Emergent Predication Structure in Hidden State Vectors of Neural Readers
This work addresses the interpretability of neural models for researchers in NLP, offering incremental insights into their internal representations.
The paper investigates whether hidden state vectors in neural reading comprehension models encode predication structures, providing evidence that these vectors represent atomic formulas linking semantic predicates to entity identifiers.
A significant number of neural architectures for reading comprehension have recently been developed and evaluated on large cloze-style datasets. We present experiments supporting the emergence of "predication structure" in the hidden state vectors of these readers. More specifically, we provide evidence that the hidden state vectors represent atomic formulas $Φ[c]$ where $Φ$ is a semantic property (predicate) and $c$ is a constant symbol entity identifier.