2.1 Related works on identification of lexical semantic changes

Existing surveys on semantic change regularly cover detection but seldom focus on characterization, which is the topic of this paper. In the context of identification, Tang (Reference Tang2018) proposes splitting LSC analysis into four steps: obtaining and preparing a diachronic corpus, identifying word senses, modeling semantic change, and presenting and validating hypotheses. This empirical, rather than historical, paradigm enables the study of semantic change through scientific and statistical methods, allowing findings to be reproduced and tested – a key element for computational research.

Tang’s framework structures the survey by describing studies related to each step. Change identification is mainly discussed in the ‘modeling semantic change’ section, which explains the difference between historiographical and empirical analysis. However, it does not cover approaches that classify changes into categories like broadening, narrowing, pejoration, amelioration, or figurative shifts.

A second survey on computational approaches to LSC (Tahmasebi et al. Reference Tahmasebi, Borin and Jatowt2018) focuses on word-level differentiation, sense differentiation, and lexical replacement. While it includes some types of change categories (e.g., metaphor, metonymy), the main change categories discussed are broadening, narrowing, and correlated types (split, join, birth).

However, one recommendation from the authors is to conduct deeper studies on methods for automatically detecting other types of change and using them to build word histories. The authors provide an in-depth analysis of the linguistic literature, grounding each domain, followed by evaluation strategies and applications. The difficulty of comparing approaches is highlighted and attributed to the absence of standard metrics and reference datasets, but also to the inherent challenges of the task itself.

Kutuzov et al. (Reference Kutuzov, Øvrelid, Szymanski and Velldal2018) survey word representation models and data for semantic change identification and describe a framework for comparing these methodologies. The authors focus on defining consistent terminology and coherent evaluation practices. In this work, they claim there’s still a gap in studying sub-classifications of change:

For the study of LSC and NLP in general, there are four main ways to represent word meaning: (i) frequency of co-occurrence, (ii) graphs of associated words or meanings, (iii) topic modeling for a list of meaningful terms, and (iv) word embeddings. Initial studies in the field employed graph- or frequency-based methods. Later approaches benefited from advances in the state of the art, replacing discrete methods with those based on static and contextual embeddings. Although contextual embeddings represented a massive improvement for a series of NLP tasks (Devlin et al. Reference Devlin, Chang, Lee and Toutanova2019), the improvement for LSC did not follow this trend (Schlechtweg et al. Reference Schlechtweg, McGillivray, Hengchen, Dubossarsky and Tahmasebi2020), showing that there is still room for improvement in this domain (Zhou, Ethayarajh, and Jurafsky Reference Zhou, Ethayarajh and Jurafsky2021; Kutuzov, Velldal, and Øvrelid Reference Kutuzov, Velldal and Øvrelid2022).

Other surveys in LSC investigate the phenomenon of change at the ontological level, trying to discover which laws govern the semantic evolution of words. For example, which kinds of words are more prone to broadening, given their current properties? In this context, Dubossarsky et al. (Reference Dubossarsky, Weinshall and Grossman2017) investigate existing evidence for claims such as (i) word frequency affects meaning change, (ii) a negative correlation exists between meaning change and prototypicality, and (iii) a positive correlation exists between meaning change and polysemy. The authors observed that laws proposed in previous studies are affected by model or data bias. When other investigators reduced the bias, the ‘laws’ either did not hold or were less precise than initially reported.

Research in this field also delves into methodological issues, including the challenges of characterizing semantic shifts. As pointed out by Hengchen et al. (Reference Hengchen, Tahmasebi, Schlechtweg and Dubossarsky2021), few studies detail the nature of these changes, mainly because it is difficult to model meaning from textual data. The paper further concludes that no particular measures for quantifying category change exist and that a consensus on a taxonomy for classifying types of change is required.

