Introducing Traveling Word Pairs in Historical Semantic Change: A Case Study of Privacy Words in 18th and 19th Century English

Abstract

In recent years, Lexical semantic change detection (LSCD) has become a central task of NLP. Because most studies in LSCD only consider the semantic change of words in isolation, in this paper, we propose a new direction for the analysis of semantic shi昀琀s: traveling word pairs. First, we introduce shi昀琀 correlation to 昀椀nd pairs of words that semantically shi昀琀 together in a similar fashion. Second, we propose word relation shi昀琀 to analyze how the relationship between two words has changed over time. As a test case, we investigate the wordprivacy (and related words identi昀椀ed by a pre-existing dictionary), as an example of a word that has shi昀琀ed semantics historically and remains vibrantly explored as a concept in contemporary humanistic discourse. We report that the termprivacy in comparison shows relatively little change initially - with correlation analysis revealing more about how key terms surroundinpgrivacy have shi昀琀ed in tandem, and explore nuanced changes through word pair analysis, suggesting a shi昀琀 toward concreteness in particular.

Keywords

• eol>semantic change
• language models
• computational semantics

1. Introduction

There is a growing body of recent work in computational approaches to lexical semantic change detection in historical linguistics research2[ 7, 25, 22, 21, 33, 8, 4 ]. There is recognition that such research will eventually evolve from tracing semantic shi昀琀s in individual words or lexemes to larger groups of words [ 16, 13 ]. Eventually, the investigation of semantic historical shi昀琀s may uncover groups of words which shi昀琀 in correlation, or larger, tectonic shi昀琀s of meaning which we have not yet perceived. As McGillivray writes, ”Truly cutting-edge computational research in historical semantics should involve the development of innovative and impactful methods, which are build to answer questions relevant to humanists” 2[ 1 ].

In this paper, we explore two methods by investigating semantic shi昀琀 around the term privacy as a test case, using a pre-existing dictionary of words relating toprivacy. Privacy is selected, 昀椀rstly, because it has had shi昀琀ing semantics over time. The Oxford English Dictionary entry1 begins with word sense 1 — ”The state or condition of being alone, undisturbed, or free from public attention, as a matter of choice or right; seclusion; freedom from interference or intrusion” — and of the seven remaining word senses and sub-word senses, six are ”obsolete” and/or ”rare”. In the long nineteenth century, ”privacy as a concept underwent important evolutions in society, both in its literature, which increasingly explored privacy as a theme, and in law, with the foundation of the recognition of privacy as a legal right separate from copyright or defamation” [ 15 ]. Privacy as a theory is so manifold that Tavani 3[ 4 ] proposes a taxonomy of theories of privacy with four categories: nonintrusion, seclusion, control, and limitation. A consistent, uniform theory of privacy has proven elusive 3[ 7 ], and this historical uncertainty over the de昀椀nition of the lexical item and theories of privacy, as well as the voluminous discourse around privacy today (for instance, in the widely-used metric of di昀erential privacy, which captures the increased risk to one’s privacy incurred by participating in a database7[]), support privacy as a word worth exploring for semantic change, and perhaps related words shi昀琀ing semantics in some correlation.

To investigate the semantic shi昀琀 of privacy, we employ contextualized word embeddings derived from a historical English BERT model, MacBERTh 2[0]. We 昀椀rst generate token embeddings for all of our dictionary words from sentences from CLMET5][, a corpus of historical English. We then follow the standard procedure of calculating semantic change for all dictionary words via centroid distances between token embedding sets of two time slices. Finally, we propose two novel methods to further break down semantic shi昀琀s in the context ofprivacy: correlated semantic word shi昀琀s and shi昀琀ing word relations, with the latter being a reframing of the nearest neighbor approach.

We report that the term privacy shows relatively little semantic change by our models, a surprising result given the presumed manifold semantics of this term, and expand the investigation of privacy by detecting traveling word pairs; we report, for example, that revealing and protecting shi昀琀 in tandem semantically, which could be due to both of these terms shi昀琀ing from more 昀椀gurative to more concrete over time.

