2.1. Contemporary language change

The development of interest in language change in living languages is not exclusively a contemporary phenomenon, as its roots can be traced back to the 19th century (e.g., the Neogrammarians or the second Wittgenstein). Regarding lexical semantics, the study of word meaning over time is usually referred to as a ‘semantic shift’ or ‘semantic change’.

The theoretical background of semantic change, originating from cognitive and psychological mechanisms, can be found in the pre-structuralist postulates of Bréal [14] who enunciated six types of semasiological change: ‘pejoration’ versus ‘amelioration,’ ‘restriction’ versus ‘expansion’, ‘metaphor,’ and ‘metonymy’. In the first axis, the meaning of a word changes to become more positive or negative. Geeraerts [15] points out several driving forces of semantic change. Largely, it is (i) subject to non-referential meaning and (ii) culturally motivated (i.e., it changes with the cultural evolution of views). Semantic change also entails a psychological process of subjectivation. Provencio Garrigós [16] evidences a process of subjectivization, from the objective meaning of a word in the voice of a speaker (depending on external factors) to a subjective meaning that describes an internal or subjective situation (subject to the speaker’s cognos).

Motivations for semantic change are of two types: ‘linguistic drifts’ (changes in the central meaning of words, mainly slow and regular) and ‘cultural shifts’ (variations in word associations, mainly culturally motivated) [17]. Half a century ago, Waldron [18] identified changes in meaning based on variations between lexical relations: “Change of meaning, then, essentially concerns a shift of relations between words on the one hand and integral senses on the other hand” (p. 139). In relation to this, Traugott [19] reflected on the conditions of semantic change, considering that behind every variation in meaning, there is a chain of causality that, in fact, belongs to extralinguistic causes (i.e., material, social, or psychological conditions). We adhere to Lyons’ [20] assertion that even if semantic change is regular and continuous, “languages change more rapidly in certain periods than they do in others” (p. 183).

Finally, the use of computational techniques to study semantic changes has emerged as a distinct research methodology over the last decade. Different linguistic theories and models have been proposed to explain rules and regularities in semantic change, such as Diachronic Prototype Semantics [15] and the Invited Inference Theory of Semantic Change [21].

2.2. Corpus-based methodologies to study semantic change

The availability of large corpora and the development of natural language processing have led to data-driven approaches that produced paradigm changes. Many research initiatives have emerged to grapple with tasks that were difficult to approach several decades ago. A state-of-the-art review by Kutuzov et al. [17] compiled dozens of papers on this subject using heterogeneous methodological approaches. Three main research communities are interested in this particular topic: (i) natural language processing/computational linguistics, (ii) information retrieval/computer science, and (iii) political science. Alessi & Partington [22] comprehensively described the tools and methodologies employed in the novel discipline of ‘modern diachronic corpus-assisted language studies’. This study reflects on the consideration that tracking contemporary language has received little attention in linguistics, which has been justified by two main factors: (i) the difficulties in developing instruments and methods to carry out this task, and (ii) the difficulty in observing changes in the short term. Corpus linguistics and natural language processing techniques provide solutions to these problems by providing sophisticated tools to process large amounts of contextualized text, which facilitates the study of the evolution of different text types over short periods of time.

The analysis of collocations is a fundamental aspect of corpus-based studies, particularly when exploring semantic change. This method plays a key role in understanding how words and phrases co-occur, and how they can evolve over time. Williams [23] revealed the crucial role that collocational networks play in the formation of meaning. The acquired connotations of a word’s meaning can ‘push’ its original semantics into a subordinate position. In relation to this, Hoey [24] proposed an expansion of the idea of ‘lexical priming’, in which a word gradually acquires meaning for each individual speaker as it becomes primed for collocation, colligation, and semantic and pragmatic associations. Repetition is a key factor, as every re-encounter that a speaker may have with a particular word reinforces this conditioning.

The concept of ‘semantic prosody’ or ‘semantic association’ is also central to the qualitative study of this work. Post-Firthian scholars Sinclair [25] and Louw [26] were the first to use the term, which refers to expressions associated mostly with sentiment-related meanings, whether positive or negative. Louw defined semantic prosody as “a consistent aura of meaning with which a form is imbued by its collocates” (p. 157), in a way that maintains evaluative harmony and listener comprehensibility. Semantic prosody is thus a non-static property of the lexicon but variable in time, which researchers can explore, explain, and predict using a text corpus. According to Bublitz [27], a set of collocations evolves to acquire a tinge of sentiment (positive or negative) over a time interval and context analysis of the term. To illustrate this, from a qualitative Political Sciences perspective, Stratigaki [28] argues that economic actors choose original terms in order to fit them into their political agenda.