In the work of Montanelli and Periti (Reference Montanelli and Periti2023), the authors review methodologies based on contextualized word embeddings. They categorize methodologies for training contextualized embeddings based on how they represent meaning, how time information is presented to the model, and what kind of learning procedure is performed. In this paper, the authors also discuss the importance of characterization:

For this survey, we argue that semantic change characterization needs appropriate methods and tools to analyze each fine-grained category. We also defend that observing variations in a single property cannot describe all forms of LSC well. For instance, changes in the angle between word vectors (Dubossarsky et al. Reference Dubossarsky, Weinshall and Grossman2017) are not the best solution to identify the narrowing or broadening of a sense (Tahmasebi et al. Reference Tahmasebi, Borin and Jatowt2018). Specific methods for identifying different types of change could lead to a better understanding of LSC than straightforward dissimilarity measures.

This survey addresses the need to depict the characterization problem in semantic change and present recent works dealing with this problem. In the next section, we define the survey objectives and the criteria for selecting papers on change characterization.

2.2 Objectives

Different terms refer to LSC, such as semantic drift or semantic shift. In this paper, we distinguish between the identification of a change and the characterization of that change. The former refers to detecting that a word’s meaning has changed, while the latter refers to describing how it has changed. We formally differentiate these problems in Subsection 5.1. The analysis of work on the identification problem has been well covered by Tahmasebi et al. (Reference Tahmasebi, Borin and Jatowt2018), Kutuzov et al. (Reference Kutuzov, Øvrelid, Szymanski and Velldal2018), and Tang (Reference Tang2018).

Semantic change occurs from the need to convey more complex concepts and ideas with a restricted set of known words (Hock and Joseph Reference Hock and Joseph2019). While some authors categorize changes by their cause (e.g., social, cultural), others use linguistic mechanisms (e.g., metaphor, metonymy) as the basis for classification. The focus of our literature review is on the characterization of change, where we look for common properties that allow for grouping identified semantic changes into categories. From a computational perspective, determining the cause of a semantic change is very complex to implement, and, to the best of our knowledge, no computational approach has been implemented yet.

Frameworks have been proposed on how to track LSC, such as the Cambridge and McTaggart views (Tang et al. Reference Tang, Qu and Chen2013), which did not initially consider computational methods. These approaches differ in how they use temporal information to identify changes: while the Cambridge view involves comparing two static corpora from different time periods, the McTaggart view involves tracking the evolution of a word’s usage over time. The Cambridge approach best suits the computational perspective, as the McTaggart approach would require historiographical work to acquire socio-cultural knowledge that goes beyond contextual analysis of word occurrences (Tang et al. Reference Tang, Qu and Chen2013, p 4–7).

This core distinction causes considerable disagreement, as empirical and historiographical methods are based on different ideas and focus on distinct aspects of research. Empirical methods rely on direct observation and measurable data, whereas historiographical methods involve interpreting past events and written records. Since these two approaches operate in different ways and aim to answer distinct kinds of questions, they often yield contrasting views.

For this review, we assume the Cambridge setting for change analysis, using linguistic figures as a basis for change categories, which aligns with current state-of-the-art methods for computational LSC (Pivovarova and Kutuzov Reference Pivovarova and Kutuzov2021). Given these considerations, we adopt the typology from Traugott (Reference Traugott2017), which is also predominant in the literature, as seen in Koch (Reference Koch2016, p 21–66), Hock and Joseph (Reference Hock and Joseph2019, p 189–222), and Campbell (Reference Campbell2013, p 221–246):

• • Broadening: A word gains new meanings, making it applicable to more concepts, for example, ‘cloud’ as a computing infrastructure.

• • Narrowing: A word’s meaning becomes more restricted, applying to fewer concepts. For example, ‘gay’ once meant ‘jolly’ or ‘festive’ but now primarily refers to homosexuality.

• • Amelioration: A word gains a more positive sense. For example, nice changed from ‘foolish, innocent’ to ‘pleasant.’

• • Pejoration: A word is used with a worse connotation. For example, Old English stincan (‘to smell sweet or bad’) changed to stink.

• • Metonymization: A change based on contiguity or association between concepts, for example, board from ‘table’ to ‘a governing body of people who sit at a table’.

