MTQA: Text-Based Multitype Question and Answer Reading Comprehension Model

Table 1

Optimal results and innovation of the models.

Model design perspective	Model	Results		Innovations
		EM	F1
Operations on external methods	NAQANet	46.20	49.24	The first model built for DROP corpus, with four decoding heads
	NABERT+	64.61	67.35	Switch to BERT encoding based on NAQANet model
	MTMSN	76.68	80.54	Add two decoding heads and cluster search algorithm based on NABERT+
	TbMS	76.91	79.92	Improve and increase the answer prediction algorithm
	TASE	—	—
Operations on the model itself	QANet + ELMo	27.71	30.33	Put the DROP corpus into the model for training and testing
	BERTBase	30.10	33.36
	GENBERT	68.20	72.80	Use transformer internal structure for decoding and secondary pretraining