Finally, we argue that terms can develop semantic prosody as they gain prominence in news discourse following a major event, gaining gradual collocation with new non-domain-specific (i.e., general language) words over time. It is important to understand that certain words, such as ‘ease’ or ‘fight’, which may be said to carry specific sentiment taken in isolation, may act simply as sentiment shifters that modify the inherent semantic orientation of the terms they collocate with, such as ‘inflation’ or ‘debt’. Thus, these shifters are to be considered general collocates, and are used to describe the economic situation, regardless of the sentiment conveyed in the context.

In this work, however, we conform to the semantic prosody assumption that a higher frequency of sentiment words of a certain polarity in the vicinity of an event word is an indicator that a shift towards that polarity is taking place regarding that word, and therefore changes in its semantic orientation can be tracked by studying the contexts of the focus terms over time.

2.3. Sentiment analysis

The wide availability of electronic text on the Internet has led to an increase in research on information extraction, a subfield of natural language processing that seeks to obtain structured data from unstructured text. Sentiment Analysis (SA) [29], also called opinion mining, is the field of natural language processing and textual analysis that studies “people’s opinion, sentiments, evaluations, appraisals, attitudes, and emotions towards entities such as products, services, organizations, individuals, issues, events, topics, and their attributes” (p. 451). The basic tasks of a SA system include polarity detection and emotion recognition [30]. A typical SA system classifies texts according to their polarity (i.e., positive or negative) or output on a scale (e.g., 0 for very negative and 100 for very positive). Discourse, syntactic, textual, and intertextual phenomena play crucial roles in sentiment analysis. Many disciplines, including linguistics, psychology, political science, and economics, make use of their applications.

There are two main approaches to SA: machine learning and lexicon-based. Machine-learning approaches rely on datasets that contain annotated sample sentences and different features that are subsequently processed and learned by a computational system. Unsupervised machine-learning approaches in the form of neural networks (specifically, Transformers-based models, such as BERT) offer current state-of-the-art results for SA tasks [31]. In lexicon-based systems, sentiment words are detected using sentiment dictionaries of varying quality and level of detail. Ideally, context is processed to account for the modification of polarity or valence. These semantic orientation modification mechanisms have been studied and formalized through various systems of rules. Polanyi & Zaenen [32] first proposed at a theoretical level what they called ‘sentiment shifters’ or ‘contextual valence shifters’, a series of rules that modify the sentiment of the expression by inverting polarity, its intensification, or its attenuation. It is also possible to detect in-text evaluative segments and different aspects of the entities to which such segments refer, a task known as aspect-based sentiment analysis [33].

The traditional criticism to machine learning approaches (e.g. [34, 35]) is that they need to be trained on the same type of data that the classifier is going to be used with, that is, a classifier trained on a movie reviews dataset will not perform well on Amazon consumer product reviews. Current state of the art in sentiment classification is offered by unsupervised machine-learning approaches in the form of neural networks that use Transformers [31]. Language models based on the Transformer architecture have been shown to improve on previous top benchmark scores across numerous NLP tasks, both in natural language understanding and generation, including sentiment analysis [36]. The Transformer architecture, being unsupervised, might be thought to alleviate the need for domain-specific datasets; however, although pre-trained language models exist that can be used “off-the-shelf”, these models still required to be fine-tuned on smaller, domain-specific, datasets using traditional supervised learning techniques.

Regardless of their accuracy, machine learning classifiers pose a second important limitation for our objectives, however, as they can help us assess what proportion of texts (or sentences) are positive or negative but tell us nothing about which words and expressions were identified that led to these classification results. A machine learning (or deep learning) sentiment classifier is a predictive model that acts as a black box which can provide no explanations as to how the classification result was arrived at, only (in certain cases) a confidence score. This is where lexicon-based sentiment analysis systems can be useful, since they determine the overall sentiment of a text by identifying specific sentiment words and phrases, which they can of course list in their output along with the classification.