• • Metaphorization: Conceptualizing one thing in terms of another, for example, using the heart to represent ‘feelings.’

Since the types in this typology form natural pairs that often involve similar computational challenges (e.g., broadening/narrowing requires sense counting; amelioration/pejoration requires sentiment analysis; metaphorization/metonymization requires a relatedness measure (Blank, Reference Blank2003; Koch, Reference Koch2016; Hock and Joseph Reference Hock and Joseph2019)), we group them into three overarching poles to better compare and categorize research. We only include approaches that compare corpora, as this typology is designed for diachronic studies rather than synchronic variation (e.g., across domains). Figure 1 illustrates the hierarchical taxonomy for LSC.

Figure 1.

Taxonomy for the poles of Lexical Semantic Change, based on the work of Blank et al. (Reference Blank2003); Koch (Reference Koch2016); Hock and Joseph (Reference Hock and Joseph2019).

Figure 2.

Change in meaning and orientation for the word awful. On the left side, we reproduce the figure from Hamilton et al. (Reference Hamilton, Leskovec and Jurafsky2016b) that shows the evolution of the word ‘awful’ in the embedding space. On the right side, we present the hypothetical function ( $\mathfrak{f}$ ) for this word over time.

We assign studies that track sense broadening or narrowing to the dimension ( $\mathscr{D}$ ) pole, where words gaining or losing senses see a change in their dimension value (see Subsection 5.2). This pole is self-complementary, as the change in the dimension’s value addresses both broadening and narrowing. A change in dimension could signal, for example, that a word requires more careful translation (Mohammed, Reference Mohammed2009).

We group changes related to metaphorization or metonymization into the relation ( $\mathscr{R}$ ) pole. These changes occur when a new sense develops from an existing one through a figurative link (see Subsection 5.3). Relations like hypernymy, hyponymy, and synonymy are also factors for semantic change (Koch, Reference Koch2016; Hock and Joseph Reference Hock and Joseph2019) as they involve a conceptual relation and can potentially be encompassed in this pole.

For the orientation ( $\mathscr{O}$ ) pole, we group studies that identify amelioration, pejoration, or any perceptual change in the sentiment associated with a word (e.g., criticism, compliment). This pole also encompasses many forms of change in perception, as some studies suggest that concepts like ‘taboo’ also affect changes in word usage (Koch, Reference Koch2016).

To illustrate a word that changed in orientation, consider ‘awful’. In the 1850s, it meant ‘full of awe’, but today it means ‘extremely disagreeable’. As the first sense is semantically positive and the second is negative, this word underwent pejoration, as illustrated in Figure 2. Notice that the word’s orientation may have first undergone amelioration before pejoration. In the same figure, the word ‘awesome,’ which has a close relation to ‘awe’, gained a more positive sense, undergoing amelioration over time.

In the next section, we describe our method for reviewing the literature based on the poles presented above.

2.3 Review criteria

In this survey, we select papers that represent and analyze words from one of the poles mentioned above in an automatic manner. We exclude works that involve manual analysis, as we focus on computational methods. For the relation and orientation poles, these searches returned many unrelated results, such as studies on synchronic metaphor detection, which do not infer diachronic change and were therefore excluded.

Querying scientific databases required an iterative process to include all relevant terms. In our initial efforts, we searched for ‘[type of change, e.g., broadening] semantic change’ in Scopus, which returned 33 articles; however, none of them were relevant. Due to the lack of consensus in terminology and the rapid pace of the field, we adopted a semantic search approach, starting with key papers and iteratively expanding our search query. Table 1 presents the final set of search terms used to collect articles.

Table 1.

Search terms used to discover articles related to change characterization

From these initial queries, we could find 11 seed papers that satisfy our criteria. We then employed a network semantic citation graph to recover more papers, Research Rabbit.Footnote ^b From the 11 initial papers, the graph suggested 83 new papers; from this, we were able to recover 7 that matched our criteria. We again included these additional 7 papers as seed, obtaining 129 papers. From 129 papers, we obtained 5 more relevant papers. In the last interaction, we obtained 151 paper suggested, but no new relevant papers were found. In Figure 3, we present the graph of documents, where green nodes represent the input papers, and blue nodes represent papers suggested by the tool.

