Computational studies of semantic change are often wide in scope, aiming to capture and quantify semantic change in language at large in a data-driven, 'hands-of' way. Case-driven, corpus-linguistic studies of semantic change, by contrast, generally aim to tackle questions about the development of specific linguistic phenomena. Due to its narrower scope, case-driven research is more restricted in terms of the data it may employ, and at the same time it requires a more fine-grained description of the targeted linguistic developments. As a result, case-driven studies face particular methodological challenges that are not at play in more wide-scoped approaches. The aim of this paper is to set out a 'hands-of' computational procedure to study specific cases of semantic change. The case we address is the development of the phrasal expression to death from a literal, resultative phrase (e.g. he was beaten to death) into an intensifier (e.g. We were just pleased to death to see her). We deploy hierarchical clustering algorithms over distributed meaning representations in order to capture the evolution of the semantic space of verbs that collocate with to death. We then describe the arising diachronic processes by means of monotonic efects, providing a more accurate picture than customary linear regression models. The methodology we outline may help tackle some common challenges in the use of vector representations to study similar cases of semantic change. We end the discussion by pinpointing (remaining) challenges that case-driven research may encounter.
Over the past decade, computational approaches to semantic change have experienced a surge in popularity. This is largely due to the rise of an increasingly powerful body of models that aim to approximate the meaning of words over time by encoding their linguistic context (or ‘distributional properties’) into (diachronic) word embeddings [see, among many others: 47, 19, 35, 18, 1, 44, 46, 29, 26, 13, 51, 11, 48, 5, 50]. A characteristic of many of these studies is that their research questions are very wide in scope: their aim is not to address questions about any specific word or construction, but rather to capture and quantify some aspect of semantic change at large in a data-driven way. As such, these studies tend to approach semantic change in bulk, with sample sizes ranging from hundreds [e.g. 35, 13, 48] to thousands of linguistic items [e.g. 18], and with specific examples of semantic change predominantly serving as (straightforward) illustrations of a more general pattern or trend.
Yet, vector-based models are also used in more narrow-scoped, case-driven research. In this
type of research, which is perhaps most common in (corpus) linguistics, the aim is to approach
a specific case study of semantic change in a largely automated, data-driven manner. The
motivation for doing so is that introspective data annotation (which is not only labour-intensive,
but potentially problematically subjective [
Despite the fact that the two approaches have an obvious common ground, case-driven investigations are clearly distinct from wide-scope computational studies in a number of ways. Most importantly, case-driven research generally emerges from a desire to tackle questions about the development of specific linguistic phenomenon (often during a specific time window). Consequently, compared to the wide-scope computational research into semantic change, casedriven research is relatively inflexible in terms of the data it may employ to attain its goals. Furthermore, the very reason why the researcher is compelled to undertake case-driven research is that the specific phenomenon under scrutiny constitutes a complex challenge. As such, while it is not uncommon for case-driven studies to use computational models and metrics proposed in wide-scoped studies on semantic change, the specific cases they scrutinize may present methodological challenges that are either not at play, or glossed over in, the studies they draw from.
The aim of the present contribution is to tackle a case-driven study by means of
computational methods. In doing so, our work ties in with earlier explorative work aimed at pinpointing
where challenges may lie for case-driven research [e.g. 49]. The specific case of semantic change
we address is the historical development of the phrasal expression to death from a literal,
resultative phrase (e.g. he was beaten to death) into an intensifier (e.g. We were bored/pleased
to death) [
To analyse the development of to death with minimal manual interference, we suggest a procedure consisting of the following steps: 1. Surveying work on the linguistic construction (and related cases) under scrutiny to (i) delineate a time window, and (ii) formulate hypotheses or expectations that can be verified by means of computational methods (Section 2); 2. Compiling (and curating) a sufficiently large diachronic corpus collection (Section 3), from which examples of the construction can be sampled (Section 3.1); 3. Computing and evaluating distributed meaning representations (Section 3.2); 4. Conducting a diachronic cluster analysis, in which we optimize the number of clusters across time for silhouette score in order to trace changes in to death’s contextual distribution (Section 4.1); 5. Conducting a sentiment analysis to capture to death’s decreasing negativity (Section 4.2); 6. Assessing the output of the statistical model against the formulated expectations (Section 5).
