A Triangulated Digital Approach to News Sentiment Analysis: Insights from Media Coverage of Saudi Women Enlistment in Military Forces

Abstract

This study investigates the emotional tone in international news coverage of Saudi women’s empowerment, with a focus on their recruitment into the military as a milestone reform. The analysis is based on 22 news articles published between 2018 and 2023 across Western, regional Saudi and Arab, and non-Western international media outlets, including coverage from Asian media contexts such as China and India. Drawing on sentiment analysis; the study employed lexicon-based tools (LIWC; Bing; and AFINN) alongside thematic analysis using Speak AI to capture both polarity and narrative framing. This triangulated approach addressed the limitations of word-level sentiment tools by integrating contextual and thematic interpretation. The findings reveal clear regional contrasts: Western media predominantly employed negative framings, emphasizing human rights concerns and ongoing gender inequality. In contrast, regional Saudi and Arab outlets highlighted empowerment, modernization, and Vision 2030 alignment, while non-Western international outlets tended to mirror these positive narratives with limited rights-based critique. Asian media presented mixed framings. These results complicate assumptions of a simple East–West divide by showing convergence between regional and non-Western portrayals. The study contributes methodologically by demonstrating how combining polarity-based sentiment tools with thematic analysis provides a more nuanced account of media sentiment, and substantively by revealing how empowerment narratives are unevenly distributed across global media systems.

1. Introduction

Language offers a window into human psychology, as it reflects emotions, thoughts, and social behavior (Tausczik & Pennebaker, 2009). Linguistic cues embedded in discourse provide valuable insights into underlying cognitive and emotional states (Essam & Abdo, 2021). With the rapid advancement of digital tools, large-scale text analysis has become increasingly precise, enabling researchers to explore psychological and social patterns in news media more systematically.

This study digitally examines journalistic discourse on the recruitment of Saudi women in the military, drawing from local, regional, and international online news outlets. To do so, it adopts a multi-layered approach combining automated sentiment analysis, qualitative interpretation of emotional tone, and AI-assisted thematic analysis to examine how this development—closely tied to women’s empowerment—is framed across media platforms.

Since the launch of Saudi Vision 2030, the government has made major strides in empowering women, generating extensive coverage in national and international media. Anticipating that the recruitment of Saudi women into the military—a landmark achievement—would be portrayed positively, this study focuses on the linguistic feature of emotional tone as reflected in the data. Previous research (e.g., Altohami & Salama, 2019; Anctil Avoine & Lida, 2016) shows that Western media has often framed Arab women through victimization and deprivation. Nonetheless, the unprecedented inclusion of Saudi women in a traditionally hypermasculine domain creates an analytical expectation that international media framing may partially shift toward a more celebratory tone (Wu et al., 2023; Peters et al., 2015; Smith & Luedtke, 2005).

This study is grounded in frameworks addressing evaluative meaning and sentiment polarity in discourse, particularly in media representations of women’s empowerment. Drawing on the Evaluation Theory (Thompson & Hunston, 2000), emotional tone is understood as a positive or negative stance expressed through lexical choices. Framing Theory (Entman, 1993) further informs the comparative analysis by explaining how media outlets selectively emphasize or downplay the perceived positivity of Saudi women’s recruitment into the military. Therefore, the Evaluation Theory (Thompson & Hunston, 2000) reveals how meaning is built linguistically, while the Framing Theory (Entman, 1993) determines the types of constructed meaning.

Since the study aims to examine how the integration of computational sentiment analysis with human-assisted, polarity-based thematic analysis enhances the interpretation of emotional tone in digital news discourse, it accordingly addresses the following research question: What emotional tones underlie newspaper articles on Saudi women’s military recruitment across local, regional, and international outlets?

2. Literature Review

2.1. Sentiment Analysis: Foundations, Applications, and Limitations

Sentiment analysis is a computation that identifies evaluative and emotional orientations in text, which in this instance are commonly operationalized as positive, negative, or neutral with polarity classifications (Gracia et al., 2020). In media studies, it allows for the widespread investigation of affective themes in news discourse. Current sentiment analysis is primarily automated, typically using lexicon-derived tools that assign words to specific affects. These lexicons allow researchers to quickly identify patterns of emotional positioning across corpora (Mohammad, 2016; Van Atteveldt et al., 2021). Despite this efficiency, lexicon-based approaches have well-established limitations. Word polarity is context-dependent especially across domains, genres and sociopolitical settings (Moreo et al., 2012; Wankhade et al., 2022). The literature often suggests that word-level sentiment scoring, when used ad hoc and in isolation from contextual interpretation, has the potential to be erroneous, as is often the case in media discourse which includes irony, modality, and evaluative framing (Kaur & Mohana, 2019; Mohammad, 2016). Even state-of-the-art computational models can return inconsistent or conflicting outputs when processed using the same dataset, highlighting the interpretative difficulties present in sentiment classification (Rodríguez-Ibáñez et al., 2023). In light of these limitations, the contemporary literature has recommended methodological triangulation. Integration of lexicon supported tools could improve robustness, while supplementing automated polarity analysis with qualitative or thematic interpretation provides better insight into narrative framing and contextual meaning (Birjali et al., 2021; AL-Sharuee et al., 2021). This is key in politically and culturally sensitive contexts, with emotional meanings influenced by time, geography, and sociocultural context (Park et al., 2021; Rodríguez-Ibáñez et al., 2023). Accordingly, the current study takes a triangulated approach, using polarity-based sentiment tools and thematic analysis. This design responds directly to the limitations of lexicon-driven sentiment analysis and allows for a more fine-grained analysis of how emotional frames are constructed around empowerment narratives related to Saudi women in various media settings.