Figure 3.

Graph of selected works (green) and related articles (blue).

In the next sections, we investigate three elements of semantic change characterization in depth: (i) the corpora ( $T$ ) analyzed, (ii) the representation method ( $\mathfrak{R}$ ) for capturing characteristics of change, and (iii) the kind of characterization ( $\{\mathscr{D},\mathscr{R},\mathscr{O}\}$ ) addressed.

5.1 Semantic change

Koch (Reference Koch2016, p 23) presents a definition of meaning change that combines semasiological and onomasiological perspectives. In Figure 10, ‘belly’ was initially associated with the source concept ‘bag’ (M1). Later, it acquired the meaning of the target concept ‘body part between the breast and the thighs’ (M2). In computational linguistics, we primarily focus on the semasiological view: our interest lies in the word itself and the evolution of its meanings, rather than the origins of the concepts it represents.

Figure 10.

Definition of semantic change adapted from Koch (Reference Koch2016, p 23, 25).

The process of meaning change is depicted on the right side of the figure. At time T1, the word ‘belly’ possessed only the meaning M1. Subsequently, at T2, it began to be used with the new meaning M2 as well. This usage of M2 increased over time (T3-T5) until it eventually became the dominant meaning (T6). Finally, at T7, the original meaning M1 was completely lost.

To formalize this definition, we say that a word $w$ had a semantic change ( $\mathscr{C}$ ), if the set of senses, given by the function $\mathscr{S}$ , evaluated in the corpus $t_1$ is different from the set of senses in $t_2$ , given that $t_1\neq t_2$ ,

(3) \begin{equation} \mathscr{C}\,(w, t_1, t_2) = \begin{cases} \text{True} & \mathscr{S}\,(w,t_1) \neq \mathscr{S}\,(w,t_2) \\ \text{False} & \text{otherwise.} \end{cases} \end{equation}

The subset of words in vocabulary $\bar {W} \subset V$ that suffered a semantic change between corpus $t_1$ and $t_2$ is given by,

(4) \begin{equation} \bar {W}_{t_1,t_2} = \{w \mid \mathscr{C}(w, t_1, t_2), \hspace {1em} \forall w \in V\}. \end{equation}

These changes in lexical meaning across corpora are caused by many different phenomena, like time period (1930 and 2020, gay), culture (Brazil and Portugal, rapariga), and domain (engineering and biology, cell).

5.2 Change in dimension

We define that a word $w$ had a changed in dimension if the set of meanings increased or decreased, with respect to the cardinality of the set, when comparing two corpus $t_1,t_2$ . It means that the word suffered a broadening or a narrowing change between these corpora, that is

(5) \begin{equation} |\mathscr{S}\,(w,t_1)| \neq |\mathscr{S}\,(w,t_2)|, \hspace {1em} w \in V. \end{equation}

Consequentially, if $|\mathscr{S}\,(w,t_1)| \lt |\mathscr{S}\,(w,t_2)|$ it suffered a broadening, and if $|\mathscr{S}\,(w,t_1)| \gt |\mathscr{S}\,(w,t_2)|$ , it suffered a narrowing.

This definition is compatible with the concept of innovative/reductive meaning change from Koch (Reference Koch2016, p 24–26), in which words acquire new senses (sememes) by being used with different meanings. It differs, however, from the more general notion of broadening, where words may also lose intension (the precision of their meaning), making them vaguer.

Under this definition, phenomena like sense splitting, birth, and joining are special cases of change in dimension for a single lexical item. For instance, birth or death of a sense is considered a change from or to an empty set. Split and join are instances of broadening or narrowing with respect to a word. For example, the word ‘plane’ gained the sense of ‘aeroplane’ after the 19th century, while previously it was only used to describe the act of planning or a flat surface. This new sense greatly increased the dimension of the word ‘plane’, as many correlated senses, like ‘traveling in an aeroplane’, were also added.

