Typical case of data annotation: how does jinglianwen technology help enterprises build data solutions

As one of the three elements of the development of artificial intelligence , The role of data is crucial .

Jinglianwen technology provides professional data scheme design for enterprises , Help enterprises quickly build data solutions that match their algorithm models , Explore new business areas .

Jinglianwen technology is the largest in the Yangtze River Delta AI Basic data service providers , Self developed data annotation platform , Establish mature annotations 、 to examine 、 Quality inspection mechanism , Support computer vision 、 Speech Engineering 、 Multi type data annotation projects such as natural language processing .

With high-quality training data and hundreds of large-scale technology enterprises around the world 、 Scientific research institutions maintain in-depth cooperation .

  User stories

1. Well known security manufacturers contraband 2d Split dimension

One 、 demand ：

50 10000 images 33 Categories of contraband segmentation labels

Two 、 Project difficulties ：

1） Large amount of annotation data , Tight schedule .

2） There are many prohibited varieties and labels , It is difficult to distinguish similar objects .

3、 ... and 、 Solution ：

1） For the annotation rule system , Configure exclusive business 、 project manager 、 Annotator 、 Quality inspector 、 Technical personnel , Work overtime to train the team , Learning professional knowledge , Adjust the marking specification and repair data .

2） The pre annotation ability and automatic quality inspection function of jinglianwen annotation platform effectively improve the annotation efficiency and quality

3） Full quantity quality inspection and two rounds of random inspection, three times of data quality inspection, and then submit data in batches , Timely confirm with customers about the problems of submitted data feedback

Four 、 Delivery results ：

Complete the delivery in full within the construction period , First pass rate 98%.

 2. Government smart city video annotation project

One 、 demand ：

2 Ten thousand vehicles are tracked 、1000+ Event type annotation

Two 、 Project difficulties ：

1） Total video duration 4000 minute , There are many annotation scenes （ Including urban roads , Expressway , Urban viaduct, etc ）, There are many types of events , Tight schedule .

2） Video frame extraction 、 Data cleaning is time-consuming .

3） There are many types of events displayed in the video data , The effect of ordinary algorithm auxiliary annotation is not obvious .

3、 ... and 、 Solution ：

1） For the annotation rule system , Arrange business 、 project 、 mark 、 Quality testing 、 The technical personnel are familiar with the background and purpose of the project 、 The rules 、 matters needing attention 、 difficulty 、 Platform operation 、 Project requirement （ Accuracy rate 、 Daily output ） Training .

2） The continuous frame pre annotation function of Jinglian text annotation platform effectively improves the annotation efficiency and quality

3） Full quantity quality inspection and two rounds of random inspection, three times of data quality inspection, and then submit data in batches , Monitor progress in real time , Check while repairing , Solve problems at any time .

Four 、 Delivery results ：

Complete the delivery in full within the construction period , First pass rate 98.5%.

Besides , In the field of computer vision , Jinglianwen technology in rectangular pull frame 、 Polygon annotation 、 Semantic segmentation 、 Instance segmentation 、 Face key points 、 Lane line 、 Video continuous frame annotation 、 I have rich annotation experience in emotional analysis and other projects .

1. National Laboratory Voiceprint recognition voice mark

One 、 demand ：

Multi person conversation voice in uncertain scenes 50000 Bar labels

Two 、 Project difficulties ：

1） The marking requirements in the early stage of the project are not clear , Modify the annotation rules repeatedly .

2） Part of the audio has multi regional accents , The effect of algorithm aided annotation is not obvious , It is also difficult to identify manually .

3、 ... and 、 Solution ：

1） Equipped with 3 Project manager and tagging team with more than years of experience in voice tagging project management .

2） Annotation rules for changes , Instantly feed back the problems encountered in the labeling process and provide multiple solutions , Plan to reserve repair time , Establish an early warning mechanism , Strictly control key time nodes .

3） The pre annotation and automatic quality inspection functions of jinglianwen annotation platform effectively improve the annotation efficiency and quality

Four 、 Delivery results ：

Complete the delivery in full within the construction period , First pass rate 98%.

  Besides , In the field of voice Engineering , Jinglianwen technology is in ASR Voice transcribe 、 Voice cleaning 、 Speech cutting 、 Emotional judgment 、 Phoneme labeling 、 Prosody tagging 、 I have rich annotation experience in machine translation and other projects .

1. National Laboratory Text label

One 、 demand ：

Text label 20 Ten thousand , Contains intention matching 、 General abbreviation 、 Integrated induction 、 Different rules such as label classification

Two 、 Project difficulties ：

1） It is difficult to mark a wide range of contents , High requirements for the quality and ability of marking personnel .

2） There are many text types 、 Lots of scenes 、 Length , Lots of labels , Cumulative labels and options up to 530 More than a

3） Accuracy requirements 99%, Tight schedule .

3、 ... and 、 Solution ：

1） Equipped with 3 In the above NLP Project managers and marking teams with project management experience .

2） Arrange to mark the team's understanding of the project background 、 Purpose 、 The rules 、 matters needing attention 、 difficulty 、 Platform operation 、 Project requirement （ Accuracy rate 、 Daily output ） Conduct training and assessment , Assessment and elimination 40%, The remaining 60% Personnel enter the formal task .

3） Analyze the project structure according to the project requirements , be based on WBS The principle decomposes the project into a tree view layer by layer according to its internal structure and the order of implementation process , Form a relatively independent 、 Project responsibilities of each unit of the project that is easy to manage and check 、 Progress, etc. shall be specifically implemented to each participant of the project , Ensure the marking quality .

Four 、 Delivery results ：

Complete the delivery in full within the construction period , First pass rate 99%.

  Besides , In the field of natural language processing , Jinglianwen technology is in ocr Transcribe 、 Statement generalization 、 Sentiment analysis 、 Named entity 、 Part of speech tagging 、 Slot extraction 、 Text matching 、 Intention matching 、 Text cleaning 、 I have rich annotation experience in information extraction and other projects .