These systems, however, do require rich lexical knowledge to achieve good results in different domains and their quality and coverage are the most determining factor in their performance. Thus, a high-quality lexicon is necessary to achieve good performance, along with a system that is able to process the linguistic context to account for the above-mentioned sentiment shifters, such as inversion (e.g., negation), intensification, and downtoning, in the form of context rules. Examples of this type of system include SO-CAL [37] and Lingmotif`[38], a multilingual sentiment analysis tool that offers these features and whose English lexicon has been shown to be high-coverage and high-quality. In this work we employ the latter and we do not expect this tool to be 100% accurate, since aspects such as connotation, irony, sarcasm, discourse structure, etc. pose serious problems to automatic analyzers of this kind (see [39] for a thorough description of linguistic aspects that intervene in the expression of evaluation). Still, we need to know what level of accuracy we have with our own domain-specific texts; thus, we carried out our own performance evaluation. We describe the tool and the experiment’s results in section 3.3.

3.1. Hypothesis

Our hypothesis is that there are significant short-term sentiment changes in economic terms once they become ‘event words’. Thus, in this study, we aim to identify and track such sentiment shifts by analyzing the sentiment of the contexts in which those event words occur. We consider that terms become event words in the news when there is a sudden rise in their relative frequency. Conceptually, the task of discovering sentiment shifts from data can be formulated as follows: given a time-annotated corpus containing news items, the task is to discover changing sentiment trends in sentences containing ‘inflation’ in different time periods and tracing its collocational changes.

Fernández-Cruz [40] considers ‘inflation’ a neutral-sentiment economic marker, as it does not have an inherent sentiment, and its classification relies entirely on the expert’s interpretation of inflation figures. Its semantic orientation can be described as a potentially negative indicator [41], as rising inflation figures (i.e., high inflation) are considered to have negative sentiment and vice versa. However, as discussed in the data analysis section, extremely low inflation figures are considered unfavorable.

3.2. Data sources

The data source is the Great Recession News Corpus (GRNC) [42], a specialized (economic/financial domain) text compilation composed of approximately 35,000 articles from the business sections of The Guardian (52.9%) and The New York Times (47.1%) published between 2007 to 2015. These two newspapers are among the top influential newspapers worldwide [41] and they are considered to be ideologically in the center-left ideological spectrum. Divided into annual subcorpora, the latest version of the corpus contains approximately 26 million words. Finally, to enable ease of analysis and replication, filenames in the corpus are encoded in the format of [YYMM-Newspaper Code-File Number] and they are specified in the examples we provide in Section 4. All datasets and results are available in [43] to ensure replicability.

We obtained two datasets: the first one consists of all sentences (N = 5,481) containing the search term ‘inflation’ in the GRNC, along with a set of sentiment scores and other textual statistics generated by Lingmotif. Secondly, we extracted a dataset of the collocates of the term ‘inflation’ by year. This time interval was selected as it permits a concise yet detailed analysis and avoids seasonality. The LogDice statistical measure [44] was chosen with a cut-off value of 7, a window size of 5 words on either side, and requiring a minimum frequency of 5 instances for the collocate in the corpus and 3 for the collocate in the given range. Function words and proper nouns were removed, as they do not convey sentiment by definition. LogDice subsumes the frequency and exclusivity of collocations and returns a normalized measure in the range 0–14, which reduces the subcorpus size bias [45].

3.3. Instruments

Lingmotif [38], the sentiment analysis tool employed in this study, is a multilingual lexicon-based suite that can be used with general languages and domain-specific texts to calculate sentiment and textual values. It is implemented as a Python 3 library that is accessed through a user-friendly web interface using a REST API. By identifying linguistic expressions of polarity, Lingmotif determines the semantic orientation of texts and sentences, and also provides rich quantitative information in the form of various sentiment and text metrics. The tool splits a text in one or more segments of the same size and returns partial and total sentiment metrics, along with a comprehensive set of text stats (e.g., number of nouns or the number of positive/negative words). Our study is based on Lingmotif’s main sentiment metric, the Text Sentiment Score (TSS), a quantitative scale in the range of 0–100. It quantifies the sentiment of the analyzed text from extremely negative (0) to extremely positive (100). A value of 50 on the TSS scale is considered neutral. Importantly, Lingmotif arrives at a neutral score in one of two situations: either no sentiment words are found or the positive and negative words found cancel each other out. This implies that even scores slightly higher or lower than 50 are strong indicators of positive or negative polarity.

The application relies heavily on Lingmotif-lex’s lexical resources [46], an English sentiment lexicon with over 26,000 single-word forms and over 41,000 multi-word items that have been manually curated to have wide coverage and be domain-neutral. Lingmotif allows the base lexicon to be overridden via plugin lexicons to address domain-specificity issues in sentiment calculation. For this study, we used SentiEcon [47], a fine-grained specialized sentiment lexicon in the financial domain. It contains 6,470 entries for both single and multiword expressions. A comprehensive set of sentiment shifters (nearly 1,000 in the current version of Lingmotif) was used to account for context-dependent valence modifications.

