We applied methods from corpus linguistics to investigate language variance between three preselected groups with converging interests in tobacco control. The 2 largest transnational tobacco companies, British American Tobacco (BAT) and Philip Morris International (PMI) [26], were selected to sample tobacco industry language. We selected health organizations from the United Kingdom, the United States, and internationally to sample institutional health communication: the National Health Service (NHS), the National Institute for Health and Care Excellence (NICE), the Centers for Disease Control and Prevention (CDC), and the World Health Organization (WHO). These English-language samples matched the English-language balanced corpora we chose: the 27 British National Corpus 2014 (BNC2014) [27,28] and the Corpus of Contemporary American English (COCA) [29]. We used both the written and spoken components of the 2 corpora, spanning genres such as informal speech (transcribed to text), newspapers, fiction, television scripts, and social media. To compile new bodies of language for querying, we constructed detailed search queries for the tobacco industry and health organization corpora.

The time frame for data collection was set between January 1, 2003, and December 31, 2023. The period from 2003 to 2023 was marked by huge upheaval in both the tobacco consumer product landscape and information ecosystems. In May 2003, member states of the WHO adopted the WHO Framework Convention on Tobacco Control. Although it did not enter into force until 2005, the adoption marked a concerted effort “for the promotion of public health” on a global scale. Simultaneously, a consistent consolidation of tobacco companies meant that the tobacco industry had also undergone some major shifts—from a number of small and medium enterprises to just a few global powerhouses [18], of which PMI and BAT are the largest.

The 2 ready-made, balanced corpora of English—BNC2014 and COCA—were chosen for the insights they might offer into the way a “general” English-speaking public talks about tobacco products and tobacco use behaviors. The BNC2014 is a 100-million-word collection representing British English, with texts dating between 2007 and 2020, and the COCA is a 1-billion-word collection representing American English, with texts dating between 1990 and 2019.

For feasibility purposes, the construction of the Tobacco Industry Corpus (TIC) was restricted to BAT and PMI, the 2 largest transnational companies based on 2021 market value [26]. The search parameters for the TIC included web pages and documents published by BAT and PMI during the set time frame. The search query was designed to capture content containing multiple terms related to smoking status and smoking behavior. Asterisks indicate “wildcard” searches, which capture various word endings following a given stem:

smok* OR quit* OR cessation OR switch* OR vap* OR experiment* OR relaps* OR “us* NRT*” OR “dual” OR addict*

Table 1 provides examples of terms captured by different parts of the search query. We conducted three separate searches: (1) a Google site search, (2) using the search function within the two tobacco company websites, and (3) searching web archive tools—the Internet Archive [30] and Archive Today [31]—to collect backdated content not captured in searches one or two. All documents were available in the public domain.

The search parameters for the Health Organizations Corpus (HOC) were webpages and documents published by 4 major health organizations: NICE, NHS, CDC, and WHO. For the HOC, we added the following to the previous search query to refine the content returned to be about tobacco:

AND (tobacco OR smok* OR vap*)

The study conducted two searches: (1) a Google site search of the named organizations and (2) the WHO Institutional Repository for Information Sharing [32]. The study excluded workbooks, research papers, replacement guidelines, and any draft documents.

When compiling both corpora, lists of returned results were saved as structured CSV files containing metadata, which included URLs, webpage headings, and publication dates. This list was manually filtered to exclude job advertisements, pages spotlighting named individuals, and pages or documents in which search terms only appeared in legal disclaimers. Region-specific or geographically local content pages were also excluded, as our focus was on organizational rather than local characterizations. Documents or webpages consisting mainly of figures or videos were also excluded from both the TIC and the HOC for reasons of feasibility. Descriptions of smoking status and smoking behavior were our primary focus, so we manually filtered the data files to ensure documents focusing on individuals (profiles or personal stories) were not included in the corpora. The TIC includes multiple document types directed at different audiences, including investors and consumers. This was intentional, as the aim was to capture the tobacco industry’s discourse at the organizational level rather than to compare communication strategies across internal genres.

After manual filtering and the application of the exclusions listed above, we used Python (Python Software Foundation) scripts to process the CSV files. The scripts included routines for downloading linked PDF documents, extracting and cleaning text from PDF documents using the pdfminer library [33], and exporting web-based content as individual plain text (.txt) files. Annotations of page numbers were included in the files to preserve cross-references to the original PDF documents. The author and publication year were included in the given file names. The sources used to compile the corpora are available via the University of Bath Research Data Archive [34]. Once processed, corpus documents were imported into LancsBox using automatic part-of-speech tagging, with LancsBox (Lancaster University) default settings applied.