2. Related Work on Semantic Change

Lexical semantic change can be de昀椀ned as ”changes in ’sense,’ the concepts associated with expressions” [ 35 ]. This change in sense may either stem from the context in which a word is used over time (diachronic or temporal shi昀琀) or across di昀erent domains (synchronic or domain shi昀琀). This work focuses on the former. Change in sense may have di昀erent origins; amongst the di昀erent types are for example pejoration and amelioration (association of a term with a negative or positive meaning, respectively) or the narrowing/restriction of a term as opposed to the broadening/generalization of a concept 3[5].

The automatic detection of semantic shi昀琀s (also referred to as lexical semantic change detection or LSCD) has gained signi昀椀cant attention in NLP in recent years [33, 28, 9]. To detect diachronic change, a large text corpus for each time period of interest is assembled. There exist two types of computational approaches: type-based (word embeddings) and token-based (language models or contextualized embeddings). Using type-based methods usually implies 1https://www.oed.com/dictionary/privacy_n?tab=meaning_and_use training one word embedding model (e.g. fastText,2[ 3 ]) for each corpus slice, aligning the vector spaces (e.g. orthogonal Procrustes) and then comparing the word vectors across time slices. For the token-based approach, embeddings are extracted from a single language model (e.g. BERT, [ 6 ]) through the sentences the words of interest are used in. As an intermediate step, clustering may be used to identify di昀erent word usages before the comparison. The distances of the token vectors of one word for adjacent time slices then indicate change, for example through average pairwise distances (APD) or prototype/centroid distances1[ 8 ].

Disadvantages of the token-based method are therefore currently the computational scalability [10] and the fact that token clusters are formed by word forms rather than semantics1[ 8 ]. On the other hand, token-based methods enable the contextualization of each token through additional information from model pre-training and joint processing of the word and its sentence. This enables a deeper look into individual word senses and their shi昀琀s.

There are multiple ways in which change may be detected in vector space. Most commonly, words are analyzed in isolation: The distance of one word to itself in another time period or domain shows the stability of that word. The lower the distance, the higher is its stability, and in turn, less change is detected. In previous studies, distances were typically translated to a binary or graded scale for evaluation [28]. There are multiple evaluation datasets that capture this notion of word change through the creation of word usage graphs (WUG)1[ 2, 29, 30 ]. However, there are also other ways in which change could be analyzed. While some studies working with word embeddings have analyzed change in terms of changing word associations in the past [11, 38], the trend towards using language models has streamlined the task towards single word analysis. Kutuzov, Øvrelid, Szymanski, and Velldal 1[ 6 ] argue that ”most current studies stop a昀琀er stating the simple fact that a semantic shi昀琀 has occurred”. Similarly, Hengchen, Tahmasebi, Schlechtweg, and Dubossarsky [ 13 ] criticize that ”an emerging or evolving concept, almost by de昀椀nition, will not be constrained to a single word. Rather, methods will probably have to be adapted to study a cluster of words”.

The notion of words shi昀琀ing semantically in groups was hypothesized by a handful of precomputational linguists. In a 1931 treatise, Stern appealed to ”昀椀nd out to what degree [synonyms’] development runs in parallel lines and is conditioned by identical factors”32[]. In 1985, Lehrer expanded this hypothesis to di昀erent word relations: ”semantically related words are more likely to undergo parallel semantic changes because of their semantic relationships. [...] If one word changes meaning, it will drag along other words in the domain”1[9]. Computation now allows us to explore these notions in greater depth.

In this paper, we propose to pick up on these previous intuitions on change detection, that is analyzing the relationship of word pairs across time on a larger scale, using contextualized word embeddings.

3. Experiments

3.1. Corpus and Dictionary

We selected The Corpus of Late Modern English Texts, version 3.1, a genre- and time-balanced text corpus of late modern English (1710-1920) of about 34M word tokens5[]. Originally, the corpus consists of three balanced slices of 10-12M tokens each. To create more subcorpora, we 1780-1815, 1815-1850, 1850-1885, 1885-1920. The splits had an impact on the overall genre balance, although the proportions of genres for all slices between 1745 and 1920 stayed mostly consistent (昀椀ction being about 45-65% of each sub-corpus, while the other genres have a share of 5-25% each). The 1710-1745 slice is a bit of an outlier in genre mix to keep in mind in the experiments below, being about 45% letters.