After describing the procedure and results, we highlight and discuss the following remaining pitfalls (Section 6): 1. Because case-driven research aims to examine a fixed (set of) linguistic construction(s) in a specific time window, researchers may run into issues of data sparsity and balance that may be difficult to circumvent. 2. The extent to which a reliable, completely automated, ‘hands-of’ approach is possible and extendable to new cases not yet analysed remains an open question. While we are optimistic about incorporating computational models into the study of case-driven semantic change, manual interference may still be desirable or even required.
The methods adopted in the present study are similar to those used in large-scale computational
studies of semantic change. Previous work has suggested various ways of improving the models
that generate (diachronic) word embeddings [e.g. 46, 44], determining (predictive) laws of
(lexical) semantic change at large [e.g. 19, 14], and developing statistical measures that help
detect diferent types of semantic change (e.g. specification vs. broadening; cultural change
vs. linguistic change) in a data-driven manner [e.g. 47, 35, 18, 11, 13, 48, 16]. In other work,
computational models are applied to map changes in specific word classes (e.g. intensifiers;
[
In terms of its focus, aims, scope and granularity, this study is reminiscent of research
in corpus linguistics and construction grammar, where a single case of linguistic change is
considered. The development of to death from a phrase that expresses the result of an action
(e.g. He was beaten/stabbed/shot to death) to an intensifying or ‘amplifying’ expression (e.g.
We were thrilled/pleased/shocked to death to see you; [
STAGE 1 Initially, to death functioned as adverbial complement of verbs expressing physical harm, which may result in death (e.g. beat, bleed, burn, etc.).
STAGE 2 Over the course of the 16th and 17th century, to death sporadically started occurring in
contexts where a literal, death-resulting reading is ruled out (e.g. That book bored me to
death). It was not until the 18th century, however, that to death was frequently used in
such non-literal, intensifying cases [
STAGE 3 Despite its persistent preferences for negative situations, to death started to expand
further [
The expansion process spawned by the grammaticalization of to death would seem to lend
itself well to computational analysis. A template for the general research design can be found in
the work of Perek [
Crucially, Perek demonstrates that, from a linguistic perspective, it is important to approach processes of host class expansion in a way that distinguishes changes in lexical diversity (measured by the number of unique lexical items that occur in a construction) from semantic diversity (measured by the semantic similarity between those lexical items). This is also relevant for the study of to death, because changes in lexical diversity alone may not be indicative of linguistic change, but of cultural change. It may be the case, for instance, that diferent modes of execution have become prevalent or obsolete, or that the specificity and lexical diversity with which causes of death are described may increase or decrease as the topic becomes more or less taboo. In these scenarios, the set of lexical items to death collocates with may indeed shrink or expand, but the semantics of the phrase do remain stable. At the same time, such cultural change may happen alongside the grammaticalization of to death into an intensifier. Thus, the reality of case-driven research may be that the distinction between cultural and linguistic change is not a matter of ’either/or’ [18], but of ’and’.
The distributed meaning representations that are fed into the clustering algorithm by Perek
[
For the purposes of the present study, we gathered a collection of diachronic English corpora, spanning the period from 1550 to 1949. These corpora include Early English Books Online (EEBO), the Corpus of Late Modern English Texts (version 3.1; CLMET3.1), the Evans Early American Imprints Collection (EVANS), Eighteenth Century Collections Online (ECCO), the Corpus of Historical American English (COHA), and the Hansard corpus (Hansard). In terms of text types, these corpora are varied, covering an array of literary works, religious and legal text and news reports. The sole exception is Hansard, which ofers transcriptions of British parliamentary debates (starting in 1800).
All corpora were submitted to the following pre-processing pipeline. First, we applied a
language identification module in order to sort out foreign text. We relied on two language
identification modules – Google’s Compact Language Identifier (v3) 1 and FastText Language
1Our code repository is accessible through the following url: https://github.com/google/cld3/releases/tag/
3.0.13.