In this survey, we define broadening as the degree of polysemy in a word. This is aligned with the NLP literature, given its application in data analysis and its measurable nature, compared to the notion of intension. We differ from the definition in Tang et al. (Reference Tang, Qu and Chen2013), where broadening is a superclass of metaphorization and metonymization, as the literature reviewed here does not perform this aggregation.

5.3 Change in relation

Metaphorization or metonymization occurs when a word takes on a new meaning that inherits qualities from its original meaning through a figurative relationship the speaker aims to convey (Lakoff and Johnson, Reference Lakoff and Johnson2008, p 3–6). The original meaning is invoked by establishing a material or abstract relation with the initial sense, where the listener is able to resolve the non-literal target concept (Fass, Reference Fass1988). Material relations are built on physical-world contiguity (e.g., replacing an institution with its location, as in ‘Washington decided…’). In contrast, abstract relations are formed through conceptual similarity (e.g., applying human body parts to objects, as in ‘the arm of the chair’).

The theory defining metaphor and metonymy is often informal, relying on subjective perception or overlapping definitions. For example, Choi et al. (Reference Choi, Lee, Choi, Park, Lee, Lee and Lee2021) define metaphor as a case where the ‘literal meaning of a word is different from its contextual meaning’, which also applies to metonymy. As discussed by Koch (Reference Koch2016, p 42–51), there is no precise rule to define a metonymy. Some computational work, such as Schlechtweg et al. (Reference Schlechtweg, Eckmann, Santus, im Walde and Hole2017), measures a word’s entropy to capture metaphorization, but it is not clear that this measure can differentiate a metaphor from a metonymy or a completely new, unrelated sense (e.g., ‘Apple’ in a technology context).

Given these considerations, we define that a word $w$ has changed in relation if the aggregated sum of the relational similarity ( $\mathit{l}: S \times S \rightarrow \mathbf{R}$ ) between its senses changes between corpora. The function $\mathit{l}$ takes two senses, $s_i$ and $s_j$ , and returns a real value corresponding to their material or abstract relation (a greater value means a stronger relation). The total relation for a word is:

(6) \begin{equation} \mathscr{R}(w, t)= \left \{ \begin{aligned} \sum ^{N-1}_{i=1}\sum ^{N}_{j=i+1}{\mathit{l}(s_i, s_j)} \hspace {3em} s_i,s_j \in \mathscr{S}\,(w, t)\\ 0 \hspace {6em} N \leq 1, \end{aligned} \right . \end{equation}

where $N=|\mathscr{S}\,(w, t)|$ .

A word $w$ increases in relation between corpora $t_1$ and $t_2$ if $\mathscr{R}(w,t_1) \lt \mathscr{R}(w,t_2)$ , and decreases if the opposite occurs. Since we lack methods to perfectly differentiate material and abstract relations, and since a word can undergo both metaphorization and metonymization, we simplify the interpretation: if a word’s figurative usage increases, it ‘increases in relation’ ; if its figurative usage decreases, it ‘decreases in relation’.

5.4 Change in orientation

A word’s change in orientation (positive or negative connotation) reflects amelioration or pejoration in its usage from one corpus to another. This is measured by analyzing how the collective perception of a word’s meaning evolves (Campbell, Reference Campbell2013, p 227–229). A classic example is ‘terrific,’ which originally meant ‘terrible’ but shifted to mean ‘formidable’ or ‘impressive.’

While some frameworks like Buechel et al. (Reference Buechel, Hellrich and Hahn2016) use a multi-dimensional Valence-Arousal-Dominance setting, most sentiment analysis research focuses on a simpler polarity setting (Cui et al. Reference Cui, Wang, Ho and Cambria2023). Therefore, we restrict our formalism to this more general positive/negative view to align with current research and applications (Susanto et al. Reference Susanto, Livingstone, Ng and Cambria2020).

