Data collection

To investigate the commonalities and differences in acknowledgments of Chinese doctoral dissertations, this study constructed a small-scale database comprising dissertation acknowledgments. The database includes 360 acknowledgment sections from doctoral dissertations, with 30 randomly selected samples drawn from each of the 12 major disciplinary categories defined by Chinese higher education. The following points provide further clarification:

All acknowledgment sections were sourced from publicly available data on CNKI (https://www.cnki.net/) for dissertations completed between 2015 and 2021. CNKI is the largest national repository for Mainland China dissertations with standardized metadata and broad institutional coverage. Prior to 2021, authors and their institutions had the discretion to decide whether to upload and make public the acknowledgment sections of doctoral dissertations. As these sections may contain personal information (e.g., names of family or friends), uploading them implies the author’s consent to public disclosure. Starting in 2022, dissertations uploaded to CNKI no longer include acknowledgment sections.

The database was constructed using publicly available CNKI data from 2015 to 2021. For disciplinary classification, this study adopts the Discipline Catalogue for Degree Conferral and Talent Cultivation (2018 revised edition) issued by the Chinese Ministry of Education, which categorizes disciplines into 13 major fields: Philosophy, Economics, Law, Education, Literature, History, Science, Engineering, Agriculture, Medicine, Military Science, Management, and Arts. Due to the limited availability of publicly accessible dissertations in Military Science, this field was excluded from the study, resulting in the inclusion of 30 acknowledgment sections from each of the remaining 12 disciplinary categories.

It should be further noted that the number of dissertations varies across disciplines in the CNKI database. During data collection, we observed that some disciplines, such as History, have significantly fewer doctoral dissertations compared to others, such as Engineering. This discrepancy is attributed to the structure of disciplinary programs in higher education, where natural sciences have notably more doctoral degree-granting institutions than humanities and social sciences. We considered this variation but chose to extract an equal number of samples from each discipline. This decision was made for two primary reasons. First, this study conducts topic modeling on the overall acknowledgments of doctoral dissertations. If sampling were based on disciplinary proportions, acknowledgments from science and engineering fields would significantly outnumber those from humanities and social sciences. Topic modeling is sensitive to data volume, and if data is collected based on disciplinary proportions, it may lead to topics more aligned with science and engineering, thereby overlooking some topics in the humanities and social sciences. Second, the study later employs hierarchical clustering to quantify the similarity of acknowledgment sections across disciplines, which requires a balanced dataset across disciplines. Therefore, extracting an equal number of acknowledgments from each discipline better aligns with the research objectives. We believe this approach does not introduce potential bias to the study results due to the uneven distribution of acknowledgment sections across disciplines in the CNKI database.

Additionally, to ensure representativeness, the selection of dissertations avoided interdisciplinary fields with significant overlap. For each discipline, the most representative subfields were chosen. For instance, interdisciplinary programs such as “Management Science and Engineering,” which can be classified under both Management and Engineering in the Discipline Catalogue for Degree Conferral and Talent Cultivation (2018 revised edition), were excluded. Similarly, for the History discipline, only dissertations explicitly categorized on CNKI under core subfields such as “Archaeology,” “Chinese History,” or “World History” were selected, as studies in intellectual history or literary history could be classified under Philosophy or Literature.

Research methods

This study employs two primary methods to analyze acknowledgment texts: topic modeling and hierarchical clustering.

For topic modeling, we utilize BERTopic [26], which is a topic modeling tool based on natural language processing (NLP), designed to uncover hidden topics from large volumes of text data. It combines the strengths of deep learning and traditional machine learning, leveraging a language model called BERT to understand the meaning of text and then clustering similar texts into topics.