Identification system [
Second, we tokenized and sentence-tokenized the remaining text using the Punkt tokenizers
provided by the NLTK package [
The resulting patchwork corpus consists of a total of 3.9B tokens, which we utilized in various ways in subsequent steps of the research process.
The attestations of to death were retrieved from the corpus collection (excepting the specialized Hansard corpus). As is common in linguistic research, the data was divided into fixed-width bins. Each bin represents a 50-year period, which results in a total of 8 bins. As not all corpora in the collection are balanced in terms of the amount of text a single author may contribute, we applied an additional sampling step to ensure that no author dominated more than 25% of the instances in a particular bin.
The total number of instances retrieved from each corpus per bin is listed in Table 1. In the bin covering the period between 1700 and 1749, the total corpus size (and hence, the token frequency of to death) was substantially lower than for other bins. To ensure that any observed diferences in the number of verb types that collocate with to death across bins is not afected by large diferences in sample size, we decided to cap the maximum number of tokens sampled per bin at 800.
After removing any duplicates, we identified the verb that collocates with each instance of to death by relying on part-of-speech tags. Each instance of to death was assumed to collocate with the verb in closest proximity (using a window of 15 words). In a number of cases, the tagger failed to find a collocate verb. These cases included instances where the copula be was used in combination with an adjective (e.g. be frozen/sick to death), which were subsequently corrected and included in the dataset. Cases where to death functioned as a prepositional modifier of a noun (e.g. on her way to death), fixed expressions (e.g. from birth to death, be nigh to death), and cases where the verb was illegible (e.g. And when my mother euen before my sighte, Was (-) to death; 1550, EEBO) were discarded. In total, 109 examples were discarded.
In order to capture semantic similarity between to death’s collocate verbs across time, we rely
on distributed meaning representations computed by the word2vec algorithm [
We trained distributed representations with a size of 200 using the gensim library [
In order to validate the resulting embedding space, we ran a number of semantic similarity
benchmarks, which allow us to contextualize the quality of our embeddings within the
stateof-the-art. The employed benchmark datasets comprise of sets of Present-day English word
pairs, each of which has been manually assigned a similarity score. The evaluation proceeds
by correlating these human judgments with the cosine similarities between the corresponding
vector representations, using the Spearman correlation coefficient. 2 We compared our
embedding space with (i) 200 dimensions Glove vectors [
As Table 2 shows, our embedding space generates scores comparable to the Glove space, while lying behind those generated by the word2vec space. Considering that our embedding space is trained on a smaller dataset and covers a large period of historical English, we take these results to validate the semantic similarity properties of the inferred word representations. For a sanity check, Table 3 shows the 20 nearest neighbours of a selection of verbs from our dataset of to death collocates based on cosine distance.
2While it is obviously not ideal to evaluate our model with respect to a Present-day English reference point, no human similarity judgements of this scale are available for historical English. In order to conduct at least some sort of sanity check, we used the of-the-shelf Present-day English spaces.
3The embeddings are available through the following url: https://nlp.stanford.edu/projects/glove/.
4We use the software package word-embedding-benchmarks [
whip strangle (0.59) cudgel (0.69) delude (0.73) knife (0.59) bludgeon (0.66) flatter (0.63) bleed (0.58) lash (0.66) perplex (0.61) slash (0.58) kick (0.59) terrify (0.60) bang (0.56) cuf (0.57) frighten (0.60) kill (0.55) spur (0.57) tickle (0.58) poison (0.55) lfog (0.56) harass (0.54) bite (0.55) bang (0.55) tire (0.54) cudgel (0.54) goad (0.55) annoy (0.52) prick (0.54) scourge (0.54) vex (0.51) mental verbs amuse scare
vex frighten (0.78) aflict (0.72) terrify (0.73) perplex (0.72) startle (0.67) harass (0.71) worry (0.55) annoy (0.69) drive (0.54) oppress (0.69)
sweep (0.52) fret (0.67) delude (0.51) grieve (0.64) astonish (0.51) terrify (0.61) annoy (0.50) pester (0.60) amuse (0.50) worry (0.58) A basic way of quantifying the host class expansion of to death is by examining the change in diversity in the set of attested collocate verbs over time. One such index of diversity is given by type frequency – shown in the last row of Table 1. However, while such diversification is potentially indicative of host class expansion, changes in type frequencies (or lexical diversity) need not indicate that to death has indeed undergone semantic change; as argued in Section 2, they may equally be indicative of cultural change. To probe into the host class expansion of to death, Section 4.1 operationalizes the process as a change in the structure of the semantic space that the collocate verbs of to death occupy. We rely on hierarchical cluster analysis over distributed meaning representation in order to not only incorporate a notion of lexical diversity into the analysis but, crucially, also take ‘semantic diversity’ into account.