2.2. Previous Research on News Sentiment Analysis

Research on news sentiment analysis has evolved along three broad trajectories: early lexicon-based approaches, hybrid and sentence-level refinements, and more recent deep learning and transformer-based models. Researchers have advanced news sentiment analysis through diverse methods and technologies. For instance, Reis et al. (2015) analyzed 69,907 headlines from The New York Times, BBC, Reuters, and Daily Mail using a hybrid approach that integrated iFeel and SentiStrength. Their findings showed that headline polarity strongly influenced readers’ responses. Similarly, Meyer et al. (2017) compared lexicon-based (General Inquirer Lexicon) and machine learning (Bag-of-Words and POS syntactic models) techniques in financial news, reporting greater accuracy from machine learning. These early studies established sentiment analysis as a scalable tool for examining media tone, yet they also revealed persistent challenges related to contextual interpretation and methodological consistency.

Subsequent research refined sentiment detection at different textual levels. Islam et al. (2017) applied sentence-level analysis with a dynamic lexicon, aggregating polarity across sentences to classify articles, while Agarwal et al. (2016) employed SentiWordNet 3.0 to classify headlines based on keyword polarity counts. These developments illustrate a growing recognition that sentiment operates beyond isolated lexical units and requires attention to structural and contextual factors within media texts.

The field has rapidly advanced, with recent research increasingly focusing on sophisticated deep learning and language models. This transition addresses limitations inherent in earlier methods and enables more nuanced, context-aware analysis. The introduction of deep learning architectures, including Convolutional Neural Networks (CNNs) and Long Short-Term Memory (LSTM) networks, represents a significant advancement. For example, Souma et al. (2019) demonstrated the effectiveness of deep learning in analyzing high-frequency news sentiment, while Alasmari et al. (2024) investigated hybrid CNN-LSTM models for specialized sentiment analysis tasks. These methodologies facilitate a more detailed understanding of sentiment dynamics within media texts; however, while deep learning improves contextual sensitivity, it also raises concerns about interpretability and analytical transparency, particularly for discourse-oriented research.

Building on these developments, recent research has shifted toward transformer-based architectures designed to capture semantic nuance at scale. A transformative advancement in sentiment analysis is the emergence of large language models (LLMs) based on transformer architectures, such as BERT (Bidirectional Encoder Representations from Transformers) and its variants. These models are pre-trained on large-scale text corpora, enabling them to capture context, subtlety, and ambiguity in language with high accuracy. Chandra et al. (2025) used LLMs to examine newspaper sentiment during the COVID-19 pandemic and found a higher prevalence of negative emotions in news coverage than on social media. The effectiveness of these models has been substantiated by comparative studies. For instance, Areshey and Mathkour (2024) conducted a comprehensive comparison of transformer models, including BERT, RoBERTa, and DistilBERT, and highlighted their strengths in sentiment classification. Joshy and Sundar (2022) further analyzed these models across diverse datasets, confirming their robustness. Applying these advanced models to large-scale events, such as the COVID-19 pandemic, has yielded important insights into the relationship between media sentiment and public perception.

Many studies show considerable methodological progress, and are mainly divided between highly interpretable lexicon-based models and more opaque deep learning models. Though accuracy in the last few years has improved, the relationship between sentiment classifications and broader narrative framing in news discourse has remained debatable. Lexicon-based tools enable systematic classification with variable results between systems or analytic structures. In contrast, deep learning and transformer-based models improve contextual understanding but raise problems surrounding computational burden, interpretability, and theoretical clarity.

While prior research into news sentiment analysis has evaluated article polarity, research on coupling computational sentiment analysis with discourse-analytic perspectives remains scarce, and none have directly examined women’s empowerment, the core issue of the current study. Moreover, non-Western contexts and gender-focused narratives remain comparatively underexplored in sentiment research. Therefore, this section reviews prior research on the representation of Arab, Muslim, and Saudi women in news discourse.

Research has long shown that Arab and Muslim women are framed through contrasting lenses. Although some studies point to moments of discursive shift in Western representations of Arab women, these changes remain limited and unstable. Visual framing research similarly indicates that representations are not entirely static; Dastgeer and Gade (2016) found that during the Arab Spring Muslim women were frequently depicted as visible political actors, particularly in regional media, suggesting moments of increased agency within global news coverage. Mustafa-Awad and Kirner-Ludwig (2017) observe that during the Arab Spring, Western and German media increasingly framed Arab women as agents of political change; however, this agency was conditional, emerging primarily when women’s actions aligned with Western liberal ideals of protest, emancipation, and public visibility. Postcolonial analyses further demonstrate that such visibility often operates through exceptionalism, whereby Arab and Muslim women are portrayed positively only when they conform to Western feminist expectations, reinforcing a “media-darling” trope within rescue narratives (Muhtaseb, 2020). Building on a critical discourse analytic perspective, Bajri and Alqurayqiri (2022) show how Western reporting frequently constructs Saudi women as subjects in need of external intervention, thereby reproducing ideological power relations. As Edam et al. (2024) argue, apparent shifts often coexist with what they term the “rescue discourse,” within which Arab women are constructed as simultaneously victimized and amenable to Western intervention. This discourse reinforces Western moral authority while marginalizing locally grounded forms of agency and alternative knowledge systems. Rather than situating inequality within political or structural conditions, it individualizes oppression and attributes it to culture itself, thereby reproducing an enduring Orientalist logic.

