Performance analysis of CRF-based learning for processing WoT application requests expressed in natural language.

Area of Science:

• Artificial Intelligence
• Natural Language Processing
• Internet of Things

Background:

• Investigates the effectiveness of a Conditional Random Fields (CRF)-based learning method.
• Focuses on identifying Web of Things (WoT) application components from natural language user requests.
• Addresses the challenge of automating WoT application composition for user-defined tasks.

Purpose of the Study:

• To develop and evaluate a CRF-based method for recognizing necessary WoT application components.
• To enable the automatic composition of WoT applications based on user requests.
• To leverage existing WoT platforms for training and evaluation.

Main Methods:

• Developed an engine to identify main acts (MAs) and named entities (NEs) from user requests.
• Trained the engine using descriptions of WoT applications (recipes) from the IFTTT platform.
• Utilized a dataset of over 270,000 publicly available recipes for training and testing.

Main Results:

• Successfully built an engine capable of identifying key WoT application components from natural language.
• Demonstrated the effectiveness of the CRF-based approach in recognizing MAs and NEs.
• Collected and processed a large-scale dataset of WoT recipes for robust model training.

Conclusions:

• Shared unique experiences in generating training and test sets from recipe descriptions.
• Assessed the performance of the CRF-based language method for WoT component identification.
• Introduced further research directions based on performance evaluation.