BERTopic is widely used in topic modeling research based on large-scale corpora [27,28]. In the study [27], researchers used Twitter posts as research material to compare four topic modeling methods: LDA, NMF, Top2Vec, and BERTopic. Based on human expert evaluations, NMF and BERTopic were found to perform the best. This is because BERTopic offers two significant advantages over traditional Latent Dirichlet Allocation (LDA) methods. First, BERTopic leverages the BERT model to generate semantically rich document and word embeddings, effectively capturing the contextual meanings of words and producing coherent and intuitive topic representations. In contrast, LDA relies on a bag-of-words model, which struggles to handle polysemy and complex semantics. Second, BERTopic employs HDBSCAN to automatically determine the number of topics, offering greater adaptability compared to LDA, which requires pre-specifying the number of topics. By applying BERTopic, this study can scientifically, efficiently, and intuitively identify the primary topics in dissertation acknowledgments.

We conducted a comprehensive analysis of acknowledgments from doctoral dissertations across 12 disciplines: Philosophy, Economics, Law, Education, Literature, History, Science, Engineering, Agriculture, Medicine, Management, and Art. A total of 360 acknowledgment texts were collected, with 30 samples from each discipline. Initially, the texts were segmented using the Python-based Jieba library for Chinese word tokenization. Subsequently, a stop-word removal process was applied to eliminate punctuation, function words, adverbs, adjectives, and other terms irrelevant to thematic content. Next, each acknowledgment text was vectorized. For this purpose, we employed the open-source acge_text_embedding model from Hugging Face (https://huggingface.co/aspire/acge_text_embedding), which was pre-trained on Chinese corpora and is capable of effectively capturing semantic information from Chinese texts.

For hierarchical clustering, we adopt two approaches:

Vector-based Clustering: Vectorized semantic representations obtained using word embedding techniques can effectively preserve the semantic content of text. As early as 2014, Baroni et al. [29] used a corpus of 2.8 billion tokens in the general English domain to systematically evaluate word vectors generated by four models. These models were tested on 14 benchmark datasets, demonstrating that text vectorization can effectively represent the semantic meaning of text. Therefore, we vectorize each acknowledgment text individually using a language model. For each disciplinary category, the average vector of all acknowledgment texts is computed, and these vector representations are used for hierarchical clustering.

Text Feature-Based Clustering: In addition to vectorizing text and performing hierarchical clustering, we also employed text readability metrics as text features for clustering. This is because, while text vectorization effectively preserves semantic information, high-dimensional vectors are not interpretable. Therefore, we introduced text readability metrics as feature parameters for clustering.

We analyze the text feature of acknowledgments by assessing the readability of Chinese texts, as literary sophistication in Chinese often correlates with reduced readability. This is due to the frequent use of idioms, historical allusions, and classical poetry, which enhance cultural depth and artistic expression but may obscure meaning for general readers due to their specialized, implicit, or context-dependent nature, thereby lowering readability. Lei et al. [30] developed AlphaReadabilityChinese, a tool for measuring the readability of Chinese texts, which evaluates readability across three dimensions—lexical richness, syntactic complexity, and semantic precision—using nine specific indicators. We apply these readability metrics to cluster acknowledgment texts from different disciplines, in order to enhance the reliability of clustering analysis from multiple perspectives.

Primary topics in the acknowledgments of Chinese doctoral dissertations: Recollection, interpersonal relationships, discipline

At the beginning of the article, we proposed Q1: What are the primary topics in the acknowledgments of Chinese doctoral dissertations, and what insights do these topics provide? Topic modeling and analysis of associated topic words reveal three key characteristics of Chinese doctoral dissertation acknowledgments: recollection, interpersonal relationships, and discipline. Below is an introduction to the topic modeling and related analyses.