Even though both Lingmotif’s core lexicon and the SentiEcon plug-in lexicon have been evaluated with good results in the past, it is necessary to ascertain whether TSS classification with this particular dataset is accurate enough. Thus, we carried out a formal evaluation study using a random sample (n = 360) from the full set of sentences (N = 5,481) in our main dataset (Confidence Level = 95%, Margin of Error = 5%). Three annotators independently labelled the sample dataset as belonging in one of three classes: positive (TSS>50), negative (TSS<50), neutral (TSS = 50). In order to avoid bias towards the lexicon-based classifier, annotators were asked to classify the sentences according to the overall sentiment they thought the sentences expressed, regardless of the number of sentiment words that were present in them. In order to ensure annotation reliability, Inter-Annotator Agreement (IAA) was calculated using the Disagree Python package [48]. Agreement was found to be very high for both Fleiss’s kappa (κ = 0.878) and Krippendorff’s alpha (α = 0.878). The gold standard was arrived at by majority vote on differing labels, as no case was present where all three annotators disagreed on the label.

Precision and recall evaluation metrics were calculated using the scikit-learn [49] Python library. With an overall accuracy of 76% and a macro-averaged F1 score of 0.73, the classification predictions can be said to be very acceptable for a multi-label classification task. The lowest scores were found to be for the neutral class, as Lingmotif’s neutral classification can be arrived at in either of two very different ways: when no sentiment expression is found or when the positive and negative expressions cancel each other out. Table 1 shows the classification with all relevant evaluation metrics.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Lingmotif’s evaluation metrics for three-way sentiment classification.

https://doi.org/10.1371/journal.pone.0287688.t001

Finally, our study also utilizes Lancaster Stats Tools [50], an online statistical tool, and The Sketch Engine [51], a web-based tool that extracts linguistic information, (in our case, relative frequencies per million tokens, collocations and concordances) from large corpora. We uploaded the GRNC to The Sketch Engine in XML format so that the subcorpora feature was available. Relative frequency was calculated by dividing the number of occurrences of ‘inflation’ in each yearly subcorpus by the total number of words in the subcorpus, and then multiplying it by 1 million.

We also use Lancaster Stats Tool to carry out Usage Fluctuation Analysis (UFA). This method attempts to examine the fluctuation of word usage manifested through collocation, that is, the co-occurrence of words in texts. The instrument calculates Gwet’s AC1 agreement statistic, a measure of inter-rater reliability that is used to assess the level of agreement between two or more raters when assigning categorical ratings to a set of items. McEnery et al. [52], the original proponents of this method, use this statistical measure to calculate the agreement between different lists of collocates in a timeline. The technique relies on two simple assumptions: (i) words co-occurring in the vicinity of other words provide insight into the words’ usage (collocation principle), and (ii) the change in the pattern of co-occurrence of words over time can identify points where their usage changes. The tool returns a set of collocation lists: consistent for collocates that do not change in the timeline), initiating (collocates that emerge after the second data point and remain until the end of the series), terminal (collocates that emerge in the last data points, and transient (collocates that emerge after the second data point and decline before the final data points).

3.4. Method

Our procedure to examine discover and study semantic orientation changes of terms over time consists of four steps, which we describe below.

In step 1, we process the sentences dataset with Lingmotif and compare the yearly TSS means. To do so, we perform annual parametric (one-way ANOVA, post-hoc tests: Bonferroni adjusted t-test, p-values) and non-parametric [53] statistical analyses with the Lancaster Stats Tools online [50]. The hypotheses are as follows: (i) H0: μ1 = μ2 = μ3… = μ9, and (ii) H1: means are not all equal.

In step 2, we observe general trends in a more detailed manner. To do so, we aggregate the text sentiment score (TSS) figures by quarterly averages and calculate the relative frequency of the terms for each of the 36 data points of the time series (corresponding to quarters). The dispersion points are fitted and analyzed using the peaks-and-troughs [54] technique in order to measure the difference in sentiment/frequency between consecutive points during the period. Peaks-and-troughs is a technique used in time series of diverse nature to identify trends in data by analyzing the high (peaks) and low (troughs) points of a chart over a certain period of time.

In step 3, we apply Usage Fluctuation Analysis to the dataset of collocations by year, which provides a concise and reliable comparison throughout a specific timeline. The parameters used in the tool were as follows: LogDice>7, Consistency Threshold 80%, 1 sampling point, Absolute Frequency cut-off = 0.