For a list of words relating to privacy, we selected Vasalou et al.’s Privacy Dictionary3[ 7 ], a dictionary of 616 words and phrases for automated content analysis on privacy-related texts. Some bene昀椀ts of this dictionary include its creation through a rigorous methodology, as well as its application to studies in a variety of 昀椀elds, e.g. [ 1, 2, 36, 3 ], including humanistic research

3.2. General Approach

To obtain the contextualized word embeddings, we turn to the MacBERTh model implemented by Manjavacas and Fonteyn [ 20 ], which is entirely pre-trained on historical English. Their training corpus consists of texts of varying genres (literary works, articles, etc.) the size of about 3.9B tokens, covering a time span from 1473 to 1950.

As proposed by Hengchen, Tahmasebi, Schlechtweg, and Dubossarsky [ 13 ], to study a cluster of concepts, keywords may be chosen either manually or in a data-driven way. To construct a list of words around the concept ofprivacy, we followed both suggestions. We turned to 1) the Privacy Dictionary created by Vasalou, Gill, Mazanderani, Papoutsi, and Joinson3[ 7 ], selecting only the unigrams and discarding phrases, and 2) we expanded this word list with 70 nearest neighbors ofprivacy from fastText embeddings trained on our 昀椀rst and last corpus slices each, to create a new dictionary of 454 words in total.

We adopt the method of using large language models (in our case MacBERTh) for detecting semantic change, i.e. retrieving token embeddings for all of our words of interest from the corpus, resulting in one embedding matrix for each of the six time slices. To obtain a stable representation, we only consider words in their respective time slice when they have more than 10 occurrences in that slice. For computational scaling purposes, we sample 100 occurrences for each word per slice.

For the embedding extraction, we use the average of the last 4 layers (9-12) in the MacBERTh model. To determine the degree of change for one word for adjacent time slices, we calculate the cosine distance of the mean of both matrices (centroid distance) with a sliding window over all slices: where , +1 1 , ..., 5 per word.

= cos ( 1

=1 ∑ wti ,

1 +1 j +1 =1 ∑ wti+1 )

j are individual word vectors of two adjacent time slices, an d represent the counts of word vectors in the two time slices. The results are 昀椀ve distance measurements

We begin our analysis with the standard procedure of binary/graded word level semantic shi昀琀 detection. First, we calculate mean and standard deviations for all 昀椀ve points in time

privacy in comparison to two other, highly unstable words. The grey dotted line indicates mean cosine distances across all 454 words. The grey area indicates the standard deviation. across all words as our point of reference:

Mean = 1 =1 ∑ 1 √ =1 Standard Deviation = ∑( − Mean )2 where represents the number of words in our dictionary (454). We then extract the words that show the highest distances across all measurements, additionally toprivacy.

Our top 昀椀ve most unstable words are interfering, amendment, suppressing, licence, and sensitive. We 昀椀nd that privacy actually shows a stabilizing trend in our corpus across the 18th and 19th century (see Figure 1). In the following experiments, we will take a closer look especially at some of these more unstable words in addition to the focus word of our dictionarpyr,ivacy.

3.3. Shi昀琀 Correlation

For our 昀椀rst novel approach, we take up Kutuzov et al.’s proposal to identify ”groups of words that shi昀琀 together in correlated ways” [ 16 ]. Correlating shi昀琀s may reveal what we dub traveling word pairs: pairs of words which shi昀琀 in meaning over time with a similar correlative pattern. We see computational change correlation as an opportunity to dive deeper into why meanings may have changed.

For this experiment, we investigate each pair of words in our list of 454, by calculating shi昀琀 detecting correlations from the stability scores. This reveals which word pairs show a similar shi昀琀 in stability scores across the 18th and 19th century. To calculate correlations, we use Pearson’s .2 The top 10 pairs are reported in Table1.