Studies focusing specifically on Saudi women further illustrate how Orientalism and rescue discourse operate across media systems. Across regional scholarship, empowerment narratives frequently intersect with modernization and reform discourses, particularly in relation to state-led initiatives and expanding public participation. Afzal and Omar (2021), for example, identify predominantly positive empowerment frames in Arab News and the Saudi Gazette during the early Vision 2030 period, though their findings remain constrained by the temporal scope of the dataset. Karolak and Guta (2020) similarly show that Saudi media framed women’s participation in the 2015 municipal elections as a historic milestone, whereas Western outlets foregrounded restriction and exceptionality. Elyas et al. (2021), however, complicate celebratory regional narratives by demonstrating that inclusion within Saudi press remains selective and strategically managed. Such internal complexities, from the Western viewpoint, are often flattened: as Bashatah (2017) finds, British newspapers often adopt frames predicated on conflict-based and Orientalist narratives, and place Saudi women primarily in the audience’s perception as victims rather than as sociocultural actors. These studies combined indicate that framing in agency-oriented narratives is not even handed down equally throughout media systems and is less the result of journalistic practice than of ideology-focused mediation. Recent scholarship has also started to frame sentiment and media framing in the broader discussions regarding representation, ideology, and geopolitical context. Another systematic overview of global media portrayals of Islam and Muslims identifies a number of common and frequently recurring binary oppositions and fear-based narratives that sustain cultural othering and shape public perceptions across Western media systems (Khan et al., 2025). Meta-analytic research further demonstrates that media portrayals of Muslims are predominantly negative and heavily concentrated in Western contexts, with comparative and visually oriented studies remaining relatively limited (Ahmed & Matthes, 2016). Such findings reinforce concerns that dominant research trajectories have privileged conflict-centered narratives while overlooking alternative framing dynamics. A more recent analysis of Saudi media highlights the increasing prominence of empowerment stories, discourse on modernization, and framing of institutional reform in English-language news and visual representations (Al-Qahtani, 2025; AlQahtani & Alghamdi, 2025). The conclusions drawn from these studies indicate that media coverage of Arab and Muslim women is influenced by dynamic ideological, regional, and narrative dynamics, indicating the importance of exploring how sentiment, framing, and discourse come into play in current news coverage.

Overall, this literature suggests both continuity and transformation: Western media often use either critical or rescue-oriented frames, while Saudi outlets are focusing more on the empowerment narratives and aligning them with modernization goals. However different the methods used or the analytical focus, complementary approaches are still necessary. This gap is addressed by the current study through sentiment analysis, where research is focused on the attitude that newspapers create in regard to Saudi women’s empowerment, specifically relating to their military involvement.

3. Methodology

This study employs a multilayered hybrid sentiment analysis within a convergent mixed-methods design, integrating quantitative and qualitative approaches. Grounded in semantic orientation, the framework quantifies and interprets emotional tone in news discourse on Saudi women’s military recruitment. Automated scoring, human-guided interpretation, and AI-driven thematic analysis were combined to ensure both computational precision and interpretive depth.

The analysis unfolded in three stages. First, quantitative sentiment scoring was performed using LIWC, AFINN, and Hu and Liu’s (2004) opinion lexicon (the Bing lexicon), which classify words as positive, negative, or neutral. These scores provided the foundation for subsequent interpretation. Second, excerpts were qualitatively examined through Keywords in Context (KWIC), enabling nuanced exploration of connotations and rhetorical framing often missed by automated tools. Finally, the Speak AI platform generated polarity-based themes, producing structured positive and negative frames for comparison with the human-led KWIC findings and numerical sentiment scores. Speak AI uses Natural Language Processing (NLP) and supervised machine learning to conduct automated thematic and sentiment analysis. It processes text through tokenization and key phrase extraction, then applies trained classifiers to assign probabilistic sentiment labels (positive, neutral, negative). Together, this triangulated design—integrating sentiment metrics, contextual interpretation, and algorithmic thematic generation—addresses the limitations of any single method. As Mohammad (2016) notes, automated text analysis tools often yield non-comparable outcomes; thus, employing multiple sentiment tools alongside qualitative interpretation enhances both validity and reliability. The contextual validation step was conducted by the primary researcher using predefined salience criteria (frequency, evaluative intensity, and discursive prominence). While intercoder reliability was not calculated, the procedure followed explicit operational guidelines to minimize subjectivity. Future research may incorporate multiple coders to enhance reliability.

The following table (

Table 1. Mixed-methods analysis in the study.