In step 4, a small number of representative collocations are selected for qualitative analysis by examining concordance lines to provide a more detailed picture of how the focus terms are characterized through the time series.

4.1. Sentiment analysis

Table 2 provides descriptive statistics of several basic text metrics by year for the first target term. Fig 3 shows the distribution plot throughout the corpus. There was a statistically significant effect of diachrony on TSS values (one-way ANOVA, F (8, 5481) = 5.75, p < 0.001; Kruskal-Wallis H (8) = 47.79, p < 0.001). The effect size was small (ω = 0.083).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Distribution plot of ‘inflation’ trough the corpus.

https://doi.org/10.1371/journal.pone.0287688.g003

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Statistic summary for ‘inflation’.

https://doi.org/10.1371/journal.pone.0287688.t002

4.2. Peaks-and-troughs

Fig 4 plots the quarterly averages of TSS. To observe the trends more clearly, we processed these figures using a regression model (Linear GAM) of sentiment scores on a regression curve with 95% and 99% confidence intervals [55]. The form of the sentiment evolution of the term is that of a rotated s-shape. We identified two peaks, corresponding to 2007 and 2013–2014 and one trough (2008–2012).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Quarterly peaks-and-troughs for TSS.

https://doi.org/10.1371/journal.pone.0287688.g004

In Fig 4, first we observe a steep increase in negative sentiment in 2007. The second phase (2008–2012) is characterized by a predominant negative sentiment and intermittent stabilization, with a discernible shift to a positive sentiment beginning to take shape around 2010. The last interval corresponds to the years 2014–2015, with a sharp rise to positive levels (TSS>50), followed by a fall to the threshold of negativity (TSS<50) in 2015. The term ‘inflation’ becomes an event word in times of change, as relative frequency in the corpus rises significantly in 2008, 2011, and 2014–2015. The stages are outlined as follows:

1. In the first period (64.84 occurrences per million tokens in the GRNC), the average TSS starts from a positive sentiment in 2007 (TSS = 51.21) so that, in relative terms, most sentences referring to inflation are positive in nature.
2. The second period (2008–2012) starts with a steep decline of 10.72% and TSS = 40.49 in 2008, after which scores stabilized, rising slightly to approximately 46 in 2009, 2010, and 2012. The fall in sentiment values is not accompanied by a sharp rise in the relative frequency of the term in the corpus; rather, it is irregular over this period (135.27 per million tokens in 2008, a peak of 193.29 in 2011).
3. The third period (2013–2015) encompasses the time after the policy changes in the European Central Bank. This reflects a pronounced upward trend in sentiment score averages, reaching TSS = 50.51 (i.e., slightly positive) in 2014, coinciding with Quantitative Easing policies, to subsequently fall again below the threshold of negativity in 2015 (TSS = 48.68), after the Eurozone’s drop to negative inflation in the first quarter. After an initial period of continuity in the level of frequency in discourse, the frequency increased in the 2013–2015 interval (2013:147.00; 2014:222.13; 2015:358.33 per million tokens).

4.3. Usage fluctuation analysis

The next step is usage fluctuation analysis (UFA), which identifies 10 consistent collocates across the nine periods. Table 3 presents a comprehensive chart of the results summarized in Fig 5, indicating the sentiment categories of the most significant collocates though the timeline. The results show that the bulk of shifts in word use (transient collocates) occurred after the outbreak of the 2008 crisis, when the term became an event word and negative collocational patterns emerged.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. A diachronic summary of relevant UFA collocates and their sentiment.

Red: negative, grey: neutral, green: positive.

https://doi.org/10.1371/journal.pone.0287688.g005

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. UFA analysis of the collocates of ‘inflation’.

Underlined words convey positive sentiment, words in italics convey negative sentiment.

https://doi.org/10.1371/journal.pone.0287688.t003

This emergence of new collocates indicate a period of pejoration in the sentences that include ‘inflation’ and it is consistent with TSS figures. The interval from 2011 onwards on both transient and terminal collocates signals semantic amelioration.

consistent collocates all fall within the domain of economics, whether specialized (e.g., ‘rate’, ‘growth’) or directionality indicators (‘low’, ‘fall’, ‘rise’, etc.). While the sentiment of these items is generally neutral, some may be linked, even if not necessarily, to negative connotations (‘pressure’ or ‘risk’).

4.4. Qualitative analysis