As explained in Section 2, the host class expansion of to death also involved increased cooccurrence with verbs with progressively more positive connotations. In order to capture this process, Section 4.2 devises a way to quantify the average polarity of verbs over time using word embeddings, and statistically describe any existing ‘positivization’ process.
At any given period, we inspect the semantic space delineated by the distribution of attested
verbs using a hierarchical cluster analysis. A known problem with automated cluster analysis
of semantic spaces is that the induced semantic clusters are not always easy to interpret. As a
result, their application in subsequent steps of the research workflow may require manual
finetuning and post-filtering [
As to death develops new, non-literal meanings, we expect the semantic space defined by the
verbs appearing in this construction to expand, with existing clusters of collocates becoming
denser and new clusters representing novel semantic fields starting to form. The silhouette score
[
For a word wi assigned to cluster Ci, the silhouette score decomposes into the quantities a(wi) – shown in Equation 1 – and b(wi) – shown in Equation 2. By measuring the average intra-cluster distance between a word and all other words in the same cluster, a(wi) captures the tightness of the clusters induced by a clustering algorithm. In contrast, b(wi) measures the distance to the nearest point in a diferent cluster. The dataset-level aggregated b(wi), thus, captures the overall separation between clusters.
a(wi) =
1 ∑ |Ci| − 1 j∈Ci,i̸=j b(wi) = min ∑ cosdist(wi, wj ) k̸=i j∈Ck
cosdist(wi, wj ) s(wi) =
b(wi) − a(wi) max(a(wi), b(wi))
The final silhouette score for a given instance is computed by an aggregation of both quantities, dividing by a normalizing factor to ensure a constant output range between -1 and 1 – as shown in Equation 3.
One risk that can be linked to the presented methodology is that the optimal number of clusters may increase because the number of unique verb types in the sample has increased (as shown in Section 3) – i.e. regardless of the semantic composition of the space representing that bin. Thus, increases in the optimal number of clusters can be due to sampling artifacts – an issue that becomes even more likely with fat-tailed distributions that are common in linguistic data. Moreover, even in the absence of sampling artifacts, we must ensure that we are not simply measuring increases in type frequency-based diversity, which, as already argued, are not necessarily indicative of linguistic change.
In order to remedy the afore-mentioned issue, we employ the following bootstrap procedure. For each period, we sample 500 verbs with replacement from the multinomial distribution observed in the dataset and compute the optimal number of clusters based on silhouette score. Repeating this process a 1,000 times per period yields a dataset with 8,000 observations (i.e. for 8 periods), which we submit to statistical analysis in order to quantify the efect of time on the optimal number of clusters. Crucially, we record the total number of distinct verbs sampled in each bootstrap iteration, which allows us to statistically control for the efect of population size on the obtained optimal number of clusters.
We rely on hierarchical (agglomerative) clustering using the cosine similarity and complete linkage,5 and optimize the number of clusters by inspecting the silhouette scores at diferent nodes in the induced merge tree until reaching the merge step that maximizes the silhouette score.6
5We made these choices on the basis of a single manual scan of the interpretability of the clusters induced from the verbs in the entire dataset.
