FAQ about NLP AutoML-Natural Language Processing(NLP)-阿里云帮助中心

This topic answers frequently asked questions about NLP AutoML.

1. NLP AutoML service invocation limits

Algorithm type	QPS per model	Maximum request size (bytes)
Named entity recognition	20	No limit
Text classification	20	No limit
Key phrase extraction	20	No limit
Short text matching	20	No limit
Relation extraction	20	No limit
Resume extraction	20	No limit
Product review analysis	20	No limit
Sentiment analysis	20	No limit

2. Length limits

Q: Is there a length limit for the text in named entity recognition predictions?
A: There is no limit. However, very long text can cause a timeout. We recommend that you keep the text under 1,000 characters. You can split longer text using line feeds ("\n") or periods ("."). Alternatively, you can use our asynchronous prediction API, which supports predictions for up to 10,000 characters. For more information, see Asynchronous prediction API example.
Q: Does splitting long text for named entity recognition predictions affect model accuracy?
A: No, it does not. Splitting text using line feeds ("\n") or periods (".") does not affect the model's prediction accuracy. We recommend that you split long samples before you make predictions.

3. RAM user authorization

3.1 Authorizing a RAM user

1. In the Resource Access Management (RAM) console, create a custom policy on the Policies page.

2. Enter the policy script.

NLP AutoML policy script

{
    "Statement": [
        {
            "Effect": "Allow",
            "Action": "nlp-automl:*",
            "Resource": "*"
        }
    ],
    "Version": "1"
}

3. On the **Users** page, grant permissions to the desired RAM user.

4. Select the custom policy that you created to complete the authorization.

3.2 RAM user permissions

An authorized RAM user has almost all platform management permissions, except for permissions related to purchases and viewing.

Platform resource	Permission	Authorized RAM user has permission
Projects	All	Yes
Model	All	Yes
Data	All	Yes
Monitoring and statistics	All	Yes
Purchase-related	All	No

3.3 Other notes

Currently, only the root account has permissions to perform sales-related operations and view sales data.
After an authorized RAM user logs in, they are redirected to the My Projects page in the console. An unauthorized RAM user is redirected to the Annotation Center.
The administrator of a data annotation project is the creator of the project or a user who is granted administrator permissions for the project.
The Alibaba Cloud account has editing permissions for data annotation tasks and can add RAM users to participate in data annotation.
The quality inspector for a data annotation project is the administrator of that project.
By default, the creator ID for all projects and models is the ID of the Alibaba Cloud account.
When the predict API is called, the caller ID defaults to the Alibaba Cloud account.
The logs that are pushed to Log Service (SLS) by the predict API include information about whether the caller is an Alibaba Cloud account or a RAM user.
Billing is calculated based on the Alibaba Cloud account.

If your average daily call volume exceeds 500,000 over the past week, contact your sales manager to increase your queries per second (QPS).

4. FAQ about named entity recognition projects

4.1 Data annotation notes

In a named entity recognition project, you must annotate all occurrences of an entity in the document. We recommend that you enable the "Annotate Same Value" feature in the upper-left corner.
Q: If an entity has been annotated in one document, does it need to be annotated again in another document?
A: Yes, it does.
Annotated entities cannot contain line feeds ("\n") or periods ("."). If an annotated entity contains these characters, it will not be recognized. However, this does not affect model training.

4.2 Advanced rules engine settings

Q: For a named entity recognition task, my document format is relatively fixed, and the fields that I need to extract have clear and consistent contexts. Is there an efficient extraction method that does not require extensive data annotation?

A: Yes, there is. The platform provides a rules engine module to improve the performance and efficiency of named entity recognition models. If you need this feature, contact us to have your account added to the whitelist. You can join our DingTalk group (ID: 44619071) for assistance. You can configure the rules engine during model creation. The rules engine interface is shown in the following figure:

You can configure rules that combine regular expressions, dictionaries, wildcards, and entities extracted by the model. You can use the rule preview feature to test the effectiveness of your rules.

Example of rule configuration:

I need to extract the name, gender, and date of birth of the plaintiff and the defendant from the following judgment document:

Plaintiff: Cheng Xiaoer, female, residing in Yuhang District, Hangzhou, Zhejiang Province.\n\nDefendant: Wang Moumou, male, born on October 1, 2019, Han ethnicity.

You can use the rules engine to configure regular expressions to extract these fields:

Click Add Rule Group, set Entity Type to Plaintiff, select Regular Expression as the rule element, and then enter `(?<=Plaintiff:)([^,]+)(?=,)` as the value.

After you configure the rules, click Rule Preview to check the results:

4.3 Common fields for named entity recognition projects

The platform can extract common fields, such as dates and mobile phone numbers. If you select these fields during model creation, the trained model can also extract them. Note: Model creation requires annotated data. To use this feature, you must annotate a minimum of 10 data entries to train a model. Make sure that you select the required common fields in the advanced settings.

Path: Named Entity Recognition Project > Model Hub > Create Model > Advanced Settings

5. FAQ about purchase and renewal

5.1 "Number of models exceeds the quantity in the new specification" error during renewal

On the renewal page, if you receive a message indicating that the number of models exceeds the quantity in the new specification, go to the My Models page. On this page, check your total model quota and the number of models in use. Make sure that the number of models in use is less than or equal to your quota.

If the number of models in use is greater than your model quota, you can perform one of the following operations. For example, this issue may occur if you used three models during a trial but purchased a quota for only one model for your paid subscription without deleting the other two models.

Delete the models that you no longer need. (Note: Deleted models cannot be recovered.)
Click Increase Quantity and upgrade your specification to increase your total model quota.

5.2 Is there a limit on the number of data annotation tasks for model training on the self-learning platform?

A: A single model can have an unlimited number of data annotation tasks.

5.3 Can I purchase a resource plan directly?

A: No, you cannot. Resource plans cover only the fees for model invocations. To train models, you must also purchase a monthly model subscription.

5.4 After my NLP AutoML plan is used up, I am charged automatically. Can I disable this?

NLP AutoML is a pay-as-you-go product. When you activate NLP AutoML, you also activate the pay-as-you-go billing method for service invocations.
Resource plans offer a lower price per invocation. The fees for invocations are deducted from your resource plan. Any usage that exceeds the plan is billed on a pay-as-you-go basis.
You cannot disable only the pay-as-you-go billing. To disable it, you must disable the NLP AutoML product.
For more information about billing, see Product Pricing.

6. General FAQ

6.1 Are the AccessKey ID and AccessKey secret universal across Alibaba Cloud services?

The AccessKey ID and AccessKey secret are valid for all Alibaba Cloud services. You must activate a service before you can use your credentials to access it.

6.2 Do I need to purchase multiple models for multiple datasets?

A: No, you do not. The number of datasets is independent of the number of models. You need to purchase multiple models only if you want to train multiple models at the same time.

6.3 How to increase QPS

If your average daily call volume exceeds 500,000 over the past week, you can request a QPS increase. To do so, join the Alibaba Cloud NLP AutoML User Q&A Group 2 (DingTalk group ID: 44619071) for assistance.