We define that a word $w$ had changed in orientation, if for some notion of orientation $\mathfrak{f}$ (feeling), there is a mapping from a word $w$ in a corpora $t$ to an orientation value, for example, in case of positive, neutral, or negative feeling,

(7) \begin{equation} \mathfrak{f}\,:\,V \times T \rightarrow \{-1, 0, +1\}, \end{equation}

the orientation changes between corpora $t_1,t_2$ ,

(8) \begin{equation} \mathfrak{f}(w, t_1) \neq \mathfrak{f}(w, t_2). \end{equation}

Analogous to the previous definition, if $\mathfrak{f}(w, t_1) \lt \mathfrak{f}(w, t_2)$ it suffered an amelioration, and a pejoration if $\mathfrak{f}(w, t_1) \gt \mathfrak{f}(w, t_2)$ .

This task relies on a subjective notion of orientation captured by the function $\mathfrak{f}$ . Generally, studies define ‘orientation’ as the proximity to positive or negative seed words. However, the problem is broader than just polarity, as words can be classified along other emotional dimensions.

5.5 Semantic change identification

Core to computational approaches, words in a corpus are not compared directly to one another; instead, they are first represented using one of the many methods reviewed in this survey. A common step taken by all authors is to represent the word ‘a’ in corpus ‘x’, denoted as $\mathfrak{R}(a,x)$ , and then compare this representation to $\mathfrak{R}(b,y)$ . For example, Moss (Reference Moss2020) represents the word ‘a’ as a Gaussian embedding obtained after training a language model on corpus ‘x’. Similarly, Ehmüller et al. (Reference Ehmüller, Kohlmeyer, McKee, Paeschke, Repke, Krestel and Naumann2020) represent ‘a’ using a co-occurrence graph derived from corpus ‘x’. Meanwhile, Inoue et al. (Reference Inoue, Komachi, Ogiso, Takamura and Mochihashi2022) partition the original corpus into 20-year intervals and, for each resulting corpus ‘x’, obtain a distribution of topics for the word ‘a’.

In this context, we now define in computational terms the task of semantic change characterization. Consider a function $f_{\mathscr{C}}$ that takes two representations $\mathfrak{R}$ (e.g., embeddings, graphs, etc.) of the word $w$ in corpora $t_1,t_2$ . We define the problem of semantic change identification as assigning a value ( $y$ ) to this word,

(9) \begin{equation} f_{\mathscr{C}}(\mathfrak{R}(w,t_1), \mathfrak{R}(w,t_2)) \rightarrow y. \end{equation}

Schlechtweg and im Walde (Reference Schlechtweg and im Walde2020) proposed two notions of semantic change: binary classification, where $y \in \{0,1\}$ , and graded (continuous) semantic change, where $y \in \mathbf{R}$ is a distance between word representations.Footnote ^c

The binary classification task aligns with classic LSC literature (Blank, Reference Blank2012, p 113) and human perception, which is often influenced by sensory thresholds (Smith, Reference Smith2008, p 34). This view is interested in the gain or loss of senses. There is no general threshold criterion for the binary setting; it’s not enough for representations to be different, the magnitude of this difference should align with human judgment.

The graded task originates from computational linguistics, where a natural way to compare objects is by measuring the distance between their representations. This approach allows for a more in-depth analysis of change, as the degree of change can be compared across a wide range of contexts.

5.6 Semantic change characterization

Deriving from the reviewed literature, we define semantic change characterization as a classification problem where we need to assign identified semantic changes to categories, such as $f_{\mathscr{D}},f_{\mathscr{R}},f_{\mathscr{O}}$ ,

(10) \begin{equation} f_x(\mathfrak{R}(w,t_1), \mathfrak{R}(w,t_2)) \rightarrow y,\hspace {2em} x \in \{\mathscr{D},\mathscr{R},\mathscr{O}\}. \end{equation}

While the first equation only measured some notion of distance between the representations, for characterization, we first apply the type function on the representation, to then compare. For example, in Buechel et al. (Reference Buechel, Hellrich and Hahn2016), we have an orientation characterization, for example,

(11) \begin{equation} f(\mathscr{O}(\mathfrak{R}(w,t_1)), \mathscr{O}(\mathfrak{R}(w,t_2))) \rightarrow y, \end{equation}

where the $\mathscr{O}$ function is the Valence-Arousal-Dominance framework and they measure the distance $f$ through the $\beta$ -coefficient of a linear regression.

