What Should/Do/Can LSTMs Learn When Parsing Auxiliary Verb Constructions?

There is a growing interest in investigating what neural NLP models learn about language. A prominent open question the of whether or not it necessary to model hierarchical structure. We present linguistic investigation parser adding insights this question. look at transitivity and agreement information auxiliary verb constructions (AVCs) comparison finite main verbs (FMVs). This motivated by theoretical work dependency grammar particular Tesnière ( 1959 ), where AVCs FMVs are both instances nucleus, basic unit syntax. An AVC dissociated nucleus; consists least two words, an FMV its non-dissociated counterpart, consisting exactly one word. suggest that representation should capture similar information. use diagnostic classifiers probe vectors learned transition-based four typologically different languages. find learns if only sequential (BiLSTMs) used architecture but when recursive layer used. explanations for why case looking closely how network happens with representations AVCs. conclude there may be benefits using parsing we have yet found best way integrate our parsers.