A calculation method for the development trend of subject keywords

Abstract

SIGNIFICANCE:

The complex network structure information such as citation network and co-authored network provides rich structural information for subject trend research, thus research on deep mining of existing static structure information to obtain keyword trend receives less attention. Both dynamic method for the extraction of domain-specific common word set and the method for calculating keywords dynamic characteristic (DC value) proposed in this paper can be used as basic methods and parameters for subject trend research.

METHOD:

Employed mature and widely used TF-IDF algorithm as the word frequency calculation tool, this paper use stable static structural information such as publishing date and journal name to propose a method of calculating the dynamic characteristic of subject keywords based on stable structure such as time, journal name and domain-specific common word set.

RESULT:

The experimental results show that the keywords’ DC value is positively correlated to the change trend of keywords based on time.

Keywords

Domain-specific common word sets dynamic characters of keywords subject trend TF-IDF

1. Introduction

In order to learn such knowledge as background information, cutting-edge researches and guidelines for the R&D of enterprises in certain fields, it is of great significance for scholars to study the research trend and development in such fields [1]. Many relevant valuable achievements have been reached by scholars worldwide, among which two main research approaches are mainly adopted, namely, the analysis method based on citation relationship and the analysis method based on word frequency [2]. And this paper focuses on the latter.

Technically, word frequency analysis is a basic technology for text content mining. Sun [3] pointed out in his doctoral thesis that as for the improvement of technological efficiency related to text content mining in specific situations, it is feasible to integrate more network structure information related to the text. There is abundant related network structure information in academic literatures, e.g. publishing date, journal name, authors (co-authorship network), citation relationships, word co-occurrence networks, and so on. Many research achievements resulted from the integration of multiple structure information. For example, the famous CiteSpace is the software based on integrating of word frequency analysis and structure information including time order and citation relationship [4]. It is worth noting that there are huge differences in the complexity of the structure information of different network structures. The structures like citation network, co-authorship network, and word co-occurrence network belong to complex network structure. Although researchers have already had some knowledge of such structural information, there are still many problems left for exploration, for example, the accuracy of overlapping community detection (the collection of keywords for judging subject classification), the communication characteristics of different communities, and the pattern of information communication through multiple communities and so on. Therefore, the introduction of such complex structural information into the research of subject development trends will inevitably bring more complexity to the latter’s research, thereby increasing the uncertainty of the latter research and hindering the judgment of its research results and methods to optimize research. However, the most specific and stable structural information related to the literature is publishing date and journal names, which are not influenced by other factors. Meanwhile, current researches only see them as background structure information and thus their potential is not fully tapped.

This paper propose a calculation method to analyze the development trend of subject keywords. Based on date, journal name and domain-specific common word set, a TF-IDF [5]-based word frequency calculation tool is used to calculate the dynamic characteristic of keywords. The experimental results shows that the dynamic values of the keywords calculated by this method are clear and reasonable and could specifically indicate the development trend of the keywords in corresponding field. And since the dynamic values are stable, they can be used as a basic parameter in other methods to further data analysis, such as cluster analysis.

2. Literature review

Based on different research ideas, two main research approaches are gradually developed in research on subject development trends.

The research approach based on citation relationship mainly studies the differences between basic and cutting-edge knowledge. Persson [6] deems the citing essays are mainly about the updated research and the cited essays constitutes the knowledge base. This research approach emphasizes the development of cutting-edge researches that base on basic knowledge. Many methods are widely used to reconstruct knowledge base such as bibliometrics, cited essay clustering. Through the visual image focusing on the its timeline, cutting-edge research is defined by Morris et al. [7] as a large number of literatures that are continuously referenced by a fixed set of advanced essays; similar work is also provided by Small’s thematic series of web services [8]; But the new research trends are explored based on the basic knowledge through many approaches as statistics analysis and text analysis. CiteSpace, the masterpiece of Chen, is an important achievement resulted from the practice of this research approach. Many ideas based on this research approach are comprehensively reviewed some work in his the essays [4].

