When individuals face an abundance of information, they often selectively choose data that reinforces their existing beliefs, ignoring opposing views and creating an 'information cocoon'. This phenomenon is not limited to social media; it is also relevant in academic circles. This study introduces a novel method for measuring information cocoons in academia from two main perspectives: depth and breadth. We utilised two models, BERTopic and Sentence-BERT, to help quantify the depth and breadth of the study. The results of the study show that the degree of information cocoon in the overall citation network is on a decreasing trend, and the information exchange in academia is gradually open and innovative. Secondly, there are diferences in the information cocoon between disciplines, and disciplines with diferent cocoon sizes have their own characteristics, whose uniqueness and complexity need to be taken into full consideration in the assessment. In addition, the study also found that there is a non-linear pattern between the number of citations of scholarly literature and its information cocoon performance. These results stress the need to understand and address information cocoon dynamics in academia, promoting strategies for a more inclusive and diverse scholarly collaborations.
In the era of big data, the explosive growth and overload
of information have led to increased network dependence,
fragmentation, and selective exposure in people’s
information behavior [
Any environment that generates information is likely to have an information cocoon, including academia. Within this system, scholars’ interaction with information can lead to the formation of an information cocoon. This manifests when researchers excessively consume similar information over time, resulting in issues like information narrowing, group polarization, reduced innovation, and research bottlenecks. This prompts questions: How prevalent is the information cocoon in academia? How can it be measured? And what variations exist among diferent groups?
Previous studies have extensively examined the
formation, impact, and ways to break out of information
cocoons[
Motivated by the existing research gaps, our primary objective is to propose a comprehensive method for measuring the information cocoon within academic environments. Specifically, we aim to quantify the evolutionary changes in the value of information cocoon within academia. To achieve this, we decompose the information cocoon into two key components: research depth and research breadth. In order to accurately quantify these aspects, we utilize BERTopic and sentence-BERT techniques. Furthermore, we intend to analyze the variations in information cocoons across diferent groups, encompassing various disciplines and citation levels.
Our analysis uncovers a downward trend in the value of the information cocoon, accompanied by disparities among diferent groups. These findings provide comprehensive and practical insights into the phenomenon of information cocoon within academia. It serves as a timely reminder for scholars to critically examine their perspectives and take proactive measures to avoid succumbing to the pitfalls of an information cocoon. By doing so, scholars can efectively optimize the information environment within academia for enhanced research outcomes.
In the academic career of scholars, it is crucial to balance
continuous horizontal and vertical development. Horizontal
development allows scholars to cover a wide range of fields,
while vertical development allows them to conduct in-depth
research in specific areas. Focusing only on horizontal
development may lead to a superficial understanding of fields
and a lack of expertise, while pursuing only vertical
development may limit the breadth of knowledge. Therefore,
scholars need to maintain in-depth study of specific fields
throughout their careers, while gaining a broad
understanding of other fields, to enhance their ability to solve complex
problems, foster a spirit of innovation, and promote the
allround development of academic research. Such a balance
not only captures the essence of the problem and provides
insights, but also integrates knowledge from diferent fields
to produce comprehensive and diverse results for academic
research[
Scholars with larger information cocoons tend to perform poorly in terms of depth or breadth of research, as evidenced by their inability to break through to innovation in a particular research direction, or limitations in their research areas.
To this end, evaluating the extent of information cocoons requires a comprehensive consideration of both depth and breadth. Only by simultaneously addressing these two dimensions can researchers better transcend the constraints imposed by information cocoons. Conversely, focusing solely on one dimension or conducting superficial analyses may lead to limitations and misconceptions regarding information. Thus, this paper is grounded in this rationale to devise methodologies and propose corresponding metrics.
Sentence BERT is a modified version of the pre-trained BERT
network that incorporates siamese and triplet network
structures. By leveraging these structures, Sentence-BERT
generates semantically meaningful sentence embeddings that
can be compared using cosine-similarity [
BERTopic is a topic modeling technique that leverages a
pre-trained transformer-based language model to generate
document embeddings. These embeddings are then
clustered, and a class-based TF-IDF process is employed to
generate topic representations[
3.1. Data
For data collection, this study utilized the Semantic Scholar
Open Research Corpus (S2ORC), an extensive dataset
comprising 81.1 million academic papers from diverse
disciplines. The corpus includes comprehensive metadata,
abstracts, and parsed references. S2ORC serves as a
centralized repository that aggregates papers from hundreds of
academic publishers and digital archives, resulting in the
largest publicly available collection of machine-readable
academic text to date [
From the S2ORC dataset, we extracted papers published within the timeframe of 2010 to 2021. The data collection process encompassed capturing various information, including article titles, first authors, reference titles, publication dates, and citation counts. To ensure a suficient level of academic expertise among scholars, we removed duplicate and incomplete entries. Additionally, first authors with fewer than six publications during the specific period were excluded. As a result of this rigorous selection process, our ifnal dataset consists of 107,775 articles.