Quantitative SA	Qualitative SA
- LIWC - AFINN - Bing lexicon-uploaded to LWIC tool	- Investigating and Verifying KWIC Polarity using LIWC contextualizer tool. - Generating Categorized Themes: Utilizing the Speak AI tool to produce themes categorized as either positive or negative.

) classifies the methods used as quantitative or qualitative.

3.1. LWIC Tool

LIWC was chosen for its capacity to identify emotional tone in journalistic discourse. As a psycholinguistic text analysis tool, it examines emotional, cognitive, and structural components of language through validated dictionaries (Gracia et al., 2020). Its ratings of positive and negative words have been shown to align closely with human judgments (Tausczik & Pennebaker, 2009).

For this study, Hu and Liu’s (2004) Bing lexicon was integrated, adding 6800 entries (2005 positive; 4781 negative). LIWC matches these items against the text to generate percentage values, with “Tone” (0–100 scale) serving as the key indicator: values above 50 reflect positivity, below 50 negativity, and scores near 50 neutrality (cf. Tausczik & Pennebaker, 2009). Positive and negative percentages further clarify the balance of sentiment.

The contextualizer tool, using a KWIC approach, allows verification of word use in context. This step ensures that polarity classifications, particularly in cases of irony or ambiguity, are consistent with rhetorical framing. Interpretation in this study thus draws on both LIWC’s automated scores and contextual validation of key lexical items.

3.2. AFIINN Tool

In addition to LIWC with the Bing lexicon, this study employed AFINN, a widely used lexicon for sentiment analysis developed by Finn Arup Nielsen (2009–2011) to evaluate microblog text. AFINN contains over 3300 English words manually rated for valence on a scale from −5 (negative) to +5 (positive). Unlike many other lexicons, AFINN also accommodates neutral scoring, making it a useful complement to LIWC in classifying sentiment polarity.

3.3. Speak AI

To enhance rigor and validate automated sentiment results, the study also employed Speak AI, a machine learning-based platform for textual analysis. Speak AI generated thematic categories labeled as positive or negative, enabling polarity-based thematic correspondence. This step verified and complemented the categorizations produced by LIWC and AFINN. In particular, the identification of negative themes offered insight into how unfavorable language was contextualized within the articles, reinforcing the human-led KWIC analysis with an additional computational perspective.

In sum, Figure 1 below encapsulates the layered analytical framework of this study, integrating sentiment analysis tools, lexical-semantic measures, and qualitative thematic interpretation into a concise overview of the research design and its sequential steps.

3.4. Research Stages

The research proceeded through several stages. After data collection, text polarity was identified at the document level, and results were evaluated using the selected tools. Human judgment was incorporated to verify classifications within sentence contexts. Verification was conducted by the first author as a corpus-assisted qualitative inspection, conducting close readings of selected instances of KWIC lexicon to see if polarity assigned to the lexicon reflected the narrative framing present within these instances. Because this was an interpretive validation, and not independent categorical coding, inter-rater reliability measures were not considered applicable. After preprocessing the dataset to ensure analytical clarity, removal of photographs, unrelated subtitles, hyperlinks, and non-relevant characters was carried out. These processes minimized the noise of any processing and allowed for more accurate and efficient sentiment classification in the coming analysis.

3.5. Data Collection

The data for this study consist of newspaper articles on Saudi women joining the military, collected from local, regional, and international news agency websites. Access to some outlets (e.g., The Telegraph, Daily Mail) required subscription. Articles were retrieved from GOOGLE by searching keywords such as Saudi, women, and female in combination with military, army, soldier, and recruitment. The inclusion criterion specified articles written in response to news on Saudi women’s military recruitment since the 2018 policy decision. Given the need to include local news in the dataset, the study selected national Saudi news outlets that represent the country internationally by publishing in English. Regarding the selection of Western news outlets, the researchers noticed that this news had attracted the attention of some famous news outlets (e.g., BBC, Bloomberg, Independent, Telegraph, The Daily Mail); therefore, the researchers decided to consider the reputation of news agencies as another selection criterion. The number of articles published by non-Arab Asian news agencies was few, so the researchers decided to include all of them (see Appendix A). In order to enhance data variation and avoid biased selections, the geographic locations of news agencies’ headquarters and their country associations were considered as selection criteria for the published articles (See Appendix A). A total of 26 articles published between 21 February 2018 and 31 December 2023 were initially identified. The search was limited to this period because the issue was at its peak during these years. Links were archived in a separate reference file, and each article was saved in Word format for analysis. Advertisements, links, and photos were removed during preprocessing.

Some items were excluded: the Saudi Gazette article on army eligibility was removed for lacking media framing; and two AlArabiya News reports were merged into one, since analysis was at the document level.

After filtering, the final dataset comprised 22 articles, judged relevant to the study’s focus on news discourse surrounding Saudi women’s recruitment into the military (see Appendix A). The modest size of the corpus is acknowledged as a limitation; however, its theme-focused nature ensures depth and coherence in capturing media framing, making it suitable for discourse-level analysis.

4. Data Analysis and Results

At the document level, articles were categorized as positive, negative, or neutral using text polarity scores generated by LIWC-22 and the Bing lexicon (

Table 2. Comparative sentiment verdicts of news articles using LIWC-22 and Bing lexicon, grouped by country affiliation.

