Topic Modeling Analysis of Mental Health Research on Female Survivors of Domestic Violence in China (2000–2024)

2.1. Data Source and Selection

This study deliberately focused on English-language publications to map the research landscape visible to the global academic and policy community. This scope allowed for an analysis of how China's DV issues are framed in international discourse and enabled a critical evaluation of cross-cultural knowledge transfer.

To construct a comprehensive corpus covering the period from January 1, 2000, to December 31, 2024, a final literature search was conducted in early February 2025 across four major academic databases: PsycInfo, PubMed, Scopus, and Web of Science. The search strategy combined keywords related to DV, mental health, and the Chinese context. The complete search strings for each database are provided in the Supplementary Material.

The initial search yielded 833 records. After removing 344 duplicates, the titles and abstracts of the remaining 489 unique records were screened independently by two reviewers. Studies not directly focused on the intersection of DV, mental health, and the Chinese female population were excluded. Any discrepancies were resolved through discussion, resulting in a final corpus of 179 publications for analysis (Fig. 1).

2.2. Text Preprocessing

The titles and abstracts of the 179 publications were compiled for analysis. A standardized preprocessing pipeline was implemented in R (v. 4.4.5) using the tm package. The process included converting text to lowercase, removing punctuation and numbers, eliminating standard English stopwords, and applying Porter stemming to consolidate word variants [44-53]. Abstracts were chosen over full texts as a deliberate methodological strategy. The goal of this study was to identify the main thematic trends in the research landscape, a task for which abstracts are particularly well-suited. While full texts offer greater coverage, abstracts provide higher information density, focusing on the core concepts and contributions of a study. Furthermore, this approach avoided introducing numerous “noise terms” (e.g., from detailed methodology sections) that may be present in full texts and potentially reduce the interpretability of the resulting topics. As the literature suggests, if the aim is to map a field's primary topics, abstracts are often sufficient [54].

2.3. Topic Modeling

To identify latent topic structures, we employed latent Dirichlet allocation (LDA), an unsupervised probabilistic model widely used for abstract and news corpus analysis [44-56]. The model was trained on the preprocessed corpus using the topicmodels package in R [55-57].

2.3.1. Model Selection

Evaluation was guided by perplexity and log-likelihood metrics [57, 58], aiming to balance model fit and semantic coherence. The selection of the optimal number of topics (K) involved a multi-criteria approach combining quantitative metrics with qualitative validation to ensure model robustness. Quantitative evaluation of log-likelihood and perplexity for models with K=10 to 40 indicated that model performance stabilized within the K=20–30 range, with diminishing returns thereafter (Fig. 2). Within this candidate range, a qualitative inspection of thematic coherence was performed. K=25 was ultimately selected as it provided the optimal balance between thematic granularity and interpretability, producing distinct, coherent topics without the conceptual conflation seen in models with fewer topics or the redundancy of those with more. Beyond statistical evaluation, the interpretability of topics was qualitatively assessed by examining the coherence of representative keywords and their alignment with existing literature on DV and mental health. This combined quantitative-qualitative validation ensured the robustness and theoretical relevance of the final model.

2.4. Temporal Trend Analysis

To examine thematic evolution, we tracked the annual prevalence of each topic from 2000 to 2024. For each year, we calculated the number of documents in which a topic was identified as the dominant theme. This count was referred to as the topic's annual 'instances' or prevalence. Annual frequencies were then visualized by category to identify temporal patterns using the ggplot2 package in R.

3.1. Descriptive Statistics

After text preprocessing, the frequency distribution of key terms was examined. As shown in Fig. (3), the word cloud of the top 50 high-frequency words highlighted dominant terms, such as woman (3.5%), violence (3.24%), partner (1.77%),health (1.63%), andabuse (1.41%). Figure 4 further illustrates the relative frequencies of the top 20 terms, reflecting the field’s lexical emphasis on gender-based violence and its mental health implications.

3.2. Topic Identification and Classification

The LDA model (K = 25) identified 25 latent topics, which were semantically grouped into four overarching categories:

• Core Mental Health Consequences

• Risk, Context, and Protective Factors

• Interventions, Treatments, and Support Systems

• Research Methodology and Policy Perspectives

As shown in Fig. (5), the largest share of topics fell under risk, context, and protective factors (32.00%), followed by interventions and support systems (24.00%) and research methodology and policy perspectives (24.00%). Core mental health consequences accounted for 20.00%. This distribution suggested a shift in research focus from mental health symptoms alone to broader structural determinants and systemic responses. Table 1 provides the full list of topic labels, representative keywords, and their category assignments.

3.3. Topic Trends Over Time (2000-2024)

To examine how topic attention evolved over time, the annual frequency of each topic was tracked from 2000 to 2024. These trends are visualized by category in Figs. (6-9). In these figures, the X-axis represents publication year (2000–2024), and the Y-axis represents the number of documents per year in which each topic appeared as the dominant theme.

As shown in Fig. (6), topics under core mental health consequences generally followed a stable or modestly upward trend. Postpartum depression and suicidal ideation rose markedly between 2019 and 2021, peaking at approximately 35 instances. PTSD and depression also increased steadily, reaching around 28 instances in the same year. Other topics, including marital abuse and anxiety disorders and correlational studies of violent behaviors, displayed fluctuating patterns with local peaks around 2021. Perinatal mental health and suicide risk remained less prevalent, peaking at around 10 instances.

In Fig. (7), topics under risk, context, and protective factors showed the most pronounced growth. Substance abuse and sexual violence/satisfaction emerged as the most frequently identified topic in 2021, with approximately 50 instances. Long-term effects of adverse childhood experiences (ACEs) and social support and help-seeking behaviors also saw substantial increases, peaking at about 40 and 35 instances, respectively. Topics related to immigrant populations, family stress, and cumulative vulnerability demonstrated consistent upward movement, particularly after 2015.