6We use the reference implementations provided by the Python library scikit-learn [
Homing in on the increasing positivity of to death, we leverage the embedding space described in Section 3.2 in order to capture the sentiment polarity of the sampled verbs. Diferences in sentiments are not straightforwardly captured by means of hierarchical clustering, as antonyms are represented by highly similar vectors. In Table 3, for instance, the positive mental verb amuse, is recognized as being similar to more negative mental verbs like delude and terrify, as well as its antonyms annoy and vex. The cluster analysis is therefore supplemented by means of sentiment scores.
A first approach to induce word-level sentiment scores is to exploit the proximity of a given verb vector to the vector for the words ‘good’ and ‘bad’. The closer to the vector for ‘good’ the more positive the sentiment of that verb. However, similar confounding efects from antonyms make this approach unfeasible. Indeed, in common word embedding spaces the vectors for ‘good’ and ‘bad’ tend to be located in the proximity of each other, and, thus, lack discriminative power for classifying words with respect to their sentiment.
In order to tackle this issue, post-hoc modifications of the embedding space such as
retroiftting [
S(wi) = 1
∑ |Ngood| wj∈Ngood 1
∑ |Nbad| wj∈Nbad cos(wi, wj ) − cos(wi, wj ) (4) where Ngood and Nbad refer, respectively, to the filtered set of nearest neighbours of ‘good’ and ‘bad’.
To test the efect of time on the polarity of to death’s collocates, we assign each verb in the dataset to the bin where they are first attested. Given that grammaticalizing structures often retain their original function, it may well be that the well-established negative use of to death vastly outnumbers and hence overshadows cases where to death has expanded to intensify new, more positive verbs. Thus, we suggest that working with the sentiment of collocate verbs that were first attested in a given bin – rather than the distribution of sentiment in each bin – captures the ongoing changes more directly and robustly.
In order to assess the efect of time on the semantic structure of the attested verbs, as well as on the overall sentiment, we fit linear regression models regressing the target outcome – i.e. optimal number of clusters or sentiment score – on the time period. We use a Gaussian likelihood for both outcomes.
7These filtered nearest neighbors were checked in order to avoid too specific terms with unstable sentiment polarity over time. For example, the top 5 neighbours of ‘good’ were ‘better’, ‘excellent’, ‘great’, ‘well’ and ‘best’, while the top 5 neighbours of ‘bad’ were ‘dangerous’, “ill”, ‘inefficient’, ‘wrong’ and ‘hard’.
A further modeling choice we make is to incorporate time period as a monotonic efect –
and not as, for instance, an ordinary linear predictor. This choice is motivated by the fact
that diachronic processes in language structure often result in patterns that resemble s-curves
[
Our implementation of the monotonic predictor follows Bürkner and Charpentier [
Here, b corresponds to the ordinary linear predictor, representing in this case the direction and size of the efect on the outcome, and the individual ζi represent the normalized distances between consecutive predictor categories. The predictor term η is then included in a linear model in the usual way: y = a + η × x. When fitted, this kind of monotonic predictor can easily be interpreted by inspecting the values assigned to the ζi parameters, since these correspond to the relative increase of each category with respect to the total increase involved by the monotonic predictor.
We deploy a Bayesian regression framework which allows us to inspect the uncertainty in the
statistical parameters of interest in an probabilistic intuitive manner. We fit our models using
the Hamiltonian Monte-Carlo sampler provided by the stan library [
In order to test the monotonicity of the efect, we compare a linear model of the efect of time period on the optimal number of clusters – LINEAR(P) – with the monotonic efect model – MONO(P). Moreover, in order to control for the size of the sampled population on the outcome, we fit additional models including the number of unique verbs in the bootstrap sample as predictor – LINEAR(P)+S and MONO(P)+S.
We compare the four models using the Widely Applicable Information Criterion (WAIC), which estimates the plausibility of the models in terms of both predictive performance and model complexity (cf. overfitting). The results of the comparison are shown in the top row of Table 4. Including time period as a monotonic efect improves the predictive power of the model over the linear efect. Moreover, controlling for sample size is even more important, as evidenced by the fact that including it results in a larger improvement in WAIC than modeling period as a monotonic efect.