Country	Article	LIWC Tone	LIWC Sentiment	Bing Positive %	Bing Negative %	Bing Sentiment
Saudi Arabia	AlArabya.SA	41.14	Neutral	1.01	0.41	Positive
	Arabnews.SA	63.66	Positive	1.50	0.50	Positive
UAE	Glufnews UAE	72.45	Positive	3.33	0.00	Positive
	Gulf News World UAE	53.96	Positive	1.09	0.36	Positive
	Falir UAE	63.53	Positive	1.37	1.37	Neutral
	The National News UAE	50.59	Neutral	0.56	0.56	Neutral
Bahrain	Gulf Insider Bahrain	53.36	Positive	1.07	0.00	Positive
Oman	The Times of Oman	29.71	Negative	1.40	0.00	Positive
USA	Bloomberg USA	23.22	Negative	0.71	1.66	Negative
UK	Daily Mail UK	20.23	Negative	2.46	2.46	Neutral
	Telegraph UK	16.99	Negative	4.51	2.82	Positive
	The Times UK	6.78	Negative	1.03	2.05	Negative
	BBC UK	36.97	Negative	1.56	1.56	Neutral
	Independent UK	23.46	Negative	2.05	1.53	Positive
	LAD BIBLE UK	20.23	Negative	1.24	2.97	Negative
India	Business Today India	44.11	Neutral	0.96	0.32	Positive
	DNA India	43.67	Neutral	0.59	0.39	Positive
	MINT India	32.15	Negative	0.00	0.00	Neutral
	Jagran Josh India	31.42	Negative	1.79	0.49	Positive
China	Asia Now China	35.11	Negative	1.47	1.26	Positive
Afghanistan	Khamma Afghanistan	60.79	Positive	2.26	0.97	Positive
Bangladesh	Business Standard Bangladesh	15.51	Negative	2.11	2.67	Negative

). Results from the AFINN analysis are presented separately. To ensure reliability, each article was evaluated individually. Building on these outputs, the LIWC contextualizer was employed for corpus-assisted qualitative inspection of keywords in context (KWIC). KWIC inspection was applied selectively rather than exhaustively. Emotionally and evaluatively salient terms were identified from the sentiment outputs, prioritizing high-frequency items alongside words whose assigned polarity appeared contextually ambiguous or potentially misclassified within surrounding discourse. These items were qualitatively examined to assess how lexical polarity aligned with broader narrative framing. Finally, that (KWIC) step was followed by AI-supported thematic categorization (Speak AI) to identify sentiment-grounded themes. Together, these steps integrate quantitative polarity measures with qualitative depth, leading to comparative framing across different news origins.

4.1. LIWC Sentiment Scores (LIWC-22 vs. Bing Lexicons)

First, sentiment analysis was conducted using LIWC-22 and the Bing lexicon, with outcomes presented jointly for comparison (Table 2). The two approaches demonstrated a high level of agreement, especially in Gulf-based coverage. For example, Arabnews.SA and Glufnews UAE were consistently classified as positive by both tools. Some of the identified positive words in Arabnews include ‘empower’, ‘equality’, and ‘expand’, which appear in the expressions ‘equality between women and men’, ‘empower Saudi women’, and ‘expand the role they play’. Similarly, some of the identified positive words in Gulfnews UAE include ‘reform’ and ‘advance’, which appear in ‘Women can now apply’ and ‘introducing reforms that allow Saudi women to advance’.

On the other hand, Bloomberg USA, The Times UK, and LAD Bible UK were jointly identified as negative, including keywords as ‘criticized’, ‘restrictive’, and ‘fighting’. Such alignment highlights convergence between the lexicons on strongly valenced cases. At the same time, divergences appeared in several articles. The Times of Oman and Telegraph UK, for instance, were rated negative by LIWC but positive by Bing, reflecting differences in how the tools weigh lexical distribution and contextual polarity. Neutral cases also emerged, particularly where the balance between positive and negative terms was close. Falir UAE, BBC UK, and Daily Mail UK illustrate this, as they received neutral classifications from one or both tools. Overall, the combined results reveal three main patterns: (1) high agreement across tools in clearly positive or negative cases, (2) divergences in borderline texts, and (3) neutral classifications where sentiment balance was more evenly distributed. These outcomes establish the quantitative foundation for subsequent qualitative analysis of discourse framing (detailed raw output comparisons of LIWC-22 and Bing scores, prior to verdict classification, are provided in Appendix B).

Although both LIWC and the Bing lexicon were run through the LIWC interface, their analyses are methodologically distinct. LIWC’s Tone score is derived from a composite algorithm that incorporates contextual weighting, whereas Bing relies on a direct count of positive and negative words. This distinction explains divergences—such as Telegraph UK being negative in LIWC but positive in Bing—and reflects the tools’ differing sensitivities.

Methodologically, this highlights the importance of lexicon choice in sentiment studies. While Bing offers accessibility and speed, it may overlook nuance that LIWC captures. Such divergences, especially in politically or culturally loaded articles, underscore the need to triangulate tools and ensure analytical transparency.