If we look fortraveling pairs, we would need to identify words which, 昀椀rst of all, semantically shi昀琀 signi昀椀cantly; otherwise, two words with no shi昀琀 over time would correlate. We de昀椀ne high shi昀琀 as a cosine distance of at least 0.02 (our highest mean plus standard deviation) in our list of measurements. 2One reviewer brought to our attention that Pearson’s might not be the ideal measurement for correlation in this case, as it cannot detect similarities between shi昀琀ed time-series, and advised to use dynamic time-warping or similar instead. We would like to thank them and will certainly take up their suggestion for future work on this matter.

Among our most correlated words that also show high semantic shi昀琀, we 昀椀nd disclosing, protecting, and revealing (see Figure 2). The three words seem to undergo a signi昀椀cant semantic change shortly past 1800. By looking at the sentences in each of the corpus slices, we 昀椀nd that revealing and disclosing show a similar trend. Both go through a shi昀琀 from the abstract to the concrete, e.g. ”revealing meanwhile a 昀氀annel shirt” 3 or ”disclosing spaces of faint yet clearest blue”4 as opposed to ”revealing his a昀airs to him” 5 and ”disclosing my Love and Esteem”6.

That same sentiment is mirrored inprotecting: While in earlier text passages, it is values or acts that need protection (”protecting and securing the trade”7,) while the later texts mostly use protecting in a physical sense (”protecting that building from the fury of the populace8”, ”stretch out a protecting arm”.9)

3.4. Shi昀琀ing Word Relations

Previous studies on LSCD working with word embeddings have analyzed nearest neighbors across time, either for qualitative interpretation or evaluation1[ 1 ] or to calculate another dimension of meaning shi昀琀 (word level change in k nearest neighbors [38]). In contrast, we suggest the close tracking of a multitude of neighbor relations as another way to discover and quantitatively interpret change, which may create opportunities to identify factors that condition similar development, as Stern hypothesized 3[ 2 ]. We thus propose a second method that further explores the idea of using nearest neighbors: to investigate which words in a prede昀椀ned dictionary become semantically closer or more distant to a keyword over time, to allow a richer interpretation a single word’s semantic history over time.

In these experiments, we selected two keywords: privacy, as this is the primary topic of our investigation and dictionary, andrevealing, because this word was salient in the results 3George Gissing, 1891 4Mary St Leger Kingsley (Lucas Malet), 1901 5Henry Fielding, 1751 6Gilbert Langley, 1745 7Samuel Johnson, 1740 8Henry Hunt, 1820 9Percy James Brebner, 1910 of our experiment above. We then calculate cosine distances between each keyword and each dictionary word, in each time slice, and detect which of these pairs display the highest divergence across time. Note that in this experiment, we trace the shi昀琀 of word relations over time as opposed to contrasting single words. This is why in Figure1 and 2, the cosine distances are comparatively small (~0.01), because we compare the word to itself across di昀erent time periods. In Figure3 and 4, we compare the distances of two di昀erent words over time, which means cosines distances are generally higher (~0.2).

The results for keywordrevealing and selected words from the dictionary list are in Figure 3, which displays the progression of cosine distances betweenrevealing and dictionary words. Amongst the words displaying the most semantic divergence ispropriety, which we interpret as evidence that revealing and propriety were semantically distant in the 1700’s and gradually became more close by the 1920’s. Our method thus yields any number of comparisons in shi昀琀 to a keyword, with a calculated ”divergence” score. As a brief observation orfevealing and propriety becoming semantically closer over time, one hypothesis is that both of these words shi昀琀ed from more abstract to more concrete meanings (similar to the example of revealing and disclosing in our method above). As another observation, we report a stable relation between synonyms revealing and disclosing over time, whereas the fuzzy antonyms revealing and protecting become closer. This shows that even though stability scores may correlate, the distances between those words might not change in the same way. Further observations on these speci昀椀c words would require more dedicated studies, but our method will provide such studies with new information about comparison of semantic shi昀琀.