The research approach centered on word frequency analysis is technically based on text content mining. As one of the important part of data mining, text content mining has a wide range of applications, such as text classification [9, 10], public opinion monitoring [11, 12], knowledge discovery [13, 14], information recommendation [12, 15] and so on. And methods like word frequency analysis, word co-occurrence analysis and clustering that are often adopted in field of library and information science, are basic in text content mining. And they are mainly applied to analysis such as hot spot discovery and popularity calculation. In recent years, text mining models such as TF-IDF [16, 17] and word vectors [18] have also been introduced into the field of library and information science, among which TF-IDF algorithm is a commonly used weighting method in the field of information retrieval. It counts the frequency of a word in an article to indicate the importance of the word to the article, and adjusts weights of the word by calculating the logarithms of the frequency of the same word in other articles. That is, the weight of a word is directly proportional to its frequency in the article and is inversely proportional to its frequency in other articles. Word frequency refers to how many times a word appears in a document, and is generally divided by the maximum word frequency from all documents for normalization to unite the order of the word frequency and prevent the weight of words from being partial towards long documents. Used to calculate the weight of the term in the document set, TF-IDF has such characteristics as intuitive model and remarkable effect, and has been widely applied in researches such as search engines, thus it is deemed as a mature word frequency analysis tool [17]. While TF-IDF focuses on structural operation on words, word vectors [19, 20] focus on the semantic expression of words and their semantic computability. By mapping the terms in the text set into an N-dimensional space, the word vector model solves two basic problems of semantic computing: a) Converting unstructured text data into values with the same length; b) Converting semantic computing into the computation of space and distance. Words that are close as values in corresponding space of words are considered to be similar in meaning. The word vector model provides more possibilities for solving many problems of NLP (Natural Language Processing). For example, the Word2Vec algorithm proposed by Mikolov et al. [19, 20] in 2013 is not only computationally efficient, but also makes the vocabulary computable at the semantic level. For example, vector (‘Berlin’) – vector (‘Germany’) $+$ vector (‘England’) $\approx$ vector (‘London’). According to the bibliometric research by Wang et al. [21] on domestic research methods in the field of library and information science, the research development trend is mainly based on the computation of the word structure, but less on the semantics.

These two different research approaches are applicable to different situations. For example, the research method based on citation relationship analysis is more likely to be used in research of the development trend based on knowledge base and its path, while the method based on text mining analysis is more likely to be used in discovery of hotspots or popularity in a particular text set. In addition, as mentioned above, the effect of the text mining is closely related to the network structure information of the text. The introduction of complex network structure information such as citation network and co-network into the research of the discipline has to some extent added complexity and uncertainty for the research to the problem; While clear and stable structural information such as publishing date and journal name can provide clear and solid structural information for research.

Based on the publishing date and journal name, the two stable parameters, this paper uses TF-IDF as the word frequency calculation tool to calculate the weight of the key words namely TF-IDF values in the document set to propose a method to extract domain-specific common word set of high-frequency like sliding window and to study the dynamic development of the key words in the word set based on timeline, that is the keyword DC (dynamic character) value.

3. Research goals and thoughts

The “keywords” in this paper refer to the actual words obtained after the segmentation of the title and abstract, not the keywords in the literature summary. This article contains two closely related research goals:

(1)
Dynamic extraction of domain-specific common word sets: Extract time-independent common keywords of high frequency based on the attribute “journal name” the literature is published on. Usually corresponding to research hotspots in specific research field, research on common keywords of high-frequency is a basic method to study research hotspots and development in certain fields. This paper tends to solve the problem of how to dynamically determine common high-frequency keywords according to the characteristics of the data set itself.
(2)
To calculate the dynamic characters of common high-frequency keywords is to study the trend developed based on timeline. The transfer of research hotspots will be reflected in the time-based change of keywords; while the use of algorithms to study the dynamic characters of common high-frequency keywords is to record the changes in corresponding research hotspots.

Based on the research goals above, the procedures to extract domain-specific common word and to calculate dynamic characters of keywords are as follows (Experiment details in 4.2):

(1)
Perform data cleaning and basic NLP processing on the original data to obtain a data set including publishing date, journal number and NLP processed text.
(2)
Divide the data set based on the journal names and calculate the TF-IDF value of each journal’s keywords.
(3)
Extract domain-specific common keywords of high frequency from journals through sliding window mechanism and score the keywords with weighted average method; the set of scoring standard tends to further identify frequency of keywords being used in the entire data space. The higher the frequency is, the higher the score is; the time-independent common keywords of high frequency will be used as the domain-specific common word set and as the reference word set for the annual selection of common keyword of high frequency for the following dynamic character calculation of key words.
(4)
Divide the data set according to the publishing year and extract the annual common keywords of high frequency according to the method adopted above.
(5)
Match the annual common keywords of high frequency obtained in the previous step with the domain-specific common word set as a reference to obtain the time-based change value of each keyword.
(6)
Calculate the dynamic character DC value of each time-based keyword according to the keyword dynamic character calculation formula proposed in this essay.