Topic analysis of doctoral dissertation acknowledgments in China using BERTopic

BERTopic topic modeling involves parameter selection, requiring multiple attempts and adjustments to obtain the best-performing model. When evaluating model quality, some studies [31] use the overall “coherence” of topics as a standard for assessing model performance, calculating the semantic similarity of topic words directly or indirectly to ensure that topic words are not randomly combined but semantically related. Some studies also involve AI as a “judge” [32] or to assist in evaluating the quality of topic modeling [33]. Due to the specificity of the research object, topic modeling of doctoral dissertation acknowledgments may result in cases where a topic is related to a specific discipline, with topic words that are not semantically related but are all terms within that discipline. Therefore, in this study, we did not use the model evaluation methods from previous studies. Instead, we repeatedly adjusted parameters and manually analyzed and evaluated the modeling results.

After extensive parameter tuning, optimal results were achieved by setting the HDBSCAN clustering model’s min_cluster_size to 10 and min_samples to 8, yielding the most coherent and interpretable topic clusters. Under these parameters, the BERTopic model identified nine core topics (Topic 0 to Topic 8). The visualization of the topic feature words and their corresponding weights (displaying only the top five feature words per topic by weight) is presented in Fig 1.

Download:

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

Fig 1. Visualization of the Top Five Topic Feature Words and Their Weights for Topics 0 to 8.

https://doi.org/10.1371/journal.pone.0335035.g001

The topic words derived from the modeling analysis indicate that dissertation acknowledgments primarily encompass three dimensions: research focus, interpersonal relationships, and expressions of gratitude. The nine identified topics exhibit both commonalities and discipline-specific characteristics:

Firstly, many topics reflect distinct disciplinary traits. For instance, keywords in “Topic4” such as “当代” (contemporary), “小说” (novel), “媒介” (media), and “关系” (relationship), and in “Topic5” such as “审议” (deliberation), “法律案” (legal case), “国际私法” (private international law), and “法学” (jurisprudence), as well as “哲学” (philosophy) and “斯宾诺莎” (Spinoza) in “Topic7,” and “主任医师” (chief physician) and “神经内科” (neurology) in “Topic8,” are interwoven within the topics. These keywords suggest that authors frequently reference their academic disciplines and research directions in their acknowledgments, often recalling significant events from their doctoral journey. This is closely tied to the lifestyle of Chinese doctoral students, who typically engage in full-time research (“全日制”), with academic pursuits constituting the central focus of their work and lives during this period. Consequently, whether expressing sentiments, recalling memorable events, or conveying gratitude, the content is imbued with distinct disciplinary characteristics.

Additionally, the keywords in “Topic0” to “Topic2” center around gratitude for interpersonal relationships. Terms such as “先生” (a respectful address for teachers), “师妹” (junior fellow student (female)), and “课题组” (research group) refer to the recipients of gratitude, while terms like “感恩” (gratitude), “温暖” (warmth), “深深” (deeply), “衷心感谢” (heartfelt thanks), and “收获” (gains) directly express the authors’ feelings of appreciation.

Notably, the keyword “评阅” (review) frequently appears in Chinese doctoral dissertation acknowledgments, reflecting gratitude toward anonymous reviewers and professors involved in the defense process, despite the authors’ lack of knowledge about their identities. Although anonymous reviewers may not see these acknowledgments, such expressions are commonplace and can be viewed as a form of ceremonial gratitude or a reflection of the authors’ scholarly decorum.

Furthermore, we calculated the similarity between different topics and visualized the results in the form of a heatmap, as shown in Fig 2.

Download:

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

Fig 2. Heatmap of inter-topic similarity.

https://doi.org/10.1371/journal.pone.0335035.g002

The heatmap reveals variations in the similarity among the topics derived from the modeling process. Notably, higher similarity scores, ranging from 0.86 to 0.89, are observed between “Topic0” and “Topic1,” as well as between “Topic2” and “Topic6.” This may be attributed to these topics’ shared focus on expressions of gratitude. Conversely, lower similarity scores, ranging from 0.73 to 0.74, are found between “Topic1” and “Topic4,” “Topic2” and “Topic5,” and “Topic4” and “Topic5.” These findings suggest that dissertation acknowledgments, as a distinct genre, exhibit certain standardized expressions, such as gratitude toward supervisors or funding support, leading to some commonality across the identified topics. However, disciplinary differences also manifest, with science and engineering disciplines often acknowledging entire research groups, while humanities and social sciences disciplines tend to emphasize solitary textual research and greater interaction with supervisors rather than collaborative project completion.