Again, we set two possibilities of change, binary and graded. In this setting, the binary form allows us to answer interesting questions in an automatic manner, for example, at what point in time the word ‘head’ started being used in a figurative manner? Another possibility is to track the precise corpus where words were pejorated, and then manually understand the socio-linguistic process of this phenomenon, saving many hours of historiography work.

5.7 Framework evaluation

The proposed framework enables a spatial understanding of a word’s change and direct statistical comparison of corpora. In Figure 11, we illustrate this analysis with a possible evolution of the word ‘heart’ between two corpora from different times (1993 and 2013). A gradual analysis of this evolution can provide a detailed picture across the three poles of change.

Figure 11.

Illustrative example of the word ‘heart’ changing over time. Metaphorization, amelioration and broadening can occur for the same word, depending on the senses it gained/lost.

To illustrate the framework’s practical aspects, we investigate two corpora: SEMCOR (Miller et al. Reference Miller, Leacock, Tengi and Bunker1993) and MASC.Footnote ^d SEMCOR is a sense-tagged corpus of 200,000 words created in 1993. MASC is a manually annotated sub-corpus of 500,000 words of modern American English released around 2013.

While the corpora were produced at different times (1993 and 2013), we do not claim this analysis is sufficient to discover all examples of historical LSC. The corpora have narrow coverage and are inadequate for robust diachronic analysis. Additionally, WordNet doesn’t cover all existing senses for a given word. We use these datasets for illustrative purposes only. For the purpose of this demonstration, we define corpus $t_1$ as SEMCOR and $t_2$ as MASC, with a time difference of 20 years between the corpora.

SEMCOR and MASC are corpora manually annotated using WordNet senses. We obtain annotations for the poles of change from two lexical resources: ChainNet, a WordNet extension linking senses by metaphor and metonymy, and SentiWordNet, which provides a positive/negative score for every sense in WordNet. This step allows us to skip sense inference and focus on demonstrating the framework. For automatic WordNet annotation, we suggest Navigli et al. (Reference Navigli, Camacho-Collados and Raganato2017). For work on annotating data for the poles of change independently of a lexical resource, we suggest de Sá et al. (Reference de Sá, Silveira and Pruski2024). From SEMCOR and MASC, we computed occurrences of ‘heart’ and ‘plane,’ two well-known words that have changed semantically.

According to SentiWordNet, ‘heart’ has two positive senses: courage (heart.n.03) and the vital part of an idea (kernel.n.03). From ChainNet, the absolute literal meanings are the organ (heart.n.02) and courage (heart.n.03). We can observe that in MASC, the literal meaning ‘organ’ (heart.n.02) has no occurrences, while figurative usage becomes dominant. The word also gained a new sense (heart.n.10, the playing card suit).

From Table 5, we can compute the equations to verify the semantic change across the corpus. We first start with the dimension test. From Equation (5) we have:

Table 5.

Distribution of WordNet senses for the word ‘heart’ in SEMCOR and MASC. The score column indicates a positive score for that sense from SentiWordNet, figurative column indicates if the meaning is figurative or literal

(12) \begin{equation} |\mathscr{S}\,(w,t_1)| \stackrel {?}{=} |\mathscr{S}\,(w,t_2)| \end{equation}

(13) \begin{equation} |\mathscr{S}\,(\text{heart},\text{SEMCOR})| \stackrel {?}{=} |\mathscr{S}\,(\text{heart},\text{MASC})| \end{equation}

