The meaning of words in a language, how it changes, and how new meanings are created can tell us a lot about the cultural history of its speakers. In the recent history of English, the word ‘gay’ has acquired a completely new meaning (‘homosexual’) and mostly lost its old one (‘happy’); ‘mouse’ now refers to an object that did not exist a century ago, the computer mouse, as well as a rodent. These changes may seem immediately understandable, and easily explained in light of recent societal shifts (the development of computers, or changes in our conception of sexuality). But what happens when we move further into the past, towards periods for which our only source of information is texts, often in languages that are no longer spoken? Can we understand cultural change through the lens of language? Scholars of ancient cultures have long practised the art of ‘reading texts slowly’, comparing the usage of single words across centuries. In recent decades, however, advances in automatic text processing have made the prospect of scaling up this approach, making it more efficient and accurate through computational means, very appealing.

Distributional Semantics (DS) is a framework that allows us to study the meaning of words and how it changes, starting exclusively from the way these words are used (Lenci 2008). The essential principle of DS is that ‘you shall know a word by the company it keeps’ (Firth 1957): the meaning of a word can be understood by studying what other words it is used together with. These associations can also reveal underlying cultural phenomena: the words ‘peoplehood’ and ‘nationhood’ may appear similar, and were created about 50 years apart, but only the former is strongly associated with the adjective ‘Jewish’ [source: ‘peoplehood’ and ‘nationhood,’ Merriam-Webster.com Dictionary, https://www.merriam-webster.com/dictionary/peoplehood; https://www.merriam-webster.com/dictionary/nationhood. Accessed 7/27/2021]. We can tell that this makes the two words different; but can we be more precise in our assessment? DS allows us to translate these patterns of association into a mathematical model, through which any two words can be directly compared. DS models are multidimensional vector spaces, in which each word is positioned based on which other words it occurs with; we can then measure the distance between any two words in the space, which indicates how closely related they are in meaning.

We can also compare the meanings of the same word through different time periods in a corpus of historical texts, thereby identifying changes in meaning that can reflect changes in the cultural context. This prospect is particularly appealing when studying past cultural shifts, especially in dead languages, where scholars can’t ask a native speaker how the meaning of a word has changed in their living memory. If we look at ancient Greek and Latin, two of the languages most widely spoken at the time of the birth and diffusion of Christianity, we can see the effects of that seismic cultural change in the trajectory of word meaning. If we look at the 50 words that changed their meanings most significantly in ancient Greek between the last few centuries BCE and the first centuries CE, we see that most of them belong to religious and/or metaphysical contexts, like ‘God’, ‘father’, ‘son’, ‘angel’, etc. (Rodda, Senaldi & Lenci 2017). These changes, however, do not affect some of the basic metaphors governing language: in Latin, words related to the concept of time are always closely associated in distributional terms with words describing space, which reflects how a concept can be used as a metaphor of the other (Nowak 2019).

The challenge when applying DS models to dead languages, which often have relatively small, ‘closed’ corpora (that are difficult or impossible to expand), is how to assess their accuracy without relying on circular reasoning. The fact that a model shows a cultural shift around the advent of Christianity may mean that the model is accurate enough for that specific purpose, but in order to be able to use it as a tool for further research we also want it to detect other, unexpected waves of cultural change. There are various ways to assess how well our mathematical predictions conform to actual meaning relations in a given language: for instance, we can compare the distributional models to existing resources, like dictionaries and ancient works of lexicography (Rodda, Probert & McGillivray 2019). We can also rely on the input of present-day expert annotators to derive quantitative data on how different usages of a word display different meanings depending on the time in which the text was written, but also its genre (McGillivray et al. 2019). Both of these are ways to meet the challenge of assessing the DS models, but they need to be scaled up and made compatible with more historical languages.

Distributional Semantics can allow us to answer questions that are relevant to various humanities disciplines relying on the analysis of historical texts: how do we trace the evolution of concepts over time? What is the relationship between vocabulary change and socio-cultural change? How does language innovation take place? A synergy between the traditional ‘slow-reading’ model of philology and the advanced processing of texts at scale may be the perfect way to address them.

 

References

Firth, J. R. (1957). A synopsis of linguistic theory, 1930-1955. Studies in linguistic analysis.

Lenci, A. (2008). Distributional semantics in linguistic and cognitive research. Italian Journal of Linguistics 20(1): 1-31

McGillivray, B., Hengchen, S., Lähteenoja, V., Palma, M., Vatri, A. (2019). A computational approach to lexical polysemy in Ancient Greek. Digital Scholarship in the Humanities, 34(4): 893-907, https://doi.org/10.1093/llc/fqz036

Nowak, K. (2019). Tempus mutatur: analysing collocations of tempus ‘time’ with distributional semantic models. Lemmata Linguistica Latina. Volume I: Words and Sounds, edited by Holmes, N., Ottink, M., Schrickx, J., Selig, M. Berlin, Boston: De Gruyter: 69-85. https://doi.org/10.1515/9783110647587-005

Rodda, M.A., Probert, P., McGillivray, B. (2019). Vector space models of Ancient Greek word meaning, and a case study on Homer. Traitement Automatique Des Langues, 60(3): 63-87.

Rodda, M.A., Senaldi, M.S.G., Lenci, A. (2017). Panta rei: Tracking Semantic Change with Distributional Semantics in Ancient Greek. Italian Journal of Computational Linguistics, 3: 11-24. https://doi.org/10.4000/ijcol.421