We then apply our method to keywordprivacy (see Figure 4). Although privacy was the starting point of our use case exploration, our method actually reveals only subtle information for the relation of our dictionary words to privacy, apsrivacy remains fairly stable in semantics, by our metrics, throughout the time periods (as shown in the General Approach). This alone is a tantalizing result that warrants further investigation, asprivacy is widely considered to be extremely rich semantically and subject to a complex evolving history. A consistent, uniform theory of privacy has proven elusive [37], and taxonomies exist which group privacy by theory, e.g. Tavani [34], who suggests nonintrusion (”being let alone”), seclusion (”one’s being secluded from others”), control (”one has privacy if and only if one has control over information about oneself”), and limitation (”one has privacy when information about oneself is limited or restricted in certain contexts”). Returning to our experiments, while the changes detected are very subtle, we observe that two of the most semantically distanced words fromprivacy — accessible and disclosing — refer to concepts of access, whereas, for example,secluded and locked draw a bit closer to the meaning ofprivacy.

Additionally, we suggest that we may be able to discover, through average changes in word relations, whether a term may have undergone semanticbroadening or narrowing. If, on average, other concepts have smaller distances to a target word, we might discoverbroadening. Otherwise, if concepts grow more distant on average, we might discovernarrowing. This can be done either by the mean changes (as indicated in all 昀椀gures so far) or majority vote across all words of interest, or the topk most volatile dictionary words. Through the latter for example, among the top ten most volatile dictionary words relations opfrivacy, we 昀椀nd that all but one word have become more close (see Table2), which could imply a generalization of the concept. On the other hand, the concept offreedom in our corpus for example has equal amounts of terms that grow closer or more distant. The same trends are mirrored in the mean distances. Through a targeted selection of sentiment keywords as dictionary words, this method could also be used to identify changes in connotation. Finally, both approaches could potentially be combined to discover correlated neighbor shi昀琀s.

4. Discussion and Future Work

We have presented two experimental approaches to expanding the investigation of historical semantic change by not only limiting the analysis to words in isolation, but instead taking changing word relationships into account. The wordprivacy has undergone relatively little change, but rather key issues surrounding the concept have. Through an analysis of stability correlations, we for example were able to show the tie betweenrevealing/disclosing and protecting, as these concepts shi昀琀 to a more concrete representation in our corpus. Through a closer analysis of the word relationships ofrevealing, we found that a shi昀琀 towards propriety aligns with this hypothesis. Our interpretation also aligns with previous 昀椀ndings concerning concreteness shi昀琀s; Hills and Adelmann provide evidence that concrete words were gradually used more o昀琀en over time [ 14 ], and Sne昀樀ella et al. add that distinct word types also undergo a shi昀琀 towards concreteness [ 31 ]. While both of these examined word types and tokens at large, this work focused on a select group of words aroundprivacy. Future work could explore this shi昀琀 in more depth, for example using abstract and concrete seed words [ 31 ] for the neighbor shi昀琀 analysis to get a larger perspective, or employ additional methods to acquire concreteness ratings of the speci昀椀c word types from the corpora.

More work will be necessary to re昀椀ne and expand on the proposed method. Most importantly, there are several limitations to our work that are also ongoing issues in the 昀椀eld of LSCD. The 昀椀rst issue is orthographic variation. Especially at the start of the 18th century, we face a high degree of spelling variation that we did not normalize, which limits the recall of embedding extraction. For future work, orthographic normalization as well as lemmatization should be considered as pre-processing steps.

The results of our study are highly dependent on the choice of corpus. Even though our corpus has balanced slices in terms of genre, the observed variation can still be dependent on a speci昀椀c context of a particular slice. This means that rather than 昀椀nding semantic shi昀琀s, we 椀昀nd unintended domain shi昀琀s because words were used in speci昀椀c contexts in certain time slices. This is a well-known underlying issue of using the data-driven, distributional approach that is subject of a larger discussion about what semantic shi昀琀s actually are; per Kutuzov et al., ”If one does not employ external data sources [...], there is no reliable way to discern ‘semantic changes’ from ‘di昀erences in the underlying textual data’: they are simply the same thing” [ 17 ].