To measure the research depth, we developed two distinct metrics named "re_depth" and "self_depth". The "re_depth" metric quantifies the research depth by assessing the disparity between the target paper and its reference list. On the other hand, the "self_depth" metric quantifies the research depth by evaluating the distinction between the target paper and previous studies published by the same author.
In the first place, citing relevant literature is of utmost importance for authors, as it allows them to build upon existing knowledge and propose new insights. The disparities in knowledge between their research and the sources they cite indicate the level of innovation within their study. This concept of aggregated knowledge at the topic or field level allows us to observe the macro-scale evolution of knowledge, emphasizing the critical nature of citing behavior. We believe that a greater diference between the target literature and the cited sources reflects a higher level of innovation in the target paper, indicating a deeper level of research. Therefore, we utilize the variance between the target publication and the cited sources as an indication of the research depth within the target paper. To quantify this variance, we utilized the Sentence-BERT model to assess title similarities between a paper and its reference list. The calculation formula is:
Ref_depth = 1 − ∑︀ =1 (, ) (1)
Here, (, ) represents the similarity between the paper and each reference, while n denotes the total number of references to this paper. refers to an article, refers to the i_th reference of .
Furthermore, the depth of research becomes evident through the evolving trajectory and intensity of individual scholars’ pursuits. Each presentation of research findings signifies a continuous journey of self-challenge and breakthrough. Scholars who achieve breakthroughs in research depth often showcase distinctions from prior research. These distinctions can manifest in the exploration of new topics or the acquisition of fresh insights within the same problem domain. To precisely evaluate this depth of inquiry, Sentence-BERT was employed in this study to quantify the divergence of each publication authored by the same individual from their prior research. The dataset was organized accordingly, categorized by author, and arranged chronologically in reverse order of publication. Subsequently, the similarity of each paper to the three papers preceding its publication time was calculated. The calculation formula is as follows: Self_depth = {︃ 1 − ∑︀+=3+1 (, ) , + 3 ≤
3 0, + 3 > (2)
In this context, (, ) represents the similarity between two specific papers authored by the same individual. The variables and correspond to distinct documents authored by the same individual.
To assess the research breadth, we introduced two distinct metrics: "ref_breadth" and "self_breadth". The "ref_breadth" metric quantifies the research breadth by examining the number of topics within the references of the target paper. Conversely, the "self_breadth" metric quantifies the research breadth by evaluating the diversity of topics addressed within the target paper.
Initially, we hypothesized that the number of topics
covered by the references serves as an indicator of the research
breadth within the literature[
Ref_breadth = ref_topic_counts 10 (3) By dividing the ref_topic_counts by 10, we harmonized this value with the scale of other indicators utilized in this study.
Furthermore, the "ref_breadth" metric ofers insights into
whether scholars have explored diverse fields of knowledge
throughout their research endeavors. Through the
utilization of BERTopic modeling, each paper is assigned
probabilities for belonging to various topic groups. In this study,
the Gini coeficient is employed to quantify the breadth of
research interests. The Gini coeficient is a widely used
measure to assess the level of inequality within a dataset
or distribution[
Gini (5) In the formulas, represents the number of papers, denotes the -th paper, and ¯ represents the average value across all papers.
In accordance with our definition of the information cocoon within an academic context, a reduction in both research depth and breadth indicates that an article is confined to a singular aspect of information, thereby increasing the likelihood of information cocooning. Conversely, an expansion in both research depth and breadth implies that an article holds the potential to transcend the information cocoon. Therefore, the expression for the cocoon value is as follows: represents the sum of the four aforementioned indicators.