4. Data set and experimental process

4.1 Data set

The data in this paper, downloaded from WOS [22] (Web of Science), are collected from the summary of the academic papers published in 12 internationally renowned journals in the field of education informationization within 11 years from 2007 to 2017. After preliminary processing, a data set of summary from 8131 papers is collected, including the title, summary, author, and publishing date of the paper; the publishing date and journal number are used to divide the data set, and the title and summary are contents to be analyzed. The 12 journals are shown in Table 1.

Table 1
Journal list of reference

SN	Journal name	Factors/factors from 5 years	Index
J01	Computers & Education	3.81/5.047	SCIE/SSCI/EI
J02	Educational Technology & Society	1.584/2.034	SSCI
J03	Computer Applications in Engineering Education	0.694/–	SCIE /EI
J04	IEEE Transactions on Education	1.727/–	SCIE /EI
J05	Educational and Psychological Measurement	1.548/–	SCIE/SSCI
J06	Etr&D-Educational Technology Research and Development	0.725/–	SSCI
J07	Journal of Educational Computing Research	0.678/–	SSCI
J08	Journal of Baltic Science Education	0.448/0.412	SSCI
J09	Journal of Engineering Education	3.047/–	SCIE /EI
J10	Internet and Higher Education	4.238/5.130	SSCI/EI
J11	International Journal of Electrical Engineering Education	0.375/–	SCIE /EI
J12	Technology Pedagogy and Education	1.066/–	SSCI

In view of the main purpose of this paper is to explain the method of extracting domain-specific common word set and corresponding method for calculating the dynamic characters of keywords, the requirements of journals are not as strict as those required by traditional bibliometric research. The proportion of journal types and the amount of literatures in Table 1 is as follows: high-cited journals (J01/J09/J10), accounting for 37.9%; common core journals (J02/J04/J05/J12), accounting for 31.5%; general journal (J03/J06/J07/J08/J11), accounting for 30.6%.

4.2 Experiment process

Based on the approach above, the experiment consists of three parts: data preprocessing, extraction and scoring of common high-frequency keywords of, and calculation of keyword DC values.

4.2.1 Data preprocessing

Preprocessing includes conventional data cleaning, NLP processing, and post-slicing word frequency calculation, among which post-slicing word frequency calculation is of high importance. Conventional data cleaning is to check if there are data errors and omissions and then correct them. NLP processing mainly refers to the processing of post-cleaning data such as word segmentation, lemmatization, and stop token filter; and then store the keyword and its publishing year and journal information in the database. The post-slicing word frequency calculation includes two procedures, firstly make slice, and then calculate word frequency of the sliced data set. Two kinds of slicing procedures are required and the corresponding word frequency calculation is performed with the TF-IDF algorithm to respectively acquire two data subsets for subsequent calculation:

•
D1: Sliced in accordance with “publishing journals”, each data set is processed with the TF-IDF algorithm to obtain the TF-IDF value of keywords in each journal. This data set is arranged in reverse order according to the TF-IDF value of each journal’s keywords. D1 is used to extract domain-specific common word sets. Partial results are shown in Fig. 1;
•
D2: The data set is sliced according to the “Publishing Year”. The sliced annual subsets are processed with the method in D1, that the TF-IDF values of each journal are calculated separately and arranged in reverse order; Then the annual ftidf values of keywords in each journal is sorted in reverse order, that is, each year, a result similar to that of Fig. 1 is obtained; therefore, D2 contains 11 annual journal keywords TF-IDF values in reverse order. It will be used for subsequent keyword dynamic trend calculations.

Figure 1.
TF-IDF values of each journal’s keywords in reverse order (partial).

4.2.2 Extracting domain-specific common words of high-frequency

This essay uses a dynamic mechanism similar to a sliding window to determine a common high-frequency keyword set; each keyword is scored. This section contains D1 data set processing. The experimental steps are as follows.

