On the structure of annotation platform

One 、 background

There are many open source annotation tools , But it only solves some links in the data annotation link . For flowable annotation operations , In addition to supporting the image 、 Text and audio and video labels , Data access also needs to be considered 、 Allocation of personnel 、 Mark progress management and sign boards .

Two 、 Core functions

1、 Images

Image annotation is really time-consuming and laborious . According to the subdivision of computer vision , Annotation tasks can also be divided into 2D Gesture tagging 、 Split dimension 、 Classification marking and object marking, etc . Therefore, the annotation function needs to meet the following requirements 、 Line 、 Add, delete, modify and check the face , Also pay attention to visual design , It is convenient to distinguish the annotation results .

2、 Text

common NLP Tasks have text categorization 、OCR、NER、 Text summary, etc , Therefore, the annotation function needs to meet the requirements of word 、 word 、 Statement and paragraph selection , And multi label display .

Others have not been contacted yet , So I won't say much .

3、 ... and 、 Platform functions

1、 Dataset Management

The basis of data set management is to add, delete, modify and query data , In addition, the data entry method should be considered , Such as local uploading and database reading .

2、 Label management

For labels , There is a large demand for customization , Therefore, in addition to the conventional template, label management , If you can support configurable capabilities , The general performance of the platform will be greatly improved .

Labels are generally defined and maintained by the employer , In addition to adding, deleting, modifying and checking , How tags interact on the annotation page 、 Whether shortcut keys are supported , It will affect the dimensioning efficiency

3、 Personnel management

For the platform , Personnel management is also the basic addition, deletion, modification and query , But we need to consider the relationship between people and data 、 The relationship between people and tasks 、 The relationship between people and groups .

For groups , In addition to adding, deleting, modifying and checking personnel management , More is role management , For example, marking personnel and inspectors .

4、 task management

When creating a dimension task , Data needs to be specified , A designated person is required , You also need to specify the workflow of the task . The workflow can be divided into ：

1、 Dimension only

2、 mark + Check

3、 mark + Check + rejected

It is recommended that the workflow be as simple as possible , The logic involved later is more complicated , For example, staffing 、 Task recycling 、 Data statistics, etc. .

Task management includes addition, deletion, modification and query , Also increase the distribution of tasks 、 Task transfer and task progress kanban .

5、 Label Page

Annotation pages should increase interaction as much as possible , To improve labeling efficiency .

For example, image annotation , Need to have the image zoom function ; Text label , The presentation of text needs to be optimized ;

6、 Kanban

Kanban can be subdivided into data kanban 、 Task Kanban and personnel Kanban .

Data Kanban is mainly used to feedback and mark key indicators , For example, mark the quantity 、 Dimension accuracy 、 Inspection quantity and inspection accuracy ;

Task Kanban is the display of task details ;

The personnel Kanban focuses on the number of labels 、 Efficiency and other indicators ;

Four 、 Interface capabilities

For mass data annotation , It is not suitable for creating tasks manually , So we need to provide API.

For the above functions ,API Can open the ability to add and query , Limit the ability to delete and modify , Prevent misoperation .