With the development of deep learning techniques, research on learning and analyzing knowledge through meaning and syntax inherent in text has been actively conducted in recent language models. It is interestingly that it starts with a very simple str...
With the development of deep learning techniques, research on learning and analyzing knowledge through meaning and syntax inherent in text has been actively conducted in recent language models. It is interestingly that it starts with a very simple structure and progresses by adding ideas one by one to overcome the limitations of the previous model. Some studies evaluate language knowledge of neural networks using exploratory tasks that train them to directly predict characteristics of sentences (e.g., syntax tree depth) or parts of sentences (e.g, language parts) using only pre-trained representations of neural networks(NNs) as inputs. However, the characteristics of the model required to make syntactic generalization are not yet well known and unclear.
Accordingly, this study examines the characteristics and evaluation methods of existing language models and then evaluates the ability of language knowledge by applying them to the LSTM language model based on the preceding paper of Hu et al., (2020) which uses language models to systematically evaluate syntactic generalization.