4.2.2.1. Extraction method for common word set with sliding window mechanism

•
Let the first $n$ keywords of each journal in D1 be high-frequency keywords; and define the set of the first $n$ keywords of all journals as $T$ ; since some keywords appear in multiple journals, $|T|\leqslant 12n$ ;
•
Assume $N_{k}\in T$ , and these $|N_{k}|$ keywords are used commonly by $k\in[1,12]$ or more journals, that is, there are $|N_{k}|$ common high-frequency keywords in set $T$ ; Obviously, when $k=$ 1, $N_{k=1}=T$ . The $k$ here can be chosen based on the purpose or experience of the analysis. The constraint relationship of above parameters can be seen in Eq. (1), where $A$ calculates the ratio between the common high frequency keyword $|N_{k}|$ and the total high frequency keyword $|T|$ , and $B$ is the ratio between the election factor $n$ and the common high frequency keyword $|N_{k}|$ , and $C$ is the result of $A$ plus $B$ . Without loss of generality, in this experiment, $k=1275\%=$ 9, that is, the common high-frequency keyword set used by 9 or more journals is $N_{k=9}$ :

$\displaystyle C_{k}=A_{k}+B_{k}=\frac{|N_{k}|}{|T|}+\frac{n}{|N_{k}|}=\frac{|N% _{k}|^{2}+n|T|}{|T||N_{k}|},\quad k\in[1,12],A_{k}=\frac{|N_{k}|}{|T|},B_{k}% =\frac{n}{|N_{k}|}$ (1)
•
It must be noted that the parameters in Eq. (1), ie, $n$ , $T$ , $N_{k}$ are not purely mathematical, but are closely related to the characters of the data set and the subject. The number of high-frequency keywords in each subject data set may vary greatly, so the calculation must base on the actual results of keyword extraction. For the D1 data set, try $n=$ 100, 200, 300, 400, 500, 1000, 1500, 2000 respectively, and get the result shown in Fig. 2, in which horizontal axis is $n$ , the left vertical axis is the corresponding value of $A$ , $B$ and the left vertical axis is for value of $C$ .
•
Figure 2 provides a clear visual explanation for selecting common high-frequency keywords: when $n=$ 1000, the $C$ value is the lowest, which means that when $n>$ 1000, the acceleration of common high-frequency keywords $|N_{k}|$ decreases faster, which makes it more difficult to capture and describe common research hotspots in the field. Therefore, for the current data set D1, $n=$ 1000 is the optimal choice. And at this time $|T|=$ 3538, $|N_{k=9}|=$ 1586, that is, the domain-specific common word set contains 1586 keywords.

Figure 2.
Testing results of the chosen factors of common high-frequency keywords in data set D1.

4.2.2.2. Overall scores OR and ORE of common high-frequency keywords

The selected $|N_{k=9}|$ common high frequency keywords are scored OR (Overall Rating) – according to the TF-IDF value of each keyword in each journal, the specific process is as follows:

(1)
Count the NZA (Non-zero average) and NZC (Non-zero count) of each keyword. NZA is the $t_{\text{th}}$ keyword’s average value of the TF-IDF non-zero value in such 12 journals as shown in Fig. 1. Similarly, NZC is the number of keyword whose TF-IDF value is not zero in 12 journals. For example, the keyword “engineering” shown in Fig. 3, whose NZA $=$ 0.128 and NZC $=$ 10. Comparing with the non-zero numbers of surrounding keywords 12, “engineering” has a low level of public participation, so the keyword is more likely to have a lower scoring than its neighbors.
(2)
Normalize the NZA and NZC of each keyword separately and represent them with Na and Nc respectively, as shown in Eq. (2):

$\displaystyle\text{Na}_{i}=\frac{\textit{NZA}_{i}-\textit{NZA}_{\min}}{\textit% {NZA}_{\max}-\textit{NZA}_{\min}},\text{Nc}_{i}=\frac{\textit{NZC}_{i}-\textit% {NZC}_{\min}}{\textit{NZC}_{\max}-\textit{NZC}_{\min}}$ (2)
(3)
Calculate the weighted average of Na and Nc and use it as the OR (Overall Rating) for each keyword through the Eq. (3), Where $t$ is the weight coefficient, and the value of different $t$ has a certain influence on the OR value, but the influence on the DC value is very small. In order to improve the discrimination degree in this experiment, $t=$ 0.7 is taken:

$\displaystyle\text{OR}_{i}=\text{Na}_{i}t+\text{Nc}_{i}(1-t),\quad t\in(0,1)$ (3)
(4)
Extract the common high-frequency keywords from the above data set D1 and rate the first 12 keywords in reverse order according to their OR values as shown in Fig. 3.
(5)
Comparing Fig. 3 with Table 2, the keyword score OR value obtained through the common keyword extraction method above is more reasonable and persuasive to explain the specific keyword’s positional relationship in a specific data set and its association with the research hotspot. For example, in the field of educational technology, keywords such as “student/learning/teacher/course/technology/model” are the core vocabulary for the concern of these journals; And the vocabularies listed in Table 2 are not obviously related to this field. Therefore, from the perspective of domain research, the OR value of a keyword is more representative of the research focus of the keyword.

