Tracing the shifting collocates of complex prepositions with diachronic word embeddings: supplementary materials
The 11th International Conference on Construction Grammar (ICCG11) takes place from Wednesday 18 till Friday 20 August 2021. It covers a broad range of topics related to Construction Grammar approaches to language. My talk is scheduled on Friday at 2.15pm. This post contains supplementary materials, namely high-resolution figures and a full bibliography.
Complex prepositional constructions in the form <in/at the middle/midst/center/heart of NP> denote a relationship of internal location between a trajector, i.e. a located entity which is the primary focal participant, and a landmark, i.e. a reference entity which is the secondary focal participant (Langacker 1987: 225–228).
Locative prepositions denoting internal location are not synonyms because each imposes specific constraints that go beyond the spatial criteria that are ordinarily used: the landmark’s internal plurality, its boundedness, the magnitude of its referent, and the degree of functional dependence between the trajector and the landmark (Gréa 2017).
I adapt the above to the study of English. Using data from the Corpus of Historical American English (Davies 2010), I assess to what extent each prepositional construction imposes its own construal and how this construal has shifted over the past 200 years.
I draw upon three key assumptions from previous research: (a) changes in the collocational patterns of a linguistic unit reflect changes in meaning and/or function (Hilpert 2008, 2011; Hilpert and Gries 2009); (b) meaning can be modeled with dense word vectors (Mikolov et al. 2013b; Mikolov et al. 2013c, Author 2019); (c) by supplementing frequency-based methods with distributional word representations (Baroni et al. 2014; Turney and Pantel 2010), one can trace semantic shifts more precisely and with greater explanatory power (Hilpert 2016; Kulkarni et al. 2015; Perek 2016, 2018).
Word vectors were obtained by running a prediction-based shallow neural model on the COHA, namely Skipgram with negative sampling (Mikolov et al. 2013a). I made a semantic vector space of the most distinctive landmark collocates of all complex prepositions across the whole period covered by the corpus (1810s–2000s). Only the most distinctive collocates of each construction appear in the vector space. Next, I divided the corpus into four arbitrary periods (1810s–1860s, 1870s–1910s, 1920s–1970s, and 1980s–2000s). I included frequency-based density plots for each period and each prepositional construction. I obtained one semantic vector space per period. The resulting maps confirm that the four prepositional constructions involve different types of constraints and that these constrains have changed since the 1810s. For example, <in the heart of NP> has shifted from profiling landmarks that denote institutional or geographic bounded areas (e.g. countries, nations, world, etc.) to interactional events whose boundedness is unclear but whose magnitude is foregrounded (e.g. dispute, problem, issue, etc.).