Additionally, by selecting our dictionary words, we limit our result space and another list of words would have produced di昀erent results. As the method is computationally expensive, this pre-selection is still necessary at this point in time. To study a di昀erent concept where a curated list of words is not available, researchers could either turn to resources such as knowledge graphs, where related words for most concepts can be retrieved up to a desired path depth (e.g. WordNet or FrameNet), or select words based on distributional approaches such as fastText entirely.

A 昀椀nal limitation concerns word frequency and polysemy. Both highly polysemous words and words with instable frequency produce a higher semantic change score with this method [28]. We set a minimum and maximum frequency limit for our experiments. However high 氀昀uctuation is still possible, especially due to orthographic variation as mentioned before, which we did not control for.

At the moment, we provided suggestions for further insights into semantic change detection and presented a demonstration of that idea. In future work, evaluation data will be necessary to support the results and methods discussed so far. We propose to construct a new type of evaluation dataset based on the idea of changing word pairs. Current evaluation data consists of words with assigned stability scores 2[ 9, 12 ]. Instead, new datasets may be created by tracing the relationship of e.g. synonyms or associated words through a similar methodology. By grading sentence pairs that contain synonyms based on their similarity instead of the same word for example, the degree of synonymy may be determined across di昀erent time periods. Existing word similarity datasets such as SimLex333 or WordSim353 could provide the basis for such word pair selection. Such word pair ratings could also be established to evaluate concreteness shi昀琀s for example.

On the historical semantics ofprivacy speci昀椀cally, we note that our 昀椀ndings do not align with an analog study on the historical semantics ofprivacy in American historical jurisprudence by Prestidge, who, based on close reading, reportedprivacy ”shi昀琀ing in meaning from the literal to the 昀椀gurative” [ 26]. In our experiments, privacy remained quite semantically stable, and we present evidence that some words relating toprivacy, to the contrary, shi昀琀ed from 昀椀gurative to literal. These 昀椀ndings are certainly worth exploring further.

We hope that the ideas on automatic semantic change analysis presented here can provide a step towards the pairwise analysis of shi昀琀s in word meaning as a new direction within the 椀昀eld.

A. Prestidge. “Semantic Change in Supreme Context: Semantics in the Privacy Line and Originalist Interpretation”. In:Brigham Young University Prelaw Review 24 (2019), pp. 117–137.

References

[1] A. Caliskan Islam , J. Walsh , and R. Greenstadt . “ Privacy detective: Detecting private information and collective privacy behavior in a large social network” . InP: roceedings of the 13th Workshop on Privacy in the Electronic Society. Scottsdale Arizona USA , 2014 , pp. 35 - 46 .

[2] F. Casillo , V. Deufemia , and C. Gravino . “ Detecting privacy requirements from User Stories with NLP transfer learning models” . In:Information and So昀琀ware Technology 146 ( 2022 ), pp. 1 - 12 . doi: 10 .1016/j.infsof. 2022 . 106853 .

D. D'Acunto , S. Volo , and R. Filieri . “ “Most Americans like their privacy.” Exploring privacy concerns through US guests' reviews” . In: International Journal of Contemporary Hospitality Management 33.8 ( 2021 ).

[4] S. De Pascale . Meaning change in token space: a token-based computational approach to diachronic prototype semantics . Kick-o昀 Conference Day of the RJ Program . 2022 .

H. De Smet , S. Flach , H. - J. Diller , and J. Tyrkkö . The Corpus of Late Modern English Texts, version 3 .1. Online . 2011 .

[6] J. Devlin , M.- W. Chang , K. Lee , and K. Toutanova . “BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding” . In:Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies , Volume 1 (Long and Short Papers). Minneapolis, Minnesota, 2019 , pp. 4171 - 4186 .