(14) \begin{equation} \begin{gathered} |\{\texttt {heart.n.03}, \texttt {kernel.n.03}, \texttt {center.n.01}, \texttt {heart.n.01}, \texttt {heart.n.02}\}| \\ \stackrel {?}{=} \\ |\{\texttt {heart.n.03}, \texttt {kernel.n.03}, \texttt {heart.n.10}\}| \end{gathered} \end{equation}

\begin{equation*} 5 \gt 3, \text{we conclude that it narrowed.} \end{equation*}

To evaluate changes in Orientation, we define the feeling function $\mathfrak{f}$ as the weighted sum of the positive scores of a word’s senses. Specifically, we compute $\mathfrak{f}$ as the sum of $\text{positive}(s_i) \cdot p(s_i)$ , where $\text{positive}(s_i)$ is the score of sense $s_i$ , and $p(s_i)$ is its associated probability. A sense is considered positive if its score is greater than 0.

We use the weighted sum as a proxy for the prototypical sense, assuming that more frequently used senses have a greater influence on the perceived feeling of the word.

(15) \begin{equation} \mathfrak{f}(w, t) = \frac {1}{N}\sum _{i=1}^N p(s_i) \text{positive}(s_i), \hspace {1em}s_i \in \mathscr{S}\,(w, t) \end{equation}

Plugging the values into our equations, we have:

(16) \begin{equation} \mathfrak{f}(\text{heart}, \text{SEMCOR}) \stackrel {?}{=} \mathfrak{f}(\text{heart}, \text{MASC}) \end{equation}

(17) \begin{equation} \begin{gathered} 1 \times 0.12 + 1 \times 0.03 + 0 \times 0.06 + 0 \times 0.0 + 0 \times .44 + 0 \times 0.35\\ \stackrel {?}{=} \\ 1 \times 0.9 + 1 \times .07 + 0 \times 0.0 + 0 \times 0.02 + 0 \times 0.0 + 0 \times 0.0 \end{gathered} \end{equation}

\begin{equation*} 0.15 \lt 0.97, \text{we observe an amelioration.} \end{equation*}

Table 6.

Distribution of WordNet senses for the word ‘plane’ in SEMCOR and MASC. The score column indicates a positive score for that sense from SentiWordNet, and figurative column indicates if the meaning is figurative or literal

Figure 12.

Change in meaning distribution for the word ‘heart’ in SEMCOR and MASC.

Figure 12 illustrates the change in meaning distribution for ‘heart.’ Under our framework, the word becomes more metaphorical (increasing in Relation, $\mathscr{R}$ ), narrowed for these corpora (decreasing its Dimension, $\mathscr{D}$ ), and more positive (increasing in Orientation, $\mathscr{O}$ ).

For the word ‘plane’ (Table 6), we note that while SEMCOR reports five distinct senses, MASC presents only two. In MASC, ‘plane’ has lost its usage as a surface (flat.s.01), a level of existence (plane.n.03), and a type of cut (plane.v.01), while the prototypical usage of ‘airplane’ has become more prevalent.

For the Relation dimension, we compute $l$ by counting the number of linked senses as described in Equation (6).

(18) \begin{equation} \mathscr{R}(\text{plane}, \text{SEMCOR}) \stackrel {?}{=} \mathscr{R}(\text{plane}, \text{MASC}) \end{equation}

(19) \begin{equation} \begin{gathered} \mathscr{R}(\text{plane}, \text{SEMCOR}) = \mathit{l}(\text{plane.n.03}, \text{plane.n.02}) = 1\\ \stackrel {?}{=} \\ \mathscr{R}(\text{plane}, \text{MASC}) = 0 \end{gathered} \end{equation}

\begin{equation*} 1 \gt 0, \text{so the sense is less metaphorical and decreased in relation.} \end{equation*}

Figure 13.

Change in meaning distribution for the word ‘plane’ in SEMCOR and MASC.

In Figure 13, we can see the narrowing effect this word suffered in this corpus. The decreased usage as a positive meaning (flat.s.01) causes a decrease in its Orientation ( $\mathscr{O}$ ).