Do Not Fire the Linguist: Grammatical Profiles Help Language Models Detect Semantic Change

Abstract

Morphological and syntactic changes in word usage-as captured, e.g., by grammatical profiles-have been shown to be good predictors of a word's meaning change. In this work, we explore whether large pre-trained contextualised language models, a common tool for lexical semantic change detection, are sensitive to such morphosyntactic changes. To this end, we first compare the performance of grammatical profiles against that of a multilingual neural language model (XLM-R) on 10 datasets, covering 7 languages, and then combine the two approaches in ensembles to assess their complementarity. Our results show that ensembling grammatical profiles with XLM-R improves semantic change detection performance for most datasets and languages. This indicates that language models do not fully cover the finegrained morphological and syntactic signals that are explicitly represented in grammatical profiles.An interesting exception are the test sets where the time spans under analysis are much longer than the time gap between them (for example, century-long spans with a one-year gap between them). Morphosyntactic change is slow so grammatical profiles do not detect in such cases. In contrast, language models, thanks to their access to lexical information, are able to detect fast topical changes.

Unsupervised Semantic Variation Prediction using the Distribution of Sibling Embeddings

“…Likewise, methods for comparing sets of target word representations computed as contextualised word embeddings have also been proposed. Such methods include comparing the average (Martinc et al, 2020;Beck, 2020;Kutuzov and Giulianelli, 2020;Laicher et al, 2021;Giulianelli et al, 2022; or each pair of embeddings (Kutuzov and Giulianelli, 2020;Laicher et al, 2021). Additionally, Aida and Bollegala (2023b) and Nagata et al (2023) have proposed methods that consider the variance in the sets of embeddings.…”

“…After that, we evaluate the performance on SCD tasks. In this paper, we use the two benchmark datasets -SemEval-2020 Task 1 (Schlechtweg et al, 2020) (covering English (En), German (De), Swedish (Sv) and Latin (La)) and RuShiftEval (covering Russian (Ru)), which have also been used in prior work on SCD Giulianelli et al, 2022;Cassotti et al, 2023). 7 Statistics of those datasets are summarised in Table 2.…”

“…SCD scores are calculated by the average cosine distance (w/ TA + CD). Giulianelli et al (2022) introduced an ensemble method combining the average cosine distance from MLMs with the cosine distance between morpho-syntactic features labelled from universal dependencies (CD & UD + CD). It is the current SoTA method in SemEval La.…”

“…Consequently, we see that all rows are considered equally important when computing the semantic change scores in this case. Much prior work on SCD use the Euclidean distance (or cosine similarity) for measuring semantic change scores of words (Laicher et al, 2021;Giulianelli et al, 2022;Cassotti et al, 2023), which (a) weights all dimensions equally, and (b) does not consider cross-dimensional correlations. However, our experimental results show that Euclidean distance is a suboptimal choice for this purpose and learning a Mahalanobis distance metric is more appropriate.…”

Efficient Continue Training of Temporal Language Model with Structural Information

“…Diachronic Language Change Diachronic language change can be mainly divided into semantic change Montanelli and Periti, 2023), morphological change (Hare and Elman, 1995;Ji et al, 2019;Giulianelli et al, 2022), and syntactic change (Kroch, 2001;Seretan, 2011;Bybee, 2017;Merrill et al, 2019). Previous work focused on discovering the words that have undergone diachronic change under the supervised settings (Kim et al, 2014;Basile and McGillivray, 2018;Tsakalidis et al, 2019;.…”