Topic evolution analysis describes the emerge, develop, and extinct of topics in a field, which can help researchers understand the history and current situation of the research field. However, the material patent text has a certain domain specificity, and the general entity extraction models cannot extract special entities effectively. Moreover, the belief that topics with high similarity have evolution relationship contradicts the rule of “first the change, then the new topic”, which cannot clearly present the dynamic changes and accumulation of topics. Therefore, we design a method to extract the material performance entities accurately and construct dynamic evolution path for material performance topics. Firstly, we propose a material entity extraction model BERT-BiLSTMCRF, which integrates syntactic dependency analysis and attention mechanism, realizing the accurate extraction of material performance entities. Secondly, we design an algorithm for identifying the evolution relationship between performance nodes based on ring boundaries, which can mine the evolution relationship between performance nodes and existing topics, realizing the dynamic accumulation and change of topics. Finally, we construct the dynamic evolution path of material performance, exploring the complex associations of material performance. Experiments in the field of metal materials confirm that the proposed method can effectively construct the dynamic evolution path of material performance topics, which makes the evolution relationships between topics more abundant and interpretable.
With the emergence of a large number of patents on materials manufacturing and materials innovation, it has become critical to explore the complex associations and evolutionary trends of material performance. Such exploration can help researchers deepen their understanding of material performance and promotes the invention of new materials [1].
Through the review of existing studies, we learned that the material performance mainly depends on their microstructure, while the manufacture process directly affect the material's microstructure. So, current researches perspectives on the evolution of material performance are mainly divided into the three perspectives: "performance evolution - microstructure" , "performance evolution - microstructure - manufacture process" , and "performance evolution - manufacture process". The specific relationships are shown in Figure 1. Among them, Perspectives I and II [2, 3, 4] involve microstructure, require high levels of expertise and experimental equipment from researchers and readers ; Perspectives III [5] usually proceed in the form of controlling variables, such as varying the temperature of a certain process, then exploring the influence of on the microstructure of the material and the evolution of the material performance, which to some extents restricts the comprehensive understanding and description of the material performance evolution.
III affect material performance
I determine material s microstructure
II change manufacture process
This paper takes the material performance as the research object. Firstly, we integrate syntactic dependency analysis and attention mechanism to construct the material entity extraction model (BERTBiLSTM-CRF), we obtaining the performance nodes of each material. And then divided the performance nodes of all materials into time batches by year. After that, we designed an algorithm based on the initial performance topics, to realize the dynamic accumulation of material performance topics and the construction of evolution path.
We considered that to do material evolution, if we collect data at random intervals, it may lead to a lack of completeness and accuracy in the final analysis of the evolution results.So we takes the concept of Germany’s “Industrie 4.0” as the background, and selects metal material as an example, which is one of the key foundational materials closely related to this concept. Then we use the Derwent Innovations Index database as the patent data retrieval platform. The patent search expression is “TS=(‘Metal materials’ OR ‘Metallic materials’ OR ‘Metal alloys’ OR ‘Metal compositions’ OR ‘Metal-based materials’) AND WC=(‘Materials Science’)”, with a time interval from 2011 to 2023, where 2011 is the year when the concept of "Industry 4.0" was first introduced. Then, the top 10,000 relevant patent texts were selected as the dataset. In addition, considering the number of patents in each period, we divide it into year batches for material performance evolution.
Under the background of the continuous improvement of the manufacture process, the processing method of materials is constantly progressing. At the same time, the change in the manufacture process of the material will bring about changes in the material performance [13].
Therefore, we defined the performance entity and manufacture process entity of metal material. In addition, we refer to the relationship shown in Fig. 1, establishing the causal relationship between the two for subsequent analysis. (Among them, the manufacture process entity and causal relationship will be used in our next step of exploring the reasons for performance evolution, so it is rarely involved in this paper.)
Then, considering the content of material patents contains a large number of technical terms, material components, we constructed an entity extraction model (BERT-BiLSTM-CRF) by combining syntactic dependency analysis and attention mechanism. The combined use of these methods provides a more accurate extraction of the material performance entities and manufacture process entities from patent contents, providing a basis for subsequent material analysis and research.
In this part, we first define six evolution types, then we designed an algorithm for identifying the evolution relationship of performance nodes based on ring boundaries. Finally, we present the detailed process of the method for constructing dynamic evolutionary path of material performance topics.