Cocoon = Avg {(1 −
Mi)} (6)
Our initial objective is to examine the phenomenon of information cocooning within the entire academic environment over time. To achieve this, we calculated the research depth and breadth values on an annual basis, followed by computing the cocoon value for each year. The corresponding findings are illustrated in Figures 1 and 2. Figure 1 demonstrates that the changes in two depth indicators (represented by the blue and orange lines) remain relatively stable, whereas there is a noticeable increase in the “ref_breath" indicator (depicted by the red line). Furthermore, Figure 2 presents the overall cocoon value, which exhibits a decreasing trend over the years. This trend signifies a continuous opening up and innovation of information in the academic environment, reflecting a positive phenomenon. Subsequently, we conducted an investigation into the variability of information cocooning across diferent research ifelds and presented our findings in Figures 3 and 4. To ensure the reliability of our results, we excluded disciplines with limited data and focused on disciplines with larger volumes for analysis. Our analysis reveals notable trends within specific disciplines. In Figure 4, art, economics, and computer science exhibit the lowest levels of information cocooning. This is evident from their higher values in research depth and self_breadth indicators, as depicted by the pink, shallow purple, and dark purple bars in Figure 3. On the other hand, geography, business, and engineering tend to have larger information cocoons, as indicated by their lower research depth and breadth values in Figure 3. Furthermore, disciplines such as education, law, and sociology demonstrate the ability to partially break through the information cocoon, thanks to their relatively higher values on one or more of the four metrics. For example, the discipline of law exhibits a higher "ref_breadth" value, albeit with a lower "ref_depth" value.
These findings emphasize the importance of considering the uniqueness and complexity of each discipline when developing strategies or policies to overcome the information cocoon at the field level. Scholars within each field should also take into account the specific characteristics of their discipline’s information cocoon when designing their research studies.
discipline
Finally, our objective is to examine potential variations among papers at diferent citation levels. To accomplish this, we gathered citation data for each paper in three representative fields: art, education, and geography, which exhibit high-level, mid-level, and low-level information cocoons, respectively. Subsequently, we classified all papers into four groups based on their citation counts: Group A comprises papers with the highest number of citations over 300; Group B consists of papers with citations ranging from 100 to 300; Group C includes papers with citations between 10 and 100; and Group D encompasses papers with the lowest citation class art business computer_science economics education engineering geography law sociology 0.6 count, less than 10. Then, we calculated the research depth value and research breadth value within each group and presented the results in Figures 5 and 6.
The analysis revealed consistent trends in the metric values across papers in groups A, B, C, and D. A clear pattern emerged between the number of citations and the degree of information cocooning. It was observed that the most highly cited papers generally exhibited higher levels of research depth and breadth, indicating their comprehensive exploration of a specific area along with extensive coverage of related domains. In group B, which comprised highly cited papers, there was a focus on academic hotspots, attracting scholars with broad interests; however, the depth of analysis may have been comparatively limited. On the other hand, less-cited papers demonstrate a narrower research breadth but exhibit a significant level of depth. These papers, which often delved into niche issues or possessed a high degree of depth, may have faced challenges in gaining acceptance due to their specialized nature.
In conclusion, our findings suggest that while extensive research can lead to a considerable number of citations, studies that exhibit both depth and breadth tend to have a greater impact. It is crucial to recognize that a lower number of citations does not necessarily imply lower quality. Instead, such papers may possess a high level of depth or focus on niche topics, holding potential for further development. Therefore, instead of solely emphasizing citation counts, evaluating research based on both research depth and breadth can provide more informative insights. Consequently, research depth and breadth can serve as indicators for scholars to reflect upon the information cocoons, as well as for the academic community to assess the influence of research.
In our study, we present an index and methodology for quantifying the scale of information cocoons within academic environments and classify them accordingly. The key findings of this paper can be summarized as follows. Firstly, we observe a gradual breakdown of information cocoon within the overall academic landscape, indicating a trend towards greater comprehensiveness and innovation. Secondly, disparities exist in terms of research depth, breadth, and information cocooning across diferent art
edu discipline geo geo disciplines. Lastly, it is worth noting that while some papers may accumulate citations through diverse research, it is the papers that possess both research depth and breadth that have the potential to truly become influential. Additionally, it is important to consider that papers with a low citation count may be a result of delving deeper into niche topics, rather than indicating lower research quality. Therefore, scholars should adeptly leverage extensive and intricate academic information, continuously evaluating whether their research processes are constrained by information cocoons. Communities can utilize the research depth and breadth metrics proposed in this study to efectively assess the impact of the research.