Table 2 shows the composition of the 2 tailor-made corpora, the TIC and the HOC.

The data were analyzed using a corpus-assisted discourse approach, which combines quantitative corpus techniques with qualitative discourse interpretation to examine patterns in large collections of text [35]. In practical terms, statistical measures were used to identify potentially meaningful patterns in how people talked about tobacco use status and tobacco use behavior, and then those patterns were inspected in their immediate context to interpret what they were doing in the discourse.

The analysis drew on 2 complementary techniques—collocation analysis and concordance analysis. For collocation analysis, we examined which words occurred unusually often near a target term to identify recurring associations in the corpus. Collocational strength was measured through mutual information using LancsBox (TIC, HOC, and BNC2014) and English-corpora.org (COCA). Collocations were calculated within a window of 5 words to the left and 5 words to the right of the target term; this is the norm in corpus research and has been shown to be productive in corpus-based discourse studies [36]. A minimum raw co-occurrence frequency of 10 was used to keep the number of collocates manageable for detailed qualitative analysis. Common function words (eg, the, and, of) were excluded using the Natural Language Toolkit’s stop-word list [37]. We then used concordancing to interpret these quantitative patterns by viewing every occurrence of a search term in its immediate context, supporting qualitative examination and interpretation of differences and similarities in discourse.

Patient engagement or involvement was not applicable or relevant to the project, as it made use of documents available in the public domain. Institutional ethical clearance was provided by the University of Bath Biomedical Ethics Committee with approval EP18/19 004. The research was conducted in accordance with the guidelines set forth in the Declaration of Helsinki.

The singular form “smoker” and the plural form “smokers” were analyzed separately to assess whether any of the corpora showed strong habits of reference. Table 3 presents the raw and relative frequencies of the terms “smoker*” as a noun and “smok*” as a verb in each corpus.

A chi-square test showed a significant association between corpus and noun form (singular vs plural; n=21,224; χ²₃=1982.01; P<.001; Cramér’s V=0.31), indicating a large effect. Bonferroni-adjusted pairwise comparisons (α=.0083) showed significant differences between all corpus pairs (all P≤.007). HOC had the highest proportion of plural forms, with the largest pairwise difference observed between HOC and COCA (φ=0.35). By contrast, the difference between BNC2014 and COCA, although statistically significant, was negligible in magnitude (φ=0.03). The predominance of plural forms in the HOC constructs smoking as a population-level phenomenon, while comparatively greater use of singular forms elsewhere individualizes the smoker.

The distribution of verb forms also differed significantly across corpora (n=97,453; χ²₉=1173.18; P<.001); however, the effect size was small (Cramér’s V=0.06). TIC showed the highest proportion of the base form “smoke,” whereas HOC favored the past tense “smoked,” consistent with the retrospective orientation of clinical and epidemiological reporting.

To explore how smoking status and behavior were constructed across corpora, we investigated the semantic themes of collocates of “smoker*” as a noun and “smok*” as a verb using the complex queries “smoker* N*” and “smok* V*.” Themes were generated through an iterative examination of concordance lines of the identified collocates, with provisional categories developed, reviewed, and refined throughout the analytic process. Table 4 shows the top collocates of “smoker*N*.”

Across all 4 corpora, the plural term “smokers” was used to describe either a class (“smokers that”; raw frequencies: TIC=32, HOC=21, BNC2014=3, and COCA=19) or a group of individuals (“smokers who”; raw frequencies: TIC=995, HOC=731, BNC2014=22, and COCA=176), with the group classification (“who”) used more often.

Otherwise, overlap in the lexical construction of smokers was limited. Of the 23 specific categorizations of “smoker* N” found among the top collocates, only 7 were shared across all corpora: “former,” “never,” “non,” “current,” “ex,” “heavy,” and “regular,” indicating that most high-ranking descriptors were corpus-specific. Different groups, therefore, rely on largely distinct vocabularies when talking about smokers.