In summary, based on the topic modeling, topic keywords, and analysis of 360 doctoral dissertation acknowledgments, we identify three key characteristics of Chinese doctoral acknowledgments:

• Recollection of personal experiences during the doctoral journey accompanied by emotional reflections.
• A strong emphasis on expressing gratitude for interpersonal relationships.
• Content marked by pronounced disciplinary characteristics.

The above analysis provides a general overview of the 360 dissertation acknowledgments. In the following sections, we will further explore the specific characteristics of acknowledgments across different disciplinary categories.

Hierarchical clustering analysis of doctoral acknowledgments across disciplines

At the beginning of the article, we proposed Q2: Does the research discipline of doctoral students influence the writing style of their dissertation acknowledgments? In this section, we employed hierarchical clustering and found that dissertation acknowledgments from similar disciplines (e.g., those with similar research methods) are grouped closely together in the clustering results.

Johns and Swales [34] argue that the structure of doctoral dissertations is significantly influenced by the disciplinary field or subfield. They note that dissertations in some natural sciences often adopt an article compilation format, whereas those in the social sciences or humanities typically follow a topic-oriented structure. Intrigued by this observation, we sought to investigate whether doctoral dissertation acknowledgments are similarly shaped by disciplinary influences. To this end, we first conducted a preliminary statistical analysis of the average word count of acknowledgments across different disciplines, as presented in the table below (Table 1):

Download:

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

Table 1. Average Word Count of Doctoral Dissertation Acknowledgments Across Disciplines.

https://doi.org/10.1371/journal.pone.0335035.t001

From the perspective of word count, the top five disciplines with the highest average word counts for doctoral dissertation acknowledgments are Education, Law, Literature, Philosophy, and History. However, the differences in word count are minimal, and it cannot be inferred that the acknowledgment content in these disciplines is more similar based solely on length. For instance, while both poetry and notices may be concise, the former is characterized by literary qualities, whereas the latter is straightforward and direct. Therefore, to further investigate the similarities, we employed vectorization and readability analysis to conduct hierarchical clustering.

Hierarchical clustering results of vectorized acknowledgment texts

Text vectorization is a fundamental technique in natural language processing, as it effectively preserves the semantic content of text while facilitating efficient mathematical computations. By transforming text into high-dimensional vector representations, such as word embeddings or sentence embeddings (in this study, the “aspire/acge_text_embedding” model was used for sentence embedding), the intricate relationships among vocabulary, syntax, and semantics can be captured, thereby retaining the text’s semantic information. Moreover, vectorized data adopts a standardized numerical format, which simplifies tasks such as similarity computation, classification, and clustering. In this study, each doctoral dissertation acknowledgment underwent Chinese word segmentation and stop-word removal. The pre-trained “aspire/acge_text_embedding” model was then employed to convert the processed texts into vector representations. Subsequently, the average vector for all acknowledgment texts within each disciplinary category was calculated. Hierarchical clustering was performed on these categories using the Unweighted Pair Group Method with Arithmetic Mean (UPGMA) as the clustering algorithm, with the average distance between clusters computed. The Euclidean distance metric was applied to measure differences between the embedded vectors, resulting in the hierarchical clustering dendrogram presented below (Fig 3):

Download:

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

Fig 3. Results of Hierarchical Clustering of Vectorized Doctoral Dissertation Acknowledgments.

https://doi.org/10.1371/journal.pone.0335035.g003

Hierarchical clustering analysis reveals that Literature, History, and Education form a closely related cluster. These disciplines are typically classified as humanities. Researchers in these fields often engage extensively with a large volume of literary texts and archival documents. Engineering and Science also cluster closely, reflecting their emphasis on rigorous computation. Similarly, Medicine and Agriculture form a tight cluster, likely due to their shared reliance on extensive experimental procedures. Management and Economics constitute another cluster, consistent with their mutual focus on commercial and financial systems.

