Is one head enough? Mention heads in coreference annotations compared with UD-style heads

We present an empirical study that compares mention heads as annotated manually in four coreference datasets (for Dutch, English, Polish, and Russian) on one hand, with heads induced from dependency trees parsed automatically, on the other hand. For parsing, we used UDPipe 2.6, a modern parser trained using the Universal Dependencies collection. We show that majority of mismatches (64%–94%) can be attributed to several classes of systematic differences in how the notion of head is treated in the respective data resources, while mismatches caused by parsing errors are relatively rare (4%–15%). Our conclusion is that consistency would be gained in (and across) coreference resources after migration to UD-style mention heads, without losing substantial information. This can be achieved with sufficient accuracy using modern dependency parsers even for coreference corpora that lack manual head annotation.