Potentially interesting further work¶
We finally discuss the potential implications of the current version of autoBOT, how and where it could be further used.
Extensions with contextual neural representations¶
Current version of autoBOT does not include any Transformer-based e.g., sentence representations. Adding this functionality is one API call away, and could notably boost the performance.
Improving the classification phase¶
Given that the current implementation of autoBOT includes the SGD-based linear learners, a natural extension at this point is the use of more involved classifiers. Current implementation (see examples) offers this functionality out-of-the-box.
Speeding up evolution¶
Current implementation of evolution is one of the most basic ones. Should more involved, potentially multi-objective scenarios be considered, it’s possible this step can be drastically improved via e.g., inclusion of Pareto front-based traversals etc.
Meta-transfer learning¶
Given that current implementation of autoBOT results in solution vectors that potentially represent the representation space suitable for a given document, we believe that using such information as priors could speed up the evolution on novel data sets from the same domain.