Table 2
Top 12 keyword list based on TF-IDF value in unsliced data set

SN Key words TF-IDF SN Key words TF-IDF SN Key words TF-IDF

1 boundary 0.987 5 publisher 0.936 9 restructure 0.903

2 teenager 0.952 6 discontent 0.934 10 badge 0.892

3 space 0.937 7 readership 0.927 11 refine 0.889

4 microelectronic 0.937 8 semantic 0.925 12 multimedia 0.885

Figure 3.
Scoring of common high-frequency keywords.

4.2.3 Calculating the time-based change of keywords

SN	Key words	TF-IDF	SN	Key words	TF-IDF	SN	Key words	TF-IDF
1	boundary	0.987	5	publisher	0.936	9	restructure	0.903
2	teenager	0.952	6	discontent	0.934	10	badge	0.892
3	space	0.937	7	readership	0.927	11	refine	0.889
4	microelectronic	0.937	8	semantic	0.925	12	multimedia	0.885

The overall score OR value calculated in the previous section is actually the OR value of each keyword in the whole statistical period (2007–2017), hereinafter referred to as OR (Global-OR). The common high-frequency keyword set $N_{k}$ and its Global-OR value constitute the time-independent domain common word set mentioned before, which is represented by $N_{k}$ -Global-OR. The specific topics in the field are composed by many keywords, and the usage of keywords in the field is reflected as the Global-OR value of the keyword. The Global-OR value is the basic reference for calculating the dynamic characters of each keyword. The specific experimental process is as follows:

(1)
Extract and calculate annual common high frequency keywords and their ORs, which can be labeled as $N_{k}$ -Year-OR, such as $N_{k}$ -2007-OR. For the data set sliced by year, that is, D2, perform the common high-frequency keyword extraction and scoring algorithm described in the previous section, respectively, to obtain the common high-frequency keywords and their OR values for each year;
(2)
Calculate the annual OR value and standard deviation of each keyword. This experiment uses the common high-frequency keyword set $N_{9}$ -Global-OR as the benchmark to match the OR values of each keyword in annual common high-frequency keyword set. Partial results are shown in Fig. 4. Among them, the keywords are sorted in reverse order of Global-OR value; Stdev column is the standard deviation of the previous columns of data, which shows the degree of dispersion of the OR value of the keyword in various years, but cannot represent the state change and development of the keyword in each year. In addition, the keyword shown in the figure is zero in a specific year, which means that the keyword is not selected for the common high frequency keyword of that year, not indicates that the keyword does not appear.
(3)
Calculate the dynamic change of the keyword based on time, that is, calculate the DC value of keywords by Eq. (4). The formula is to expand the difference between the annual OR values of the keyword and its 11 annual OR averages in certain proportion, and then sum the difference of each year. $t$ is the keyword number, $a1$ , $a2$ , … $a11$ corresponds to the OR value of the keyword in each year, that is, the OR value of 2007 corresponds to $a1$ , and the OR value of 2008 corresponds to $a2$ , …. The OR value of 2017 corresponds to $a11$ ; $b$ and $l$ are constants that expand the difference value. In this experiment, $b=$ 3, $l=$ 1.

