Introduction

Studies on the categorical organisation of conceptual knowledge suggest that animate and inanimate concepts are represented differently in mind. While the categorical organisation of conceptual knowledge is a contentious issue as most researchers view such semantic dis- tinctions to be the two ends of a continuum (e.g., Brysbaert, Warriner & Kuperman, 2013), the animate-inanimate distinction seems to be well-grounded based on neuropsychological evidence. Caramazza & Shelton (1998) review several studies on selective brain damage (see also Capitani, Laiacona, Mahon & Caramazza, 2003) as well as providing original evidence from a brain-damaged subject who had the inability to name animate objects. After a rigor- ous examination on a battery of tasks, they conclude that the obtained patterns cannot be explained solely on sensory/functional grounds (e.g., animate objects yield similar sensory responses), giving support to a categorical distinction between animate and inanimate con-

due to British-American English differences, a few words were replaced (e.g.,movie ticket→theater ticket). Lastly, wherever possible, we chose similar words which would possess the same featural representations (e.g., icepop→lollipop_{). For more details on the datasets used see Appendix B.}

cepts. Caramazza & Shelton (1998) further argue that such distinctions are to be expected for categories for which there are evolutionary survival pressures to be developed separately (Gelman, 1990).9

An issue related to the above is whether this suggested distinction between animate and inanimate concepts is manifested in any way in linguistic usage. This implication is relevant to our models as if this distinction is reflected in the distributional patterns of words; then we would expect the dsms introduced above to be able to capture it. The results presented in Table 5.1 show that neither WordNet nor the neural embeddings have any trouble distin- guishing between animate and inanimate concepts. However, while WordNet synsets are marked for animacy as they are all descendants of living_thing.n.01(see Fig. 1.4), it does

not necessarily mean that the differences in the dsm are a result of different distributional patterns between animate and inanimate concepts. The categorisations usually used in these experiments are too restrictive (animals vs. furniture pieces) giving rise to alternative explana- tions. In what follows we show some examples from English and Bantu languages showing that animacy often determines word order and morphological marking.

Looking at the Bantu languages, we see that they use different morphological markers to distinguish between animate and inanimate nouns. Examples (5.13) and (5.14) show the distributional patterns in Congo-Swahili in which adjectives, connectives and the verb have to agree with the grammatical marker of the noun which is widely determined by its semantics. (5.13) Batoto

2-children ba-mingi2-many b-a2-con ikidem.7 ki-pande7-piece ba-li-kwa-ka-po2-pst-be-asp-loc ‘Many children from this area were there.’

(5.14) Bi-le

8-dem bi-ntu8-thing bi-ote8-all bi-li-kwa-ka8-pst-be-asp mw-a18-con KalumbuKalumbu bi-l-ingia8-pst-enter umudem.18

mu-nyumba

18-9.house

‘All the stuff that was in Kalumbu’s (house) was transferred into this house.’

where 2 and 8 denote the corresponding noun class (2 = living things; 8 = things). Such constructions in which elements of the sentence agree with a semantically determined noun class are well attested in many languages from distinct language families such asDyirbal

(Dixon, 1972) andBininj Gun-Wok(Evans, 2003). In these languages, distributional patterns can signal the presence or absence of an animate or inanimate concept.

Regarding word order,animacydistinctions are quite common in the world’s languages. From a typological point of view, the presence or absence of an animacy feature is important

9According to Caramazza & Shelton (1998), these survival pressures stem from the fact thatanimals_are potential predators but also a source of food, whereasplants_{are sources of food and medicine (p. 20).}

in determining word order choices in dative alternation and agreement patterns in languages from different language families (Evans, 1997; Hawkinson & Hyman, 1974; Morolong & Hyman, 1977; Polinsky, 1996). In the case of English, while animacy is not marked overtly as in Swahili (see above), Bresnan & Hay (2008) and Hinrichs & Szmrecsányi (2007) present evidence that in English varieties spoken in New Zealand as well as in the us and the uk, animacy is a significant predictor of the choice of syntactic paraphrases such as dative and genitive alternation.

Finally, current syntactic theories (Chomsky, 1995; Gazdar, Klein, Pullum & Sag, 1985) pro- pose that during the computation of the grammar of a sentence, certain semantic features are ‘checked’. In other words, the elements of a sentence need to have some semantic congruence which is determined during syntactic parsing. While the necessity of this ‘checking’ varies between theories as well as between languages, it seems that theories of syntax converge on the fact that some ‘core’ conceptual distinctions are available to the speaker at all times. In the present context, this becomes necessary as various researchers have argued that a [±animacy]

feature exists between genetically diverse language in determining the arguments of a dp (see Adger & Harbour, 2007, for an application in Kiowa and the references therein for applications to other languages).

Considering the above discussion, we expect that the neural embeddings would be able to capture the effects observed in the behavioural results in that the grammatical system would be learnable. However, since for the dsm,animacyis only indirectly inferred from the distributional patterns of the words would not expect such high generalisation rates as obtained by WordNet (where animacy is marked as a distinct feature).