Two pretraining objectives that have been successful for pretraining neural networks used in transfer learning NLP are autoregressive (AR) language modeling and autoencoding (AE). Autoregressive language modeling is not able to model deep bidirectional context which has recently been found to be effective in several downstream NLP tasks such as sentiment analysis and question answering. On the other hand, autoencoding based pretraining aims to reconstruct original data from corrupted data. A popular example of such modeling is used in BERT, an effective state-of-the-art technique used to address several NLP tasks. One advantage of models like BERT is that bidirectional contexts can be used in the reconstruction process, something that AR language modeling lacks.

Pre-trained language models have been successfully used in response generation for open-domain dialogue. Four main frameworks have been proposed: (1) Transformer-ED using Transformer encoder and decoder separately for source and target sentences; (2) Transformer-Dec using Transformer decoder for both source and target sentences; (3) Transformer-MLM using Transformer decoder that applies bi-directional attention on the source side and left-to-right attention on the target side with masked language model objective; and (4) Transformer-AR that uses auto-regressive objective instead. In this study, we compare these frameworks on 3 datasets, and our comparison reveals that the best framework uses bidirectional attention on the source side and does not separate encoder and decoder. We also examine model discrepancy, and our experiments confirm that the performance of a model is directly impacted by the underlying discrepancies. We then propose two correction methods to reduce the discrepancies, and both improve the model performance. These results show that discrepancies is an important factor to consider when we use a pre-trained model, and a reduction in discrepancies can lead to improved performance.