LinkedIn today released DeText, an open source framework for process-related ranking, classification and language generation tasks in natural language. It uses semantic matching and uses deep neural networks to understand members’ intentions in search and recommendation systems. According to LinkedIn, it can be applied to a variety of tasks in general, including search and recommendation rankings, classifying multiple classes, and understanding queries.
According to Weiwei Guo, Senior Engineering Manager at LinkedIn, DeText was developed with sufficient flexibility to meet the needs of various production services. It is based on state-of-the-art algorithms that are integrated in an end-to-end model in which the variables are updated together. However, it tries to reconcile its overall effectiveness with high efficiency.
“With the framework, users can better use models and embeddings in real applications,”
DeText contains several components, all of which can be customized using pre-installed templates:
- An embedding layer that converts a sequence of words into a matrix, a series of numbers arranged in rows and columns. (Matrices are often used to represent the data that is fed into AI models.)
- Text encoding models that map text data in fixed-length embeds, or numerical representations from which algorithms can learn.
- An interaction level that generates features based on the text embeds above.
- Feature processing that combines traditional features with the interaction features (deep features) in jointly trained broad linear models and deep neural networks. (In this context, properties refer to individual measurable properties and characteristics of the observed phenomena.)
- An MLP layer that combines broad and deep features.
To run DeText, a development environment with the required dependencies, including Python, must be created and started. After installation, however, a sample model for the sample data set from the GitHub repository can be trained.
“Deep learning-based processing of natural language has the potential to deepen understanding of human intentions through search and recommendation systems. However, the ability to use models effectively in commercial applications remains unwieldy due to the high computing load, especially when it comes to evaluating results and classifying text, ”continued Guo. “DeText can be viewed as a wireless drill that allows users to easily exchange and optimize natural language processing models depending on the application.”
The use of AI by LinkedIn is widespread. In October 2019, Microsoft’s own platform pulled back the curtains for a model that generates text descriptions for images uploaded to LinkedIn that were created using Microsoft’s Cognitive Services platform and a unique data set derived from LinkedIn. LinkedIn’s “Recommended Candidates” function learns the recruitment criteria for a specific role and automatically displays relevant candidates in a special tab. The AI-driven search engine uses data such as the kind of things that people post on their profiles and the searches that candidates perform to make predictions for the most suitable jobs and job seekers. In addition, LinkedIn’s AI-driven moderation tool automatically detects and removes inappropriate user accounts.