Three papers accepted at EMNLP 2024.

The paper “Scope-enhanced compositional semantic parsing for DRT” summarizes Xiulin Yang’s MSc thesis, done in collaboration with the University of Groningen in the context of the Erasmus Mundus Master LCT. Yang applied the compositional, neurosymbolic AM parser to semantic parsing in the DRT formalism; her parser performs especially well on test sets with complex sentences.

In the paper “Strengthening structural inductive biases by pre-training to perform syntactic transformations”, Matthias Lindemann shows that transformers achieve better accuracy on syntax-heavy tasks after an intermediate pretraining step in which they are exposed to syntactic transformations. In contrast to data augmentation methods, the pretraining step is carried out only once, without knowing the target task, and yields a single transformer that can be finetuned to the individual tasks. The paper thus adapts Lindemann’s earlier work on injective a finite-state bias to the case of syntactic structure.

Finally, Yuekun Yao’s paper “Predicting generalization performance with correctness discriminators” was accepted as EMNLP Findings. This paper tackles the question of estimating the accuracy of a parser or semantic parser on unlabeled test data, i.e. test sentences without gold annotations. Unlike previous work, Yao estimates lower and upper bounds for the true accuracy of the parser; across a number of different datasets, these bounds are almost always accurate and tight.

Congratulations to all authors!