[7] C. Dwork . “ Di昀erential privacy” . In: International colloquium on automata, languages, and programming . Springer. Venice, Italy, 2006 , pp. 1 - 12 .

[8] K. Franco , M. Montes , and K. Heylen . “ Deconstructing destruction: A Cognitive Linguistics perspective on a computational analysis of diachronic change” . InP:roceedings of the 3rd Workshop on Computational Approaches to Historical Language Change . Dublin, Ireland, 2022 , pp. 23 - 32 .

[3] [5] [9] [10] [11] M. Giulianelli , M. Del Tredici , and R. Fernández . “ Analysing Lexical Semantic Change with Contextualised Word Representations” . In:Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics. Online , 2020 , pp. 3960 - 3973 . doi: 10 .18 653/v1/ 2020 .acl-main. 365 .

M. Gruppi , P.-Y. Chen, and S. Adali . “ Fake it till you make it: Self-supervised semantic shi昀琀s for monolingual word embedding tasks” . In: Proceedings of the 35th AAAI Conference on Arti昀椀cial Intelligence . Vol. 35 . Online , 2021 , pp. 12893 - 12901 .

W. L. Hamilton , J. Leskovec , and D. Jurafsky . “ Diachronic Word Embeddings Reveal Statistical Laws of Semantic Change” . In: Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) . Berlin, Germany, 2016 , pp. 1489 - 1501 . doi: 10 .18653/v1/ P16 -1141.

[12] A. Hätty , D. Schlechtweg , and S. Schulte im Walde. “SURel: A Gold Standard for Incorporating Meaning Shi昀琀s into Term Extraction” . In: Proceedings of *SEM 2019 . Minneapolis, Minnesota: Association for Computational Linguistics, 2019 , pp. 1 - 8 . doi1: 0 .18653/v1 / S19 -1001.

[13] S. Hengchen , N. Tahmasebi , D. Schlechtweg , and H. Dubossarsky . “ Challenges for computational lexical semantic change” . In: Computational approaches to semantic change . Berlin: Language Science Press, 2021 , pp. 341 - 372 .

[14] T. T. Hills and J. S. Adelman . “ Recent evolution of learnability in American English from 1800 to 2000” . In: Cognition 143 ( 2015 ), pp. 87 - 92 .

[15] E. Ketzan , J. Edmond , and C. Vogel . “ Need a Good Book about Privacy? Evaluating Dictionary-Based Corpus Query for Detecting the Topic of Privacy in Literary Texts” . In: Journal of Computational Literary Studies (in press).

[16] A. Kutuzov , L. Øvrelid , T. Szymanski , and E. Velldal. “ Diachronic word embeddings and semantic shi昀琀s: a survey” . In: Proceedings of the 27th International Conference on Computational Linguistics. Santa Fe , New Mexico, USA, 2018 , pp. 1384 - 1397 .

[17] A. Kutuzov , E. Velldal, and L. Øvrelid . “ Contextualized embeddings for semantic change detection: Lessons learned” . In: Northern European Journal of Language Technology 8 ( 2022 ), pp. 1 - 14 . doi: 10 .3384/nejlt.2000- 1533 . 2022 . 3478 .

[18] S. Laicher , S. Kurtyigit , D. Schlechtweg , J. Kuhn , and S. Schulte im Walde. “Explaining and Improving BERT Performance on Lexical Semantic Change Detection” . In:Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Student Research Workshop . Online, 2021 , pp. 192 - 202 . doi: 10 .18653/v1/202 1.eacl-srw. 25 .

[19] A. Lehrer . “ The in昀氀uence of semantic 昀椀elds on semantic change” . In: Historical Semantics - Historical Word-Formation . Ed. by J. Fisiak . Berlin, New York, Amsterdam: Mouton de Gruyter, 1985 , pp. 283 - 296 .

[20] E. Manjavacas and L. Fonteyn . “ Adapting vs. Pre-training Language Models for Historical Languages” . In: Journal of Data Mining & Digital Humanities Nlp4dh ( 2022 ). doi: 10 .462 98/jdmdh.9152.

