The impact of knowledge recombination on innovation is a topic of ongoing debate. While recombination can foster novelty and broader market appeal, it also risks failure due to the challenges of integrating disparate knowledge categories. This study addresses this tension by proposing that the impact of knowledge spanning (a specific form of recombination) on question appeal is contingent on the questions' knowledge hierarchy (level of abstraction). The authors argue that questions with higher levels of abstraction are less susceptible to the negative effects of excessive knowledge spanning. Existing research often focuses on scientific or technological innovation; this study examines knowledge recombination in the everyday context of a large online knowledge market (Zhihu.com), a setting where ordinary users ask and answer questions, contributing social knowledge rather than purely technical solutions. The free nature of knowledge exchange on Zhihu.com allows for a less constrained study of knowledge spanning, free from the constraints seen in fee-based knowledge markets where financial incentives might skew results. The study aims to investigate the interplay between knowledge spanning and knowledge hierarchy in determining question appeal, measured by the number of followers a question attracts. This research fills a gap by explicitly modeling the knowledge space and its hierarchical structure to understand how these factors interact with knowledge recombination.

The study draws on theories of knowledge recombination, highlighting the work of Schumpeter, Nelson and Winter, and Koestler, who emphasize the role of diverse idea mingling in stimulating creativity and innovation. The concept of "bisociation" is introduced, emphasizing the blending of seemingly incompatible frames of thought. However, the literature also notes the risks associated with recombination, with some research suggesting an inverted U-shaped relationship between the extent of recombination and success. The authors reference work on small-world networks, Bourdieu's field theory of science, and Kuhn's work on the tension between convergent and divergent thinking to contextualize the challenge of balancing tradition and innovation in knowledge creation. The literature review also examines prior research on knowledge recombination in question-and-answer websites, discussing various factors influencing question appeal such as user attributes, content features (level of detail, clarity), and knowledge networks. The study highlights the limitations of previous research in explicitly modeling the knowledge space and the need for new methods to capture the extent of recombination.

The study utilizes data from Zhihu.com, a large Chinese Q&A website. The dataset includes 312,053 questions, each containing information on the number of answers, followers, content, and categories. The authors employ word embedding models (Word2Vec) to quantify knowledge spanning. Each category is represented as a vector in a high-dimensional space, and the distance between category vectors represents the degree of knowledge spanning for a given question. The hierarchical structure of knowledge is captured through Zhihu.com's officially maintained knowledge tree. The distance of each category from the root node in the tree represents the knowledge hierarchy level. The appeal of a question is measured by the logarithm of the number of followers. The authors construct nonlinear regression models to test their hypotheses, including control variables such as question title length and the duration the question remained active. The word embedding approach allows the researchers to represent complex semantic relationships between categories and questions, offering a robust way to model the knowledge space and measure the distance of knowledge spanning. The knowledge tree provides a clear and systematic way to quantify the hierarchical structure of knowledge. To address the skewed distribution of the key variables, logarithmic transformations are applied.

The study's key findings support the following hypotheses: 1. **H1 (Inverted U-shaped relationship between knowledge spanning and question appeal):** There is an inverted U-shaped relationship between knowledge spanning (log-transformed) and the appeal (log-transformed) of questions. The appeal of questions increases up to a certain point, after which the effect reverses. This finding is consistent across different categories of questions on Zhihu.com. The regression analysis shows negative and significant coefficients for the squared term of the log-transformed knowledge spanning variable. 2. **H2 (Positive effect of knowledge hierarchy on question appeal):** Knowledge hierarchy positively influences the appeal of questions. Questions with higher levels of abstraction (located higher in the knowledge tree) tend to attract more followers. The regression analysis shows positive and significant coefficients for the knowledge hierarchy variable. 3. **H3 (Moderating effect of knowledge hierarchy):** The inverted U-shaped relationship between knowledge spanning and question appeal is moderated by knowledge hierarchy. For questions with lower knowledge hierarchy (more concrete questions), the inverted U-shape is more prominent. As knowledge hierarchy increases (more abstract questions), the inverted U-shape weakens and ultimately disappears. This interaction effect is revealed through significant interaction terms in the regression model. The visualization of the interaction effects shows clearly that the inverted U-shape only emerges for questions with a low level of knowledge hierarchy.

The findings support the notion of an optimal level of knowledge spanning for maximizing question appeal. Excessive spanning leads to reduced appeal, while insufficient spanning results in a lack of novelty. The moderating effect of knowledge hierarchy suggests that the tolerance for knowledge spanning is higher for abstract questions compared to concrete questions. This implies that online knowledge markets may naturally be more conducive to knowledge recombination than previously assumed, particularly for higher-level, abstract questions. The authors discuss the implications of this finding for online knowledge market design, suggesting that maintaining a well-structured hierarchy of categories can encourage knowledge spanning by facilitating the creation and discovery of abstract questions. The study also highlights the benefits of using geometric and network perspectives to analyze knowledge recombination, particularly the use of word embedding models to represent the knowledge space.

This study provides substantial evidence for an inverted U-shaped relationship between knowledge spanning and question appeal, moderated by knowledge hierarchy. The findings contribute to theories of knowledge recombination and offer practical implications for the design of online knowledge markets. Future research should explore knowledge spanning in answering questions, integrate hierarchy and similarity measures into a unified model (perhaps using hyperbolic space embeddings), and investigate other knowledge market types (fee-based).

The study is limited to a single online knowledge market (Zhihu.com), which may not be fully generalizable to other platforms. The study focuses on knowledge spanning in question-asking and does not examine knowledge spanning in answers. The measurement of knowledge spanning relies on categorical information and might not fully capture nuances in semantic similarity. Finally, the study assumes that innovation is solely driven by recombination and does not consider other sources of creativity.