4.2. AFINN

The AFINN lexicon was employed to triangulate the sentiment findings obtained through LIWC-22 and the Bing lexicon. Articles were categorized according to their cumulative polarity scores: values greater than 0 indicated positive sentiment, values below 0 indicated negative sentiment, and a score of 0 was considered neutral. The results, presented in

Table 3. Sentiment classifications of news articles using the AFINN lexicon, grouped by country affiliation.

Country	Article	AFINN Score	AFINN Sentiment
Saudi Arabia	AlArabya.SA	+9	Positive
	Arabnews.SA	+15	Positive
UAE	Glufnews UAE	+27	Positive
	Gulf News World UAE	+11	Positive
	Falir UAE	+13	Positive
	The National News UAE	+7	Positive
Bahrain	Gulf Insider Bahrain	+6	Positive
Oman	The Times of Oman	+3	Positive
USA	Bloomberg USA	0	Neutral
UK	Daily Mail UK	−23	Negative
	Telegraph UK	−1	Negative
	The Times UK	−5	Negative
	BBC UK	0	Neutral
	Independent UK	−12	Negative
	LAD BIBLE UK	−15	Negative
India	Business Today India	+11	Positive
	DNA India	+13	Positive
	MINT India	+1	Positive
	Josh India	+15	Positive
China	Asia Now China	+3	Positive
Afghanistan	Khamma Afghanistan	+9	Positive
Bangladesh	Business Standard Bangladesh	−43	Negative

, largely align with the earlier LIWC–Bing outcomes. AFINN confirmed the positive orientation of most Gulf and Asian outlets—for example, Arabnews.SA, Glufnews UAE (+27), Khamma Afghanistan, and Jagran Josh India—while reinforcing the negative sentiment of several Western sources, such as The Times UK and Business Standard Bangladesh (−43). A few divergences emerged, most notably Bloomberg USA and BBC UK, which were rated as neutral rather than negative, suggesting a moderating effect in AFINN’s scoring scale.

Taken together, the integration of LIWC, Bing, and AFINN highlights three consistent patterns: (1) Gulf and Asian outlets predominantly express positive sentiment, (2) Western outlets lean negative, and (3) a small set of cases, particularly Bloomberg USA and BBC UK, are neutral (Figure 2). These outcomes strengthen the robustness of the analysis by demonstrating convergence across tools while also showing how different lexicons weigh borderline cases.

4.3. KWIC

To further interrogate the results of the sentiment analysis and explore how emotionally classified words function within their discursive contexts, this step applies a qualitative, KWIC-based reading of selected high-frequency positive and negative terms. The aim is to assess whether such words retain, reinforce, or subvert their assumed polarity when embedded in full narrative structures across news articles.

Following the extraction of both positive and negative keyword lists using the LIWC-22 contextualization tool, the analysis proceeded in two stages. First, keyword lists were filtered by frequency, and each item was examined in its immediate left- and right-hand context. Only a subset of sentiment bearing keywords was examined through KWIC. Selection prioritized high-frequency evaluative terms alongside items whose polarity appeared contextually ambiguous or potentially misclassified by the lexicon outputs. In some cases, this analysis was extended to the full article to capture surrounding sentences and narrative flow. The same procedure was repeated with the Bing lexicon, allowing a comparison across the two tools.

Positive Keywords in Context. Several words identified as positive were employed descriptively rather than evaluatively—for example, criteria and system in administrative contexts, or responsible in procedural discourse. In contrast, terms such as reforms and empowerment reinforced positive framings of women’s recruitment, particularly in Gulf-based outlets, while glamorous carried ironic undertones in Western coverage.

Negative Keywords in Context. Words such as combat and opposition were frequently used in descriptive, factual ways, rather than as evaluative judgments. Other words such as restrictive and controversial, framed regulation or debate without necessarily signaling disapproval. A stronger evaluative force appeared in Western outlets through terms like criticism and controversy, aligning with the more negative sentiment scores seen earlier.

Lexicon Coverage and Divergences. Bing provided broader evaluative coverage (e.g., criminal, burden), often sharpening polarity, whereas LIWC focused more narrowly on affect categories. These distinctions explain some of the divergences noted in the quantitative analysis.

Overall, KWIC re-evaluation demonstrates how polarity scores can oversimplify discursive nuance. By situating sentiment-laden words within their narrative frames, this phase clarifies why certain articles diverged across tools and highlights the interplay between lexical polarity and rhetorical strategy. Full keyword lists and contextual extracts for both LIWC-22 and Bing (positive and negative) are provided in Appendix C.

4.4. AI Sentiment-Based Thematic Categorization

Speak AI was used to generate polarity-oriented thematic outputs for each article. These themes provided the basis for determining overall sentiment by comparing the number of positive and negative themes, with the majority count forming the verdict. Articles with an equal distribution were coded as neutral, while in cases of near parity, a qualitative override was applied when one salient theme outweighed weaker opposing ones. Following this article-level coding, the Speak AI results were examined comparatively across outlets to identify recurring positive themes in local and regional coverage and recurring negative themes in international coverage. A detailed summary of these findings, including theme counts, verdicts, and justifications, is presented in Appendix D, where the table organizes results under the headings Region, Article, Positive Themes (count + nature), Negative Themes (count + nature), Initial Verdict (count-based), Final Verdict (qualitative override applied), and Justification for Final Verdict.

