編碼

The science of Neufast system

Our mission is to provide you and your candidates with an enjoyable and highly accurate screening experience—without bias or stress. We achieve this by taking a deeply scientific, multi-layered approach that ensures our screening tests are always valid, reliable, and fair.

Visual, Audio and Textual Based Comprehensive Analysis

Candidate’s apparent features, prosody features and semantic features are extracted from his/her videos based on computer vision, speech recognition and natural language processing techniques.

Visual channel for apparent feature extraction

Apparent features such as facial expression and gesture are implicitly extracted by computer vision technique.

Visual channel for apparent feature extraction

Apparent features such as facial expression and gesture are implicitly extracted by computer vision technique.

Audio channel for prosody feature extraction

Prosody features such as tonality, pace and rhythm are implicitly extracted by audio analysis technique.

Visual channel for apparent feature extraction

Apparent features such as facial expression and gesture are implicitly extracted by computer vision technique.

AI Model Architecture

Our AI model further analyzes these features in a comprehensive way with complex task-oriented adaptive weighting scheme to infer candidate’s Impression Score and Work Performance Score.

​1. Collected visual, audio, and text data will be sent to their respective channels

  • Channels are constricted by deep learning models.

​2. Apparent(visual), prosody(audio), and semantic(textual) features will be extracted by these channels correspondingly.

  • Apparent features: faical expression, gesture...

  • Prosody features: tonality, pace, rhythm...

  • Semantic features: syntactic, contextual meaning...

Visual, Audio and Textual channels

(Deep Learning Models)

Apparent

features

Prosody

features

Semantic

features

Customized Regression

​3. Extracted features will be sent to the a customized regression layer for predicting the impression/competency scores.

Impression / Competency Scores

Principles of our AI model
Principles of our AI model

Neufast embraces good data ethics in operation and the development and use of AI. To ensure the fairness of AI model, the development process is under internal governance:

Data Preparation
  • No manipulation of the training data collection leading to a biased model in favor of some certain group.

  • No data preprocessing that changes the data distributions for different groups

Model Building & Training 
  • All key aspects of the model building process are reviewed

  • Salient numerical results are copied into the sheet

  • All significant deviations from the standard model training process are well documented

Model Testing

Adverse Impact Assessment is done to evaluate if our trained AI model complies with the UGESP's four-fifth rule with respect to different groups in either:

  • Demography

  • Educational background

  • Or other groups defined by clients