Social circle theory suggests that the social circle formed around a person reflects the closeness of his or her social relationships. That is, a person's intimate social circle usually consists of relationships with a high degree of relevance; then followed by the normal friends circle and the strangers circle. In addition, there may exist such a part of people in the sea of people: they are temporarily outside your normal friends circle, but there are certain similarities between each other, and they may become your friends or even intimate friends in the future, so this paper defines them as potential friends. Therefore, centered on the individual, their affinity rank order is: intimate friends, normal friends, potential friends, strangers, and the position belongs to: within the intimate friends circle, outside the intimate friends circle within the normal friends circle, outside strangers circle within the normal friends circle, outside the strangers circle, specifically as shown in Figure 2. intimate friends circle
This paper defines six evolution types based on existing studies. Among them, the four types of develop, evolve, emerge, and fuse are derived from four different social relationships in the social circle theory, and the two types of extinct and split refer to the existing studies to ensure the diversity of evolution types. In addition, this paper also improves the fuse and split types, by further refining the different contributions of each theme in them, which helps to consider the dynamic interactions between themes in more detail. See Appendix B for details.
We refer to social circle theory and improve the model proposed by Zhang et al. [14], proposing algorithm for identifying the evolution relationship of performance nodes based on ring boundaries, the specific algorithm and its correspondence are shown in Figure 3. Firstly, for the existing performance topics, the centroid of each topic is calculated, and the maximum Euclidean distance between each topic’s patent and its centroid is taken as the topic boundary. The topic boundary is extended outward by a ratio less than 1 to obtain the outer ring boundary and shrunk inward by the same ratio to obtain the inner ring boundary. After several comparison tests, we finally set the ratio in this study to 0.2.
Strangers
Potential Friends
Normal Friends Mutual Friends
Intimate Friends intimate friendscircle normal friendscircle Strangerscircle
Then, hierarchical clustering is introduced to obtain the different types of performance topics in the batch, and merge similar topics that exceed a threshold (we set it to 0.8 in this paper). For a topic in the previous batch, if its number of topics obtained more than two in this batch, the evolution type is considered as split. Furthermore, in the construction of the evolutionary path, if a topic has no evolution relationship with the following topics, we consider it as extinct type.
The construction of the dynamic evolution path of metal material performance mainly includes the following steps, which are shown in Figure 4. Firstly, after the extraction of performance entities of each material, we get the performance node of each material, and then, all material performance nodes are divided into time batches according to the year. Secondly, the K-Means algorithm is used to cluster the first batch of data to obtain the initial performance topics.
Subsequently, for performance nodes in subsequent batches, the algorithm for identifying the evolution relationship of performance nodes (see Section 3.3.3 for the specific algorithmic process) is used to identify their evolution relationships with each performance topic. Then, hierarchical clustering is introduced to obtain the performance topics of different evolution types in this batch, and merge similar topics that exceed a threshold. Finally, incremental iterations are carried out in the above manner to obtain the material performance topics at different year batches, thereby achieving the dynamic construction of the material performance evolution path. (see Appendix D for the entire picture, where Cluster stands for topic) One time slice
Cluster1 Cluster2 Metal Material Patent Abstract Text
In the construction of the evolution path of metal material performance, we use examples from the years 2021-2023 to obtain the results. Specifically, the number of performance clusters in 2021, 2022, and 2023 are 66, 55 and 60. The result of the evolution path are shown in Figure 5, where yellow, red, and green represent the years 2021, 2022, and 2023 respectively. (see Appendix E for the entire picture, where Cluster stands for topic).
This paper proposes an algorithm for identifying the evolution relationship of performance nodes based on ring boundaries, which can not only realize the dynamic accumulation and construction of metal material performance evolution path, but enrich and improve the topic evolutionary analysis method. Currently, we are combining the manufacture process entities of each material to further analyze the causes of the evolution of
Strangers Potential Friends Intimate Friends
fuse Cluster3
Cluster2 only develops from Cluster1 Cluster2 only evolves from Cluster1 Cluster4 is treated as an emerge Cluster which has no evolutionary relationship with previous Clusters. cluster3 comes from cluster1 and cluster2 together. / /
Cluster3 topic boundary
One time slice Cluster1
Cluster2
Metal Material Patent Abstract Text fuse Data flow Cluster3 Cluster1