The results from this stage of analysis show that regional media predominantly construct a positive narrative, framing the reforms as a landmark achievement. For example, the analysis of Gulf News (UAE) reveals a strong positive verdict, with six salient themes of “empowerment, modernization, leadership support, recognition, and international standards” decisively outweighing four minor negative themes related to “social resistance and age restrictions.” This pattern, also seen in The National News (UAE), constructs a narrative of opportunity and top-down empowerment, where reforms are presented as gifts from a benevolent leadership.

Non-Western international media, such as Khamma Afghanistan, display a similar orientation, highlighting progress and empowerment while giving limited attention to structural or human rights issues.

In contrast, Western international media—including LAD Bible UK, Telegraph UK, and Independent UK—present a more critical markedly different framing. Coverage in these outlets consistently contextualizes the reforms by foregrounding human rights violations and the suppression of activists. For Telegraph UK and Independent UK identifies dominant negative themes of “human rights concerns, activist repression, tokenism, and the male guardianship system.” In this narrative, positive reforms are not taken at face value but are portrayed as insufficient or a public relations tool to mask ongoing repression, leading to a final “Negative” verdict for these outlets.

Furthermore, treating Western media as a monolithic bloc overlooks significant variations in their editorial lines, a missed opportunity in the initial analysis. While a general trend of skepticism prevails, the degree and focus of criticism differ. For example, the data shows that while the Daily Mail UK’s negative verdict is driven by themes of “repression and/or political motivations” the BBC UK’s negative-leaning verdict stems from themes of “guardianship, symbolic change, and limited roles.” This indicates that narrative strategies are not uniform; some outlets may focus more heavily on overt human rights critiques, while others adopt a more nuanced tone that still implicitly questions the depth of the reforms. This intra-regional variation highlights the complexity of media framing and warrants deeper investigation.

Across the dataset, these divergent narrative strategies are clear. Regional and non-Western reporting emphasizes opportunity, modernization, and economic participation. In contrast, Western sources, despite their internal differences, consistently highlight inequality, tokenism, and the symbolic nature of reforms, thereby constructing a narrative of doubt about the sincerity and substance of the changes.

5. Discussion

Analysis of news articles on Saudi women’s inclusion in the military reveals strong divisions in how various media outlets construct sentiment. Regional and Asian outlets tend to present the issue more positively emphasizing empowerment and modernization, whereas Western media outlets are often concerned with criticism, emphasizing human rights violations and the continued existence of gender inequality. Different sides’ analyses demonstrate how sentiment is not only articulated in individual words but also embedded into broader narrative strategies, underscoring that newspapers play a key role in framing how audiences will perceive an issue through evaluative framing.

The lexicon-based tools (LIWC, AFINN, Bing Lexicon) classified the three emotive orientations as positive, negative, and neutral. There was a clear geographical pattern to be observed: Alarabiya, Arab News, Gulf News, National News, Times of Oman, noted its borderline variation across tools, as shown in the analysis; Business Today and DNA scored very low or zero for negativity; Western outlets such as Bloomberg, Daily Mail, Telegraph, BBC, Independent, and LAD Bible produced much higher negativity scores, with Bloomberg and BBC classified as neutral by AFINN. This division is also evident in recent reviews of coverage in global media and reports, which suggest that representations of Muslims and Arabs continue to be polarized, with negative and critical framings being more likely adopted by Western outlets (Ahmed & Matthes, 2016; Jamil, 2020; Khan et al., 2025). These findings are in line with previous findings that sentiment is geographically patterned and culturally contextual (Dastgeer & Gade, 2016; Park et al., 2021; Rodríguez-Ibáñez et al., 2023). This regional split also aligns with earlier work on Saudi women’s empowerment that examined press portrayals of women’s military enlistment through discourse analysis (Ghobain et al., 2024). Recent studies on Saudi media confirm this trend, as well as how the English-language coverage increasingly engages women’s rights reforms as both a sign of modernization and a challenge to traditional norms (Al-Qahtani, 2025), while the visual analyses show how the portrayal of Saudi women entering into unfamiliar professional arenas increasingly focuses on narratives of empowerment (AlQahtani & Alghamdi, 2025). In a wider sense, these patterns reflect existing analyses of Western rescue discourses of Arab women’s representations (Bajri & Alqurayqiri, 2022) that are further discussed below.

While those polarity scores offer an overview, they do not capture the full implications of how sentiment is constructed through language. Automated tools may unintentionally miss how words are worded with specific meanings: words like “combat,” “opposition” and “prison” can be coded negatively but their evaluative nature depends, as the contextual re-analysis of keywords demonstrates, on their place in a sentence. Contextual interpretation is still important given that lexicon-based systems often struggle with context, irony and discursive nuance (Kiritchenko & Mohammad, 2018, Patil et al., 2023). Kaur and Mohana (2019) show that contextual and metadata-driven approaches improve accuracy and that word level analysis can hide intended meanings. Rodríguez-Ibáñez et al. (2023) also note that various sentiment methodologies applied to the same dataset yield varied results indicating that monitoring, especially sensitive analyses, is important. By employing a combination of several lexicon tools with thematic analysis and researcher interpretation, the present study overcomes the highlighted methodological issues. Whereas lexicon-based outputs provide polarity indicators, thematic clustering through Speak AI and qualitative verification ensures that contextual and discursive nuance is maintained. This triangulation of approaches is consistent with that employed by Birjali et al. (2021), who argue that various sentiment techniques capture distinct dimensions of emotional content, and AL-Sharuee et al. (2021) who demonstrate that multi method approaches incorporating contextual analysis significantly enhance reliability compared to single tool techniques. Through integrating thematic sentiment analysis, this study not only pinpoints polarity differences but it also shows how they are developed in narrative framing.

