id	author	title	date	pages	extension	mime	words	sentence	flesch	summary	cache	txt
cc08hd79z5w	Justin Donald DeBenedetto	Multiset and DAG Automata for Abstract Meaning Representation	2021		.txt	text/plain	537	20	44	Specifically, we: Define a new translation from weighted multiset regular expressions to weighted multiset automata more direct and compact than previous work as well as a new composable representation of partial runs of multiset automata more efficient than previous work Show that the Transformer's sinusoidal positional encodings can be viewed as a multiset automaton Prove that complex-weighted multiset automata with only self-loops can approximate real-weighted multiset automata and extend DeepSets to compute the forward weights of complex-weighted multiset automata enabling it to handle a new task Modify and implement an extended DAG recognition algorithm to use complex diagonalized multiset automata in place of positional encodings in a Transformer network for AMR-to-text generation Demonstrate that these improvements now allow such a system to train on a GPU, opening new opportunities for future systems based on DAG and multiset automata These methods build upon existing work, utilizing in particular multiset automata and directed acyclic graph (DAG) processing algorithms.	cache/cc08hd79z5w.txt	txt/cc08hd79z5w.txt