Figure 8 displays the trend for interventions, treatments, and support systems. This category reflects growing diversity and an increasing focus on digital and nontraditional approaches. Symptomatology and alternative therapy exploration and web-based health management and treatment experienced significant growth after 2018, peaking at around 40 and 25 instances in 2021. Community intervention and media advocacy reached a local peak in 2019, while interventions and screening tools showed steady gains over the two-decade span.

As shown in Fig. (9), topics under research methodology and policy perspectives followed a pattern aligned with major policy milestones. The application of quantitative analysis and statistical models rose sharply after 2015, peaking at approximately 40 instances in 2021. Policy implications from the Chinese experience saw major surges around 2016 and again in 2021, potentially reflecting the impact of China’s Anti-Domestic Violence Law. Other topics, such as assessment tools and clinical practice guidelines, experimental methods, and questionnaire-based resilience measurement, remained moderately represented, typically ranging between 15 and 25 instances. Database-driven evaluation and substance abuse showed local peaks in 2015 and 2020.

4.1. Thematic Evolution: A Paradigm Shift toward Structural Determinants

To fulfill the promise of a detailed theoretical interpretation, we analyzed the distribution of the 25 identified topics across the four levels of the social-ecological model (Table 2). This analysis revealed a substantive paradigm shift, moving from a primary focus on the individual level toward a more comprehensive, multi-level understanding.

As Table 2 illustrates, the research landscape has evolved significantly. Early scholarship predominantly documented individual-level psychological sequelae, including depression, anxiety, and PTSD, establishing prevalence but offering limited insight into causal mechanisms. From 2016 onward, however, thematic complexity increased markedly, reflecting a transition toward upstream social determinants aligned with the social-ecological model [31, 32]. The growth in research on relational factors is particularly evident. For instance, increased attention to the long-term effects of adverse childhood experiences (ACEs) [59-62], family relational stress [30, 63], and migration-related vulnerabilities exemplifies a growing recognition that individual-level trauma is embedded within broader structural contexts [64-67].

This acceleration coincides with China's Anti-Domestic Violence Law [35]. While topic modeling cannot establish causation, several mechanisms plausibly explain this pattern: landmark legislation legitimizing DV as a research priority, unlocking funding streams [38], and facilitating institutional access for researchers. Alternative explanations include the global #MeToo movement [68], China's participation in the WHO Global Plan of Action [69], and methodological advancements in trauma research. Disentangling these influences requires citation network analysis and systematic policy document reviews, which are beyond the scope of the present study.

From a social-ecological perspective, this thematic evolution revealed growing attention to relational (marital conflict, family dysfunction) and community-level factors (social support networks, migration stress). While topics at the societal level related to policy and research methodology have accelerated, other crucial structural determinants, such as gender norms, economic inequality, and the detailed functioning of the judicial system, remain analytically underexplored. This international comparison further highlighted a key divergence: the broader international literature allocates substantial attention to perpetrator psychology and batterer intervention programs [70, 71], whereas our corpus detected minimal offender-focused research, reflecting a victim-centric bias consistent with other East Asian contexts [42].

4.2. Critical Research Gaps and Underlying Mechanisms

Three major omissions constrain the field's capacity to inform prevention and intervention.

4.3. Translating Evidence into Policy and Practice

The findings highlight three evidence-informed priorities for improving DV responses in China. First, integrating mental health screening into DV service pathways is essential. PTSD, depression, and anxiety collectively accounted for 12% of the research corpus, yet most shelters and legal aid centers still lack trained professionals and standardized screening procedures [81]. Embedding validated mental health assessments at intake and referral points, alongside training staff in trauma-informed communication, could enable early identification and more effective referral for survivors. Second, the lack of perpetrator-focused research mirrors a policy infrastructure void [38-40]. Pilot programs adapted to the Chinese context and systematically evaluated with evidence-based frameworks are necessary to develop national standards for offender intervention. Third, attention should be directed toward high-risk populations, such as migrants, pregnant women, and rural residents. Although research interest in these groups has increased, targeted service models remain limited. Integrating DV screening into prenatal care, deploying community health workers in migrant-dense areas, and establishing elder abuse task forces may address current service disparities.

4.4. Methodological Contributions and Limitations

The methodological contributions of this study include several key strengths. First, the use of LDA enabled scalable analysis of 179 abstracts, providing comprehensive coverage of the literature. Second, the initial extraction process was largely objective, minimizing confirmation bias inherent in manual reviews [82]. Third, LDA allowed precise temporal tracking, revealing patterns, such as thematic shifts, following 2016. Finally, the approach facilitated interdisciplinary synthesis by linking concepts across domains, for example, connecting adverse childhood experiences with adult DV research. Topic modeling in this context complements, rather than replaces, traditional literature reviews, serving as a valuable tool for preliminary scoping before in-depth synthesis.

Despite these strengths, several limitations should be acknowledged. Restricting the analysis to English-language publications may have introduced topical, methodological, and cultural framing biases [35, 83, 84]. Relying solely on abstracts may have limited access to detailed methods, contextual information, and negative findings, prioritizing breadth over depth [50]. Finally, the analysis has been constrained by both technical and interpretive challenges. On a technical level, the preprocessing pipeline, while robust, occasionally yielded source-specific artifacts (e.g., 'pubm', 'cus') stemming from database formatting; their presence highlighted the challenge of balancing computational efficiency with semantic fidelity in natural language processing for social science research [85, 86]. On an interpretive level, topic labeling involves subjective judgment. Both challenges were mitigated through our systematic dual-author review process, where such artifacts were manually excluded, and a consensus on topic labels was reached.