The thematic analysis conducted through Speak AI revealed a marked contrast in how outlets framed Saudi women’s military participation. Regional and non Western international media emphasized empowerment, modernization, and expanded career opportunities, frequently linking these developments to economic transformation and Vision 2030. This framing aligns with recent analyses of Saudi reform narratives that identify empowerment discourse as a dominant pattern in national media (Afzal & Omar, 2021; Al-Qahtani, 2025; AlQahtani & Alghamdi, 2025), and supports Elyas et al. (2021), who observed that locally based Saudi newspapers adopt an empowerment oriented perspective. Western international outlets, by contrast, foregrounded human rights discourse and the persistence of gender inequality. Even when acknowledging reforms, they tended to frame them as limited, symbolic, or insufficient. This pattern reflects earlier findings on British press coverage that relied heavily on conflict driven and Orientalist frames positioning Saudi women primarily as victims (Bashatah, 2017). It also resonates with broader scholarship demonstrating that Western media frequently employ rescue discourse in representations of Arab women (Bajri & Alqurayqiri, 2022; Muhtaseb, 2020). Muhtaseb (2020), for instance, shows that Arab and Muslim women are positively portrayed in U.S. mainstream outlets only when they conform to Western feminist ideals, reinforcing exceptionalism and reinterpreting empowerment through a culturally critical lens. Similarly, Edam et al. (2024) argue that Western coverage often sensationalizes Arab women’s experiences by emphasizing persecution and cultural conflict. This divergence also contrasts with Mustafa-Awad and Kirner-Ludwig (2017), who found that German and English media during the early Arab Spring portrayed Arab women as active agents of change, suggesting that Western framing is neither uniform nor static but context dependent. At the same time, the present findings align with Karolak and Guta (2020), who observed that Saudi media framed women’s participation as historic progress, whereas Western outlets emphasized restriction and exceptionality.

These competing framings not only vary in style but also reflect the wider discursive processes through which power relationships in global media are reproduced. From a critical discourse analytic perspective, Orientalist logics work through the representation of Western institutions as normative evaluative authorities, which in turn permits a rights-based critique to serve as the dominant interpretive frame and sidelines locally grounded narratives of agency and reform. Rescue discourse emerges through recurrent representational patterns that emphasize deficit, intervention, and cultural exceptionalism, thereby recontextualising empowerment initiatives within familiar hierarchies of global knowledge production. By contrast, regional coverage mobilizes developmental and modernization narratives aligned with national reform agendas such as Vision 2030, framing empowerment as institutional progress rather than cultural rupture. These divergent framings illustrate how sentiment is discursively mediated through ideological positioning, where evaluative language, thematic emphasis, and narrative structure collectively sustain asymmetries in how women’s empowerment in non-Western contexts is interpreted and legitimized.

6. Conclusions

This study examined how local, regional, non-Western, and Western media frame Saudi women’s empowerment through coverage of their recruitment into the military between 2018 and 2023. The findings reveal a clear divergence in sentiment: local, regional and non-Western outlets highlight empowerment, modernization, and opportunity, often aligning with national reform agendas such as Vision 2030, while Western outlets emphasize human rights violations, ongoing gender inequality, and tokenism. A small number of outlets, such as Bloomberg USA and BBC UK, were classified as neutral in some analyses, reflecting borderline variation across tools, but the overall trend was one of Western negativity versus regional positivity. By showing that non-Western international outlets often mirror regional framings rather than Western critiques, the study complicates assumptions of a simple East–West divide.

At a theoretical level, the study confirms the interplay between the two theories framing theory (Entman, 1993) and evaluation theory (Thompson & Hunston, 2000). These theories interpret how meaning, stance and judgment can be constructed in media discourse, combining the macro-discursive level and micro-linguistic mechanisms used for shaping these news frames. Methodologically, the study demonstrates the value of combining lexicon-based sentiment analysis with thematic approaches. Tools such as LIWC, AFINN, and Bing Lexicon provided a quantifiable sense of polarity, but their limitations became evident when words with potentially negative connotations appeared in neutral or even positive contexts. The integration of Speak AI added a thematic layer, capturing the narrative strategies that shaped overall sentiment. This triangulated approach not only reduced the risk of misclassification but also illustrated how computational methods can be combined with contextual interpretation to yield a more nuanced account of media sentiment.

The study had some limitations. It examined 22 articles between 2018 and 2023. The analysis was limited to written texts, without including visual or video materials. The study was limited to English articles, which did not allow us to compare the language used in English and Arabic articles, a point that could be done in future research. However, these limitations did not weaken the study, due to the transparent and enhanced validity of the sentiment analysis. Finally, future linguistic studies may extend this topic to include the semiotic aspects of these articles.