MIT Press

Table 2:

BLEU scores on NIST dev set MT06 from rerankers which are incorporated with various language models. In the language model column X: Y means the language model X is trained on dataset Y. A bigger language model improves the doc-reranker but does not help the sent-reranker.

Proposal model	Language model	Sent-reranker	Doc-reranker
Sent-transformer	LSTM: NIST	49.92	50.24
	transformer-XL: NIST	50.29	50.56
	transformer-XL: NIST + Gigaword	50.19	50.93

Doc-transformer	LSTM: NIST	50.75	51.20
	transformer-XL: NIST	51.27	51.68
	transformer-XL: NIST + Gigaword	51.33	51.99

Proposal model	Language model	Sent-reranker	Doc-reranker
Sent-transformer	LSTM: NIST	49.92	50.24
	transformer-XL: NIST	50.29	50.56
	transformer-XL: NIST + Gigaword	50.19	50.93

Doc-transformer	LSTM: NIST	50.75	51.20
	transformer-XL: NIST	51.27	51.68
	transformer-XL: NIST + Gigaword	51.33	51.99

View Large

This Feature Is Available To Subscribers Only