Animacy encoding in English

Abstract

We report on two recent medium-scale initiatives annotating present day English corpora for animacy distinctions. We discuss the relevance of animacy for computational linguistics, specifically generation, the annotation categories used in the two studies and the interannotator reliability for one of the studies.

Empirical evaluations of animacy annotation

“…The annotation manual (Teleman, 1974) states that a markable should be tagged as human (HH) if it may be replaced by the interrogative pronoun vem 'who' and be referred to by the personal pronouns han 'he' or hon 'she'. There are clear similarities between the annotation for human reference found in Talbanken05 and the annotation scheme for animacy discussed (Zaenen et al, 2004).…”

“…Examples include metonymic usage of nouns, as in (5) and nouns occurring in dereferencing constructions, such as predicative constructions (6), titles (7) and idioms (8) It is interesting to note that the main variation in annotation stems precisely from difficulties in determining reference, either due to bleak type level properties such as for the abstract nouns, or due to properties of the context, as in the reference shifting constructions. The small amount of variation in the human annotation for animacy clearly supports a type-level approach to animacy, however, underline the influence of the linguistic context on the conception of animacy, as noted in the literature (Zaenen et al, 2004;Rosenbach, 2008).…”

“…If we examine the errors for the inanimate class we indeed find evidence of gradience within this category. The errors contain a group of nouns referring to animals and other living beings (bacteria, algae), as listed in (9), as well as one noun referring to an "intelligent machine", included in the intermediate animacy category in Zaenen et al (2004). Collective nouns with human reference and organizations are also found among the errors, listed in (11).…”

“…Within the group of inanimates, further distinctions are made between concrete and non-concrete inanimate, as well as time and place nominals. 2 The annotation scheme described in Zaenen et. al.…”

See No Evil, Say No Evil: Description Generation from Densely Labeled Images

“…We define semantic types by finding synsets of labels in objects that correspond to high level types, a list motivated by the animacy hierarchy (Zaenen et al, 2004). 8 Type features indicate the type of the object referred to by an alignment variable as well as the cross product of types when an alignment variable is on conditioning side of a distribution (e.g.…”