When utilizing AI, training data must be prepared for learning. In many cases, annotation work is necessary to prepare this training data.
But what exactly is annotation work? Furthermore, through what specific processes is it carried out?
This article explains the content, necessity, and specific work processes of annotation.
|
【Table of Contents】 |
Annotation is the process of adding "labels" to data such as text, audio, and images, which is necessary to create the training data required for developing AI.
For example, when developing an AI that recognizes traffic lights for autonomous driving, identified image data must be prepared in advance so the AI can recognize which image shapes are traffic lights. To learn, AI needs this kind of "training data."
On the other hand, such data cannot usually be obtained naturally. Therefore, it is necessary to create training data by labeling actual images of things like traffic lights. This work is annotation.
There are two main methods for carrying out annotation: "manual" and "automated by machine."
When performed manually, it is possible to create high-quality training data, but it naturally requires human resources and a certain period of work.
The other approach is the automation of annotation work by machines. Due to the limited areas where it can be utilized and the difficulty of ensuring accuracy at the current level of technology, it is presently only available for limited purposes. Its main use is in the form of tools that support manual annotation.
The main types of data subject to annotation are as follows.
In image data, training data is created by adding information such as the names of people or objects in photos, segmentation for region extraction, and facial expressions or emotion recognition. In video data, training data is created to achieve tracking and prediction of trajectories or actions. Representative image annotations include the following:
Identifies objects within an image and adds meaningful labels such as "human," "car," or "bicycle" according to the target.
Adds labels to specific regions within an image. For example, it identifies the meaning of the selected region, such as "this part is a human," "this part is a dog," or "this part is a road."
The task of classifying an entire image into a specific category. By labeling the entire image—for example, "Does the image depict a dog or a cat?", "Is the landscape a coast or a mountain?", or "Is the dish Japanese or French?"—it provides criteria for determining the primary features and content of the image.
An annotation method for detecting specific landmarks, primarily used in facial recognition. This task involves classifying an entire image of a face into specific categories.
It makes it possible to read emotions from facial expressions. Also known as facial expression recognition AI, precise annotation allows for the reading of subtle emotional changes, such as the degree of joy, anger, sadness, enjoyment, or interest. It is utilized in stress checks within companies and immigration inspections at airports.
Annotation of audio data is essential for the development of speech recognition AI. To build an AI that converts audio data into text data, training data with text information corresponding to the spoken content is required.
For example, in the development of emotion analysis AI, training data with emotional information such as "happy" or "sad" added to the audio data is necessary. Here, complex annotation is required to analyze the richness of emotions latent in speech, taking into account changes in voice quality and tone.
Annotation of text data is a central task in natural language processing (NLP). To extract some information from text, training data corresponding to the extraction method is required.
For example, in the task of "Named Entity Recognition," which extracts named entities such as person names, organization names, dates, and amounts, named entity information is identified and marked up. In "Dependency Parsing," which extracts relationships between words, dependency relationships between phrases are specified.
When extracting event information, annotation is required to clarify roles such as subject, verb, and object in a sentence. These annotations are indispensable for improving an AI's ability to understand natural language and use it to take meaningful actions.
3. Why is Annotation Important?
The main reason annotation is necessary is that AI needs appropriate training data to learn information and improve accuracy through "supervised learning." Training data refers to data generated by humans adding appropriate labels to data such as photos, audio, and text.
In supervised learning, an AI learns combinations of data and their corresponding labels to understand what each piece of data represents. For example, for an AI to understand whether an image represents a dog or a cat, a dataset where "dog" or "cat" labels are added to each image is required.
Therefore, the less training data there is, the lower the AI's ability to accurately classify or predict new information. This is because when training data is scarce, there are fewer "examples" for the AI to learn an appropriate model.
Conversely, the more training data there is, the more "examples" the AI can learn, making it possible to increase the accuracy of predictions for new data.
In this way, annotation is an important task that supports the improvement of AI accuracy through supervised learning, and is necessary to ensure a sufficient quantity and quality of training data.
Below, we introduce the specific annotation process. The explanation here covers manual annotation. The general process flow of annotation is as follows.
| ① Consideration of Execution Method ② Requirement Definition ③ Execution of Annotation ④ Delivery of Training Data and Utilization in AI |
First, the method for executing annotation is considered. While manual execution is the basis, the scope that can be automated through tools depending on the content of execution is also considered. In this step, the following challenges exist:
・Ensuring necessary human resources
・Finding an outsourcing partner for annotation work
・Selecting which tools and methods are most suitable
When executed manually, annotation work requires a certain amount of human resources. In AI development, it is often difficult to provide all resources for annotation in-house.
Consideration of an outsourcing partner for annotation work is also necessary. Additionally, if using tools to support annotation, which tools and methods to use for execution is considered.
When outsourcing annotation work, a requirement definition is created to specify what kind of work will be outsourced.
Specifically, the content of the work is organized as "how labels will be added" to "what kind of image, audio, and text data."
Additionally, the volume of work, such as how many photos, audio clips, and text data will be labeled, is also considered. Based on this information, cost estimates and actual work will be carried out, so requests should be organized as specifically and accurately as possible.
When actually executing annotation, the establishment of rules is important. Annotation work is high-volume and is generally performed by multiple people.
To ensure annotation quality, clear and easy-to-understand rules are created, and work is carried out in compliance with those rules. It is even more effective to review annotation rules after a certain period has passed once work has begun. This makes it easier to respond to cases where classification cannot be performed well under the rules initially created.
As a quality improvement measure, double-checks by multiple annotators are also effective. Although this incurs costs, it leads to the assurance of quality.
Receive delivery of the training data created by the outsourcing partner. Since the created training data will be used in AI development, it is received in a format that is easy to input into the AI. We recommend specifying the delivery method and file format in advance.
In this article, we explained the overview and specific work processes of annotation. Recent AI development often uses algorithms classified as supervised learning, and in AI development, the preparation of training data must also be considered.
At the same time, it is also true that it is difficult to notice the importance of annotation work if one is not familiar with AI development. To improve the quality of AI, awareness of annotation work is also necessary.