$\displaystyle\text{DC}_{t}=\sum_{i=1}^{11}\left(a_{i}-\frac{\sum_{j=1}^{11}a_{% j}}{11}\right)*(i+b)^{l},b,l\geqslant 1,t\in[1,|N_{k}|];i,j\in[1,11]$ (4)
(4)
By drawing a graph, the meaning of Fig. 5 can be understood more intuitively. As shown in the upper part of the graph, the curve shows the values of each keyword in each year, and the virtual line is the linear trend line corresponding to the keyword curve; as shown in the figure, the DC value is directly proportional to the trend and slope of keyword curve: the positive and negative DC value respectively show the rising (positive) or falling (negative) of the trend line; the absolute DC value shows the angle of inclination of the trend line. For example, keywords “web” ( $-$ 0.993) and “graduate” ( $-$ 0.0696) are negative, so the trend line is downward. Trend line corresponding to “graduate” falls very slowly, almost the same as the horizontal line. The “web” angle is much larger than “graduate”, that is, the “web” decline rate is larger, which shows that in the past decade, the “web” related topics in the field of educational technology have gained less and less attention. It is consistent with the situation that web technology gradually disappears from the research frontier and is seen as daily use because web technology gradually becoming the basic technology in the field of educational technology; keywords “game” (1.5, obvious upward) and “implementation” (0.0569, the upward trend is not significant) are of the same situation. The DC value of the keyword “game” is also consistent with the fact that educational game topics have gradually developed over the past decade and have attracted a lot of attention from researchers [23].
(5)
The relationship between the DC value and the slope of the trend line are illustrated in the data sheet of the lower part of Fig. 5, in which horizontal axis is the year and vertical is the DC value of keywords. The DC column is the DC value corresponding to the keyword calculated according to Eq. (4), and linest column is the slope of best fitting straight line of each keyword in 2007–2017 calculated by the Linest function in Excel. The slope of the trend line, the “DC $\div$ linest” column divides the DC value by the corresponding linest value, and if the quotient of the two is 110, it means that the DC value is exactly 110 times the slope of the trend line. Excel’s Linest function uses the least squares method to calculate the line that best fits the existing data [24], the trend line. The DC value obtained by the calculation method in this paper is fixed to the slope of the linear trend line, which can be regarded as the proof that the keyword DC values numerically describe the dynamic trend of the keyword year by year.

5. Discussion

In this paper, the three dimensions of the structural information – the publishing date, the journal name and the domain-specific common word set are introduced into the dynamic character calculation of the keyword to gain the dynamic character DC value which is logical and can clearly indicate the change character of keyword in the field. These keyword-related structural information used in this paper is simple and stable in structure, and the value of each keyword obtained by TF-IDF in certain data set is also stable, which promises the stability of the DC value. Moreover, the DC value is fixed to the slope of the trend line calculated by the Linest function in Excel, which further proves that the DC value can numerically describe the development trend of the keyword. It is worth noting that although DC can be obtained by multiplying the slope of the trend line by a fixed ratio, the selection of keywords and the calculation of DC values are closely related to the characteristics of the subject domain, and thus are not merely simple values. Therefore, it may worth considered to make the DC value which is stable as the basic attribute of the keyword, to make it as a basic parameter in the research of the subject development trend to participate in the research of other methods, such as clustering.

Figure 4.

Annual OR value of common high-frequency keywords (partial).

Figure 5.

Time-based dynamic change of keyword DC value.

On the other hand, there are some problems remained unresolved in the common high-frequency keyword extraction and in the keyword DC value calculation proposed in this paper. For example, this experiment uses natural words as keywords, but the academic terms have not been included as keywords to calculate their corresponding DC value. In theory, the common high-frequency keyword set calculated with academic terms as keywords, or the domain-specific common word set has clearer logic and relevance. Secondly, in this experiment, the keywords are treated equally, and no other conditions are used to weight the keywords. The results obtained in the data set of this paper can be basically verified such as typical keywords “web” and “game” in Fig. 5; But whether it is applicable to other fields remains to be verified. The problems above are expected to be further explored and verified in the follow-up study. In addition, because the experiment is carried out under the support of some basic research technology, such as NLP processing technology, there are still a small number of keywords in the experiment are difficult to be differentiated perfectly due to the techniques of word segmentation and lemmatization, thus there are a few cases like “learn” and “learning”. These issues are beyond the scope of this article and will not be repeated.

6. Conclusion

Based on the mature TF-IDF algorithm, this paper proposes a method for calculating the dynamic characters of keywords in specific fields with the structural information from 3 dimensions as publishing date, journal name and domain-specific common word set; The key word DC value calculated with this method has clear logic and domain relevance, and can clearly indicate the change characters of the keyword in the field. The calculation method of this paper is clear in logic and simple and efficient in algorithm. And the DC value of the keyword dynamic characters may be used as the basic parameters for the research of trend development of the subject.

References

Wang

Y.F.

, A Synthetic Research of Bibliometric method and Content analysis method. (In Chinese), Ph.D., Nanjing University of Science and Technology, Nanjing, 2007.