[21] B. McGillivray. “ Computational methods for semantic analysis of historical texts” . In: Routledge International Handbook of Research Methods in Digital Humanities. London: Routledge , 2020 , pp. 261 - 274 .

[22] B. McGillivray , S. Hengchen , V. Lähteenoja , M. Palma , and A. Vatri . “ A computational approach to lexical polysemy in Ancient Greek” . In:Digital Scholarship in the Humanities 34.4 ( 2019 ), pp. 893 - 907 .

[23] T. Mikolov , K. Chen , G. Corrado, and J. Dean . “ E昀케cient estimation of word representations in vector space” . In:arXiv preprint arXiv:1301.3781 ( 2013 ).

[24] G. R. Milne , B. Kaplan , K. L. Walker , and L. Zacharias . “ Connecting with the future: The role of science 昀椀ction movies in helping consumers understand privacy-technology trade-o昀s” . In: Journal of Consumer A昀airs 55.3 ( 2021 ), pp. 737 - 762 .

[25] V. Perrone , M. Palma , S. Hengchen , A. Vatri , J. Q. Smith , and B. McGillivray. “GASC: Genre-Aware Semantic Change for Ancient Greek” . In: Proceedings of the 1st International Workshop on Computational Approaches to Historical Language Change. Florence, Italy , 2019 , pp. 56 - 66 .

M. A. Rodda , M. S. Senaldi , and A. Lenci . “ Panta rei: Tracking semantic change with distributional semantics in ancient Greek” . In:IJCoL. Italian Journal of Computational Linguistics 3 . 3 - 1 ( 2017 ), pp. 11 - 24 .

“SemEval-2020 Task 1: Unsupervised Lexical Semantic Change Detection” . In:Proceedings of the 14th Workshop on Semantic Evaluation. Barcelona (online) , 2020 , pp. 1 - 23 .

D. Schlechtweg , S. Schulte im Walde, and S. Eckmann . “ Diachronic Usage Relatedness (DURel): A Framework for the Annotation of Lexical Semantic Change” . In:Proceedings of NAACL-HLT. New Orleans, Louisiana , 2018 , pp. 169 - 174 .

“DWUG: A large Resource of Diachronic Word Usage Graphs in Four Languages” . In: Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing.

Online and Punta Cana , Dominican Republic: Association for Computational Linguistics, 2021 , pp. 7079 - 7091 . doi: 10 .18653/v1/ 2021 .emnlp-main. 567 .

A. Vasalou , A. J. Gill , F. Mazanderani , C. Papoutsi , and A. Joinson . “ Privacy dictionary: A new resource for the automated content analysis of privacy” . InJ:ournal of the American Society for Information Science and Technology 62.11 ( 2011 ), pp. 2095 - 2105 .

M. Wevers and M. Koolen . “Digital begri昀sgeschichte: Tracing semantic change using word embeddings” . In:Historical Methods: A Journal of Quantitative and Interdisciplinary History 53.4 ( 2020 ).

[31] B. Sne昀樀ella , M. Généreux, and V. Kuperman . “ Historical evolution of concrete and abstract language revisited” . In: Behavior Research Methods 51 ( 2019 ).

[32] G. Stern . Meaning and change of meaning; with special reference to the English language . Bloomington: Indiana University Press, 1964 .

N. Tahmasebi , L. Borin , A. Jatowt , Y. Xu , and S. Hengchen . Computational approaches to semantic change . Berlin: Language Science Press, 2021 . doi: 10 .5281/zenodo.5040241.

H. T. Tavani . “ Philosophical theories of privacy: Implications for an adequate online privacy policy” . In: Metaphilosophy 38.1 ( 2007 ), pp. 1 - 22 .

[35] E. C. Traugott . Semantic change. Oxford research encyclopedia of linguistics . 2017 .

[36] K. A. Vakeel , S. Das , G. J. Udo , and K. Bagchi . “ Do security and privacy policies in B2B and B2C e-commerce di昀er? A comparative study using content analysis” . In: Behaviour & Information Technology 36.4 ( 2017 ), pp. 390 - 403 .