Iterative sorting of collocates into themes permitted the exploration of language patterns that either recurred across corpora or were distinctive to particular organizations. For example, the categorization of some tobacco users as “novice smokers” in the HOC, or the classification of tobacco use behaviors such as “otherwise” in the TIC, as shown in Figure 1.

The HOC and COCA both included the term “novice smoker*,” but this term was exclusive to the WHO in the HOC and found across 3 genres in COCA (fiction, academic, and web). The threshold for this categorization is not clearly described in either case; it is variably used by the WHO, preceded by either “young and” or “new or.”

In the TIC, “otherwise” as a collocate of “smok*V*” was used formulaically in a phrase repeated more or less verbatim from one document to another to talk about a group of people (variably described) “who would otherwise continue to smoke.”

Our analysis highlights a high degree of variation in the language used to describe smoking and smokers among the tobacco industry (BAT and PMI), national and international health bodies and services (WHO, NICE, CDC, and NHS), as well as in samples of general English language use (BNC2014 and COCA).

The HOC included more fine-grained classifications for smoker* N* than the other corpora, but each corpus we queried contained descriptors that were not shared with other corpora. This indicates significant and continuing misalignment in the ways in which different groups talk about tobacco use, in spite of calls for change [11,14,15]. It also provides clear evidence of an ongoing “jingle-jangle” problem noted by previous work [11], highlighting key patterns of language held by individual organizations.

The prominence of partially nonoverlapping vocabularies in our analysis, including the differentiation of plural (smokers) and singular (smoker) forms between the corpora, highlights the progress yet to be made in aligning shared understandings of tobacco use risk and determining structural solutions to overcome tobacco-related harms. Indeed, it is evident that addressing tobacco use as a population-level matter is difficult when only 1 group (HOC) frames it as such.

Frequency analysis highlighted that the plural or singular construction of smoker* as a noun diverged between the general English corpora (BNC2014 and COCA) and the tailor-made corpora: “smokers” were significantly more common in the HOC compared to the TIC, BNC2014, and COCA. The more common reference to smokers in the plural by the HOC, as well as the less frequent use of either “people who” or “people that” (smoke*) across all the corpora, indicates that habits of using person-first phrasing have not yet been established in general English, by health organizations or by the industry [13].

Our results show that BAT and PMI consistently frame smokers as consumers (“adult” and “ASU30”) and as players in legal dramas (“dependents” and “estates”) over dimensions seen in the other corpora (“mothers”). All corpora included categories of behavioral or status descriptions focusing on time or frequency (“never,” “daily,” and “former”). This is an important reminder that tobacco is a commercially driven industry. The characterization of the agency of the business demonstrated a focus on the provision of access to products and information, while the agency of individuals was emphasized in their will to quit or to “continue to smoke.” The characterizations of smoker agency by the TIC were complex, emphasizing the agency of individuals in their phrasing, such as “won’t” and “choose not to,” for example, while deterministically characterizing the smokers’ futures by talking about those who “would otherwise continue to smoke,” presenting continued consumption as both chosen and structurally anticipated.

It was not obvious from our analysis that tobacco use uptake was a central concern for any of the groups. It was reported indirectly by PMI, which noted it as a concern of others regarding its “noncombustible” product portfolio acting as a “gateway to smoking.” Other terms we reported as top collocates, such as “novice,” “early-experiment,” and “early-onset,” may be useful in future work exploring the narrativization of smoking trajectories by different groups and the role of personal agency and discourse in navigating those trajectories.

The high number of product terms in the general English corpora bolsters existing arguments that product variation is a key driver of nebulous understandings of tobacco use behaviors and combustible tobacco products [8]. The prevalence of associated actors and behaviors in the general English corpora also indicates that smokers and smoking are consistently linked to other behaviors, such as drinking, by general audiences. The plural form “smokers” commonly collocates with “drinkers,” which is indicative of a habitual grouping of people according to their behaviors in general English.

Person-first language (people who or that) is not being equally taken up by different groups, calling into question the effectiveness of calls for its use, especially among nonspecialist audiences (those speaking general English).

The uniquely appalling toll the tobacco industry has had on human health is writ large in our collocational analysis of the TIC; key examples include repeated mention of a legal case focused on a “deceased” smoker and the need to categorize “adult” smokers as a distinct use category (which would not be necessary if there were no nonadult smokers). The data also highlight specific categories of “smoker” that may be a key strategic focus for the industry—ASU30—and confirm key areas of concern for the company, such as PMI’s “diminished ability” to convert adult smokers to their “RRPs,” which was listed as a risk in business reports from 2019 until 2022. This “diminished ability” contrasts with the same organization’s reliance on capacity and resource-based framings in their discussions of tobacco-related harm [38] and highlights how access to policy arenas and an ability to shape regulatory policy are key to addressing business risks for transnational tobacco companies.