Using the most strongly predictive model – i.e. MONO(P)+S – we can visualize the (monotonic) efect of time period on optimal number of clusters using the posterior predictive distribution.
Figure 1 depicts the posterior predictive distribution of the optimal number of clusters using a counter-factual triptych plot, statistically controlling for the sample size at diferent percentiles. Overall, we observe a clear monotonic efect, resembling an s-curve, with a leap starting in the 1750 bin. The shape of the efect remains stable across the three sample size percentiles. Due to the positive linear efect of sample size on optimal number of clusters, the range of the outcome (i.e. the y-axis) increases across plots in the triptych. Moreover, the distribution of uncertainty varies from plot to plot. At smaller sample sizes, the uncertainty in the predicted number of clusters is larger towards the later time bins, whereas for larger sample sizes the most uncertain predictions come from the earlier bins. This is likely due to the fact – depicted in Figure 2 – that the sample size in pre-1800 bins is always smaller than in post-1800 bins. However, by counter-factually controlling for sample size, we can observe that the statistical model predicts a constant efect shape regardless of the sample size.
Similarly to the experiments in Section 5.1, we now compare the efect of time period on sentiment using a linear predictor – LINEAR(P) – and a monotonic efect – MONO(P). We use
Period < 1800
FALSE TRUE the standardized average sentiment polarity of the verbs as the outcome. The results in terms of WAIC are shown in the bottom row of Table 4. Modeling time with a monotonic efect produces an improvement over the linear predictor, although in this case the diference with respect to the linear efect model is smaller than in the cluster analysis experiments. The left plot shown in Figure 3 does indicate a slight jump starting in the 1750 bin. However, the large credible intervals observed do not rule out a merely linear efect. Moreover, as the plot in the right hand-side of Figure 3 shows, a considerable amount of variance in the dataset is left unexplained by the model. While statistically controlling for other predictors – such as, for example, document topic or genre – could improve the fit, the current model does show a predominantly linear upward efect of moderate size – about 1 standard deviation – of time on average sentiment.
The results of the statistical analyses are in line with expectations in that the optimal number
of verb clusters increases substantially over the course of the 18th century, when the meaning
of to death expanded to non-literal, intensifying uses (STAGE 2). The predicted shift away
from negative polarity (STAGE 3) also appears to be captured by the statistical model, albeit
weakly. Still, as even in present-day English to death is predominantly attested with negative
collocates [
With an eye on aiding future applications of the models and methods adopted in the present
study, we highlight some important remaining problems.
6.1. Data sparsity and balance
§While there is no shortage of Historical English corpora, corpora that span all the way from
the Early Modern period up to Present-day English are rare. A notable exception is the suite
of the Penn-Helsinki Corpora [
An issue with this patchwork solution is that individual time bins are likely not represented
by a comparable number of texts and text types, which may have consequences at later steps
in the procedure. In the present case, the patchwork corpus sufered from data sparsity in
the 1700-1749 bin, which in turn forced us to cap the maximum number of tokens per bin.
Furthermore, because of the inconsistency with which text types are labelled across the diferent
corpora, it is very difficult if not impossible to smoothly ensure register and genre consistency
across bins. For the present case, such text type inconsistency is indeed very unfortunate: the
time bin in which the host class expansion of to death has taken of also appears to be the time
bin in which the COHA corpus starts, which introduces newspaper and magazine texts into the
sample.9 At the same time, some of the text collections included in the patchwork corpus (such
as ECCO) may contain reprints of older texts, which may have led to an overrepresentation
of older usages in certain time bins. As such, a limitation of the procedure presented here is
that it devotes relatively limited attention to balancing data and/or controlling for text and
text type variation across time bins. A possible solution could be to refrain from working with
corpus patchworks, and turn to the Google Books Corpus (1500 - 2008) or other large library
dumps. Yet, even then issues of overrepresentation (and mislabelling) of texts and text types
may remain [
In short, a substantial challenge for case-driven research is that the careful data curation it
requires may lead to an impasse. Even when following strict procedural guidelines and sanity
checks [
8A somewhat larger corpus covering a very wide time span is the OED quotation database, estimated at 35M
words [
9In the 1800 bin, no tokens are included from newspaper texts, and 82 out of 800 tokens (10.25%) were found in magazine texts.
