A course project for Advanced Formal Language Theory class, Spring 2022 at ETH Zurich.

The project aims to explore the SoPa model, based on "SoPa: Bridging CNNs, RNNs, and Weighted Finite-State Machines" by Roy Schwartz, Sam Thomson and Noah A. Smith, ACL 2018, which is characterized by the patterns matching mechanism.

As in the SoPa experiment, 100 movie reviews from SST dataset with binary labels are also used in our ablation study.

We executed the study on three levels, including automaton level, automata level, as well as on the SoPa as a whole. For reproducility, please refer to README_ reproducible_code.md file.

We conclude valuable insights regarding how parameters affect the model performance through multiple experiments.

• First, increasing patterns’ width improves model performance while the length has little effect. Second, in case exclud- ing ε-transitions, the bigger max-steps-forward is, the better performance is yielded.

• Also, there is a downward trend in test accuracy as we increase the value of shared_sl because it decreases the complexity of the model.

• Third, the log-space max-times semiring achieves the best accuracy.

• Finally, training the patterns on multiple documents improves the gradient com- putation and hence the convergence rate.

For questions, comments or feedback, please email [email protected]

Code | Report