The clustering results, derived from vectorizing textual semantic information using a language model, reveal significant internal differences in acknowledgment content across disciplines, which can be broadly categorized into four groups:

Humanities disciplines that emphasize textual analysis and critical reflection;

Social Sciences disciplines that focus on societal and commercial systems;

Science disciplines centered on rigorous computation;

Engineering disciplines that prioritize practical operations.

Furthermore, the analysis reveals considerable similarity in doctoral dissertation acknowledgments within each of these disciplinary categories. This pattern suggests that disciplinary specialization influences researchers, at least in the composition of their dissertation acknowledgments, resulting in a correlation between the similarity of acknowledgment content and the degree of disciplinary proximity.

Results of hierarchical clustering based on acknowledgment readability data

Beyond the influence of disciplinary specialization on acknowledgment content, we also examined whether disciplines shape researchers’ writing styles through clustering based on readability metrics.

Lei et al. [30] proposed a framework for analyzing text readability based on metrics such as Lexical Richness, Syntactic Richness, Noun Semantic Accuracy, Verb Semantic Accuracy, Noun and Verb Semantic Accuracy, Content Word Semantic Accuracy, Noun Semantic Richness, Noun Semantic Clarity, and Noun Semantic Noise. They posited that text readability can quantitatively reflect textual characteristics. In their study, Lei et al. [30] utilized works by Chinese martial arts novelists Jin Yong and Gu Long as source materials, demonstrating that nine readability metrics derived from a Chinese text readability tool effectively clustered the authors’ works into two distinct categories. They further suggested that these metrics exhibit high applicability in fields such as authorship identification. We contend that Chinese text readability analysis tools can substantially capture statistical features of texts. Moreover, employing readability to differentiate the literary quality of texts is highly relevant for Chinese texts, as authors with strong literary competence often incorporate allusions, idioms, and poetic expressions, which tend to reduce readability. Accordingly, this study calculates the average readability values of acknowledgments in doctoral dissertations across various disciplinary categories, using these values as the basis for hierarchical clustering. The results are presented in Fig 4.

Download:

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

Fig 4. Results of Hierarchical Clustering of Doctoral Dissertation Acknowledgments Based on Readability.

https://doi.org/10.1371/journal.pone.0335035.g004

Notably, the disciplines of Literature, History, and Philosophy form a closely related cluster, aligning with the traditional Chinese notion that “Literature, History, and Philosophy are inseparable”(”文史哲不分家”). Economics is the closest discipline to this cluster, likely due to its research often involving political, cultural, and policy-related studies, which share significant overlap with the reading materials of these disciplines. This similarity may, to some extent, influence researchers’ writing styles.

Agriculture and Engineering form another closely related cluster, possibly due to their shared reliance on extensive experimentation. Medicine follows closely, as medical research also involves substantial experimental work. This likely influences scholars’ writing styles, characterized by concise and focused articulation of key points, akin to writing experimental reports.

Overall, the 12 disciplinary categories are grouped into three clusters: natural science and technology disciplines (e.g., Agriculture, Engineering, Medicine); typical humanities disciplines (e.g., Literature, History, Philosophy); and management-oriented disciplines (e.g., Economics, Law, Management). During their doctoral studies, researchers extensively engage with discipline-specific materials (e.g., experimental reports, academic papers, books), which they use to produce scholarly writing. The clustering results suggest that closely related disciplines share significant commonalities. It appears that immersion in a specific discipline influences researchers’ cognitive and writing styles, as evidenced by the distinct disciplinary characteristics observed in their acknowledgment texts.