10Additionally, even diachronic trends in balanced diachronic corpora may in a strict sense also be artefacts,
as genres and registers are also subject to change. With respect to newspaper and magazine text, for instance, it
has been shown that the changing “readerships and purposes of magazines versus newspapers result in diferent
historical-linguistic patterns of use” [
The discussion of to what extent manual interference is needed or desirable in case-driven studies of semantic change is far from trivial, and, ultimately still undecided. In the spirit of the ‘data-drivenness’ discourse in preceding work [e.g. 41, 16], the procedure presented here aimed to minimize manual filtering and annotation – but such manual interference has not been entirely absent. In collecting the collocate verbs of to death, for instance, a substantial number of cases involved structures where the collocate of interest is not the verb be but its accompanying adjective. Similarly, cases where the verb form in closest proximity to to death (e.g. we could prevent Scipio from pummel ling the dreaded wizard to death, COHA 1840) were corrected manually.
Furthermore, there are various points where further ‘manual meddling’ could be considered.
In many instances that were retained in the dataset, to death has neither resultative nor
intensifying meaning.11 Given the limited relevance and potential efects on the output of the
statistical analyses of irrelevant cases, it may be worth flagging or even excluding them from
the dataset, as done in Margerie [
Finally, with respect to the cluster analysis, a fully hands-of approach also implies that we trust the word embedding space to reflect meaningful semantic parameters, and that the resulting clusters capture, at least roughly, relevant properties of the underlying process. In the present case, it is reassuring to see that the narrative that emerges from our data analysis appears to align with what earlier linguistic research has proposed. However, it is not guaranteed that the verb clusters that were fed into the statistical analysis correspond with the semantic verb classes proposed in earlier research (e.g. actions of physical harm vs. mental verbs), or even with any groupings that are meaningful to humans. Additionally, while the bootstrapping procedure described in Section 4.1 renders the procedure more robust, it also makes it more difficult to examine which verbs constitute what cluster at which points in time. Given this lack of full transparency, the (as of yet unanswered) question becomes how to progress towards a method that reliably and robustly supports exploratory data analysis in cases not yet analyzed [also see 49, 20], and to what extent limiting manual involvement to an absolute minimum is warranted in specific case-driven studies.
Drawing on the vast (and growing) body of computational research on semantic change, this study examined how computational models can be employed to track the host class expansion of grammaticalizing constructions, such as to death. By adjusting our scope to one specific and 11For instance, positive verb collocates such as love are attested earlier than expected in examples such as He swore he wou’d love me to death (EEBO, 1700), where to death most likely functions as a time adverbial (‘he swore he’d love me until death’) and not as a resultative (‘he swore me he’d love me resulting in death’) or an intensifier (‘he swore he’d love me a lot’). These structures could, of course, have contributed to to death’s acquisition of intensifying meaning (loving someone until death implies loving them a lot), and hence be relevant to include. In other cases, however, the relevance of the query hit in relation to the semantic development described appears to be much less clear (e.g. the first who turns his back to death ; EEBO, 1800).
12A possible, yet costly solution here would be to rely on multiple annotators, preferably with expertise in the historical language variety at hand [20]. relatively complex case, the procedure we outlined caters to case-driven research, which operates at a level of specificity and granularity that is not abundantly common in computational approaches to semantic change. Besides outlining the procedure, we flagged its current limitations and issues, which will hopefully entice further case-driven computational humanities research that will help reflect on and ultimately tackle the challenges that remain.
The work for this study has been made possible by the Platform Digital Infrastructure (Social Sciences and Humanities) fund (PDI-SSH). We want to thank the anonymous reviewers for their valuable suggestions, as well as Folgert Karsdorp for his advise regarding monotonic efects.