id	author	title	date	pages	extension	mime	words	sentences	flesch	summary	cache	txt
work_w42xlvpc3nagplvbetz6f74zyy	Tal Linzen	Assessing the Ability of LSTMs to Learn Syntax-Sensitive Dependencies	2016	16	.pdf	application/pdf	11727	2555	79	model to encode or approximate structural information; nevertheless, it succeeded in recovering the majority of agreement cases even when four nouns of the language model trained without explicit grammatical supervision performed worse than chance on the We note that subject-verb number agreement is only one of a number of structuresensitive dependencies; other examples include negative polarity items (e.g., any) and reflexive pronouns model's error rate was affected by nouns that intervened between the subject and the verb in the linear Figure 2: (a-d) Error rates of the LSTM number prediction model as a function of: (a) distance between the subject and the verb, in dependencies that have no intervening nouns; (b) presence and number of last summary, we conclude that while the LSTM is capable of learning syntax-sensitive agreement dependencies under various objectives, the language-modeling models predict a singular verb even though the number of the subject conservation refugees should be	./cache/work_w42xlvpc3nagplvbetz6f74zyy.pdf	./txt/work_w42xlvpc3nagplvbetz6f74zyy.txt