S.Q.

and Bai

, The analysis of research trend in library and information science based on hotspot recognition of timing keywords (2000–2009) (In Chinese), New Technology of Library and Information Service 5 (2011).

Sun

H.F.

, Research on Text Mining in Complex Information Networks. (In Chinese), Ph.D., Beijing University of Posts and Telecommunications, Beijing, 2017.

Chen

C.M.

Chen

Hou

J.H.

and Liang,

Y.X.

CiteSpace II: Detecting and visualizing emerging trends and transient patterns in scientific literature. (In Chinese), Journal of The China Society for Scientific and Technical Information (2009).

Salton

Y.C.T.

, On the construction of effective vocabularies for information retrieval, in: The Proceedings of the 1973 Meeting on Programming Languages And Information Retrieval – SIGPLAN’73, Gaithersburg, Maryland, 1973.

Persson

, The intellectual base and research fronts ofJASIS 1986–1990, Journal of the Association for Information Science and Technology 45(1) (1994).

Morris

S.A.

Yen

G.G.

and Asnake

, Time line visualization of research fronts, Journal of the Association for Information Science and Technology 54(5) (2003), 413–422.

Small

, Visualizing science by citation mapping, Journal of the American Society for Information Science and Technology 50(9) (1999).

Pan

Z.G.

, News Web Texts Classification Based on Contents, (In Chinese), Master, Hefei University of Technology, Hefei, 2010.

10.

Zhang

, Design and Implementation of a Text Classification System Based on KNN Algorithm, (In Chinese), Master, Huazhong University of Science and Technology, Wuhan, 2011.

11.

Y.Y.

, Study on the Relationship between Emotional Tendency and Quoted Companies’ Performance Based on Text Mining Technology, (In Chinese), Master, China University of Geosciences, Beijing, 2016.

12.

Han

P.W.

, Research on Customer Experience Satisfaction Based on Text Mining Technology, (In Chinese), Master, Heilongjiang University, Heilongjiang, 2016.

13.

Cheng

C.H.

, Key Techniques of Text Mining on Criminal, (In Chinese), Master, Zhejiang University, Hangzhou, 2010.

14.

Cui

, Application of Data Mining in Chinese Medical Records, (In Chinese), Master, University of Electronic Science and Technology of China, Chengdu, 2007.

15.

Cui

, Research on Personalized Recommendation System Based on Text Mining, (In Chinese), Master, Shanghai Normal University, Shanghai, 2018.

16.

Yin

Z.M.

, Research on Key Technologies of Electronic Journal Literature Recommendation in Digital Library, (In Chinese), Master, Anhui University of Technology, Ma Anshan, 2016.

17.

Xiong

, Research on Diversity based on TF-IDF Recommendation Algorithm, (In Chinese), Master, Yangtze University, Jingzhou, 2018.

18.

Liu

, Research on Technology Trend Prediction Method Based on Word Vector, (In Chinese), Master, Beijing University of Technology, Beijing, 2017.

19.

Mikolov

Chen

Corrado

and Dean

, Efficient Estimation of Word Representations in Vector Space [Online].

20.

Mikolov

Sutskever

Chen

Corrado

and Dean

, Distributed Representations of Words and Phrases and their Compositionality [Online].

21.

Wang

S.G.

and Zhang

, A study of journal articles concerning research method in the field of LIS in China: Bibliometric and content analysis (1981–2015), (In Chinese), Library Journal 8 (2017).

22.

https://clarivate.com/. Web of Science. Available: http://www.webofknowledge.com.

23.

Martí-Parreño

Méndez-Ibáñez

and Alonso-Arroyo

, The use of gamification in education: A bibliometric and text mining analysis, Journal of Computer Assisted Learning 32(6) (2016), 663–676.

24.

Microsoft and Office. (2018). Linest Function. Available: https://support.office.com/en-us/article/linest-function-84d7d0d9-6e50-4101-977a-fa7abf772b6d?ui=en-US&rs=en-US&ad=US.

A calculation method for the development trend of subject keywords

Abstract

SIGNIFICANCE:

METHOD:

RESULT:

Keywords

1. Introduction

2. Literature review

3. Research goals and thoughts

4.1 Data set

Table 1 Journal list of reference

4.2.1 Data preprocessing

4.2.2.1. Extraction method for common word set with sliding window mechanism

4.2.2.2. Overall scores OR and ORE of common high-frequency keywords

References

Table 1
Journal list of reference