apparatus according to claim 11, wherein when an input of the reward function is the second sample comment information, an output is a result of authenticity discrimination by the discriminator model for the usefulness label of the second sample comment information; and
the training unit is configured to determine the desired reward of the predictor model in the training operation, and iteratively update the parameter of the predictor model based on an error of the desired reward as follows:
multiplying the reward function by a prediction result of the predictor model to calculate the desired reward, and iteratively updating, in response to determining that the desired reward does not reach a preset reward value condition, the parameter of the predictor model.

### Claim 15 (depends on 14)

The apparatus according to claim 14, wherein
when the input of the reward function is the first sample comment information and the discriminator model discriminates the usefulness label of the first sample comment information as a real label, the value of the reward function is 1;
when the input of the reward function is the first sample comment information and the discriminator model discriminates the usefulness label of the first sample comment information as a false label, the value of the reward function is 0.

### Claim 16 (depends on 11)

The apparatus according to claim 11, wherein the first sample comment information further has a usefulness index label, the usefulness index label is used to characterize a degree of usefulness, and the predictor model is further used to predict a usefulness index label of a piece of comment information;
the training operation performed by the training unit further comprises:
predicting a usefulness index label of the second sample comment information using the predictor model;
determining a prediction error of the predictor model for the usefulness index label; and
the training unit is configured to iteratively update the parameter of the predictor model in the training operation as follows:
iteratively updating the parameter of the predictor model based on the error of the desired reward and the prediction error of the predictor model for the usefulness index label alternately.

### Claim 17 (depends on 11)

The apparatus according to claim 11, wherein the operations further comprises: acquiring comment object information corresponding to the training samples; and
the predictor model comprises an information preprocessing network and a classification network;
wherein the information preprocessing network is used to convert a training sample and comment object information corresponding thereto into a mathematical representation, and the classification network predicts usefulness label of the training sample based on the mathematical representation.

### Claim 18 (depends on 17)

The apparatus according to claim 17, wherein the information preprocessing network comprises a first bi-directional long short-term memory network and a second bi-directional long short-term memory network;
the first bi-directional long short-term memory network converts the training sample into a first vector representation, and the second bi-directional long short-term memory network converts the comment object information corresponding to the training sample into a second vector representation;
the information preprocessing network integrates, based on an attention mechanism, the first vector representation of the training sample and the second vector representation of the comment object information corresponding thereto into the mathematical representation corresponding to the training sample.

### Claim 19 (depends on 1)

An apparatus for determining the usefulness of comment information, comprising:
at least one processor; and
a memory storing instructions, the instructions when executed by the at least one processor, cause the at least one processor to perform operations, the operations comprising:
acquiring a set of comment information;
inputting the comment information in the set of comment information into the information assessment model generated by the method for generating an information assessment model according to claim 1, to obtain usefulness labels of the comment information; and
presenting, based on the usefulness labels of the comment information, the comment information in the set of comment information.

### Claim 20

A non-transitory computer-readable medium, storing a computer program thereon, wherein the program, when executed by a processor, causes the processor to implement the Operations, the operations comprise:
acquiring training samples, the training samples comprising first sample comment information with a usefulness label and second sample comment information without a usefulness label;
acquiring a predictor model and a discriminator model respectively constructed based on a generative network and a discrimination network in a generative adversarial network, and pre-training the predictor model using the first sample comment information, the predictor model being used to predict a usefulness label of a piece of comment information, the discriminator model being used to discriminate authenticity of a usefulness label; and
training the predictor model and the discriminator model by iteratively performing a plurality times of training operations, using the trained predictor model as an information assessment model;
wherein a training operation comprises:
predicting a usefulness label of the second sample comment information using the predictor model; using the usefulness label of the first sample comment information as a real label, using the usefulness label of the second sample comment information as a false label, discriminating authenticity of the usefulness label of the first sample comment information and authenticity of the usefulness label of the second sample comment information by using the discriminator model; iteratively updating, based on an error of the discrimination result of the discriminator model, a parameter of the discriminator model and a reward function, the reward function being associated with the parameter of the discriminator model and a parameter of the predictor model; determining a desired reward of the predictor model based on the reward function and the predictor model, and iteratively updating the parameter of the predictor model based on an error of the desired reward.