In a world where even health-focused organizations like the WHO have been subject to critique for their inconsistent and ambiguous language [39], we must endeavor to remain vigilant to linguistic patterns and the epistemic power of organizational actors. The predictability of downstream impacts of product availability and regulation [7,19] may prove elusive given the lack of common or shared phrasing between consumers, policymakers, and tobacco manufacturers, especially in relation to electronic devices [9]. Our results show that these terminological inconsistencies persist across groups, even when the analysis is highly tailored to focus on a single topic (tobacco). These differences are potentially dangerous at both the individual and population levels. They have been shown to lead to misunderstandings of product harm [4,40] and undermine the consumers’ ability to make informed choices about tobacco product use. A powerful tobacco industry can leverage information gaps and misalignments of understanding to its own advantage [41-44] and begin to embed expectations for those who use or regulate tobacco—including expectations about product choice and product use trajectories, for example, in the portrayal of the use of “RRPs” as behaviorally equivalent to smoking.

Our work shows that linguistic clarity and a common understanding of what forms of tobacco use constitute “smoking,” as well as what it means to use tobacco (to be “a smoker”), are lacking between key groups. Given that information-seeking is one of many processes of interpersonal communication that can impact personal health [45], and the recognized association between exposure to tobacco-related media content and tobacco use [46], understanding how flows of information in the public domain interact with (or indeed contradict) one another is paramount to the reduction of structural health inequalities. This includes the reduction of informational access barriers and working collaboratively to overcome imbalances resulting from the distortion of “sensemaking resources” [47], whether by intent or by design.

Public information campaigns are one mechanism by which international-level health organizations attempt to mitigate the deleterious effects of tobacco use, and yet we have noted that these organizations use highly variable language in their communications. The tobacco industry, too, fosters its own characterizations of what smoking is and who smokers are. In the background, we have general English—a multitude of different genres (books, news articles, and so on) where individuals are exposed to different characterizations of tobacco products and their use. By highlighting the linguistic variation that characterizes public-domain discussions of tobacco use across key groups, we hope that our work can contribute to public health efforts to stem flows of misinformation and help disarm a tobacco industry with a history of concealing known harms associated with tobacco products and manipulating policy design and implementation for their own ends.

As our summaries of the corpora show, there was a significant variance in the size of the corpora we deployed in our analyses, and while we endeavored to mitigate the effect of this by reporting relative frequencies in addition to raw frequencies and by foregrounding other relatively adjusted scores such as mutual information, we recognize that this variance impacts the data. We worked to match our corpora to one another as best we could, including in size and date range, but it is important to point out that the dates of our tailored corpora and the prebuilt corpora do not exactly overlap.

We intentionally excluded personal narratives focused on individuals from the TIC corpora to protect the privacy of named individuals, but existing publications show that these can be valuable in exploring personal narratives relating to tobacco use, especially with regard to reducing hermeneutical injustices through the inclusion of different sense-making practices and resources [47].

The inclusion of documents from organizations within a limited time frame is inherently constrained with regard to the carrying forward of research conclusions. All corporate entities are subject to change, including significant internal restructuring, which can involve reductions in the available specialized expertise used to shape public policy. Such changes may pose a threat to expertise-led information provision in key informational spaces available to consumers. Monitoring linguistic change specific to certain disease areas or product landscapes, such as tobacco, could therefore be helpful in shaping the design of population-relevant and impactful public information campaigns that can be safeguarded against such organizational volatility.

Possible valuable future work could seek to query product-specific narratives or focus on the development of narratives in web-based communication to determine how personal technology use, such as information-seeking or dispersed content production, has impacted the information landscape. It would also be valuable to compare discourse across document types within TIC—for example, between customer-facing webpages and more formal corporate or legal reports—to assess how genre and audience influence narrative strategies. Tracing the development of specific terms over time and considering their distribution across genres could also provide insight into the dispersion of industry-constructed language into popular usage and identify key informational asymmetries and key waypoints in their evolution.