CN109408782A - Research hotspot based on KL distance similarity measurement develops behavioral value method - Google Patents

Abstract

The invention discloses a kind of research hotspots based on KL distance similarity measurement to develop behavioral value method, it combines the thematic timing with publication of publication, propose timing publication topic model TS-JTM, to realize that the tense hot spot of academic journals extracts, the theme snapshot publication research hotspot evolution model based on time series is established on this basis, simultaneously, utilize probability distribution KL distance similarity measure, the detection method that theme in measurement adjacent moment theme snapshot develops behavior is proposed, the fine granularity that research hotspot in publication develops is analyzed with realizing.

Description

A method for detecting evolution behavior of research hotspots based on KL distance similarity measure

technical field

The invention belongs to the technical field of document subject analysis and detection, in particular to a method for detecting evolution behavior of research hotspots based on KL distance similarity measure.

Background technique

With the continuous development of scientific research and exploration, the research hotspots in the academic field change accordingly. Due to the mutual penetration between disciplines and the application of new technologies, the academic research hotspots evolve over time. In this process, there are some Old research questions will disappear, and new research questions will continue to emerge, and some research questions will fission or merge with other research questions over time. These behaviors lead to the evolution of disciplinary research hotspots. Therefore, it is necessary to analyze the evolution of research hotspots in the academic field and grasp the evolution trajectory of research hotspots to predict the development trend of research hotspots. It can not only help scholars understand the current hot research issues, but also assist researchers and managers to grasp the development law of scientific research. The research results and progress of researchers are concentrated in the academic journals they publish academic papers. These academic journals collect a large number of academic research results in different categories. Due to the periodic publication of the journal, it essentially records the development process of the research field in which this journal is located. Therefore, it is very meaningful to discover the evolution of research hotspots over time by extracting the topics of publications.

In literature topic analysis, Author-Topic-Model is a commonly used topic clustering analysis method. ATM models the author's interest in literature and can analyze the author's academic preference ^[1] . The author topic model is a three-layer Bayesian probability model, which includes a three-layer structure of words, topics, and author interests. This model can be directly mapped to the topic model of the publication, that is, the publication selects a topic with a certain probability, and the topic generates the topic word with a certain probability. However, the evolution of topics over time is an important factor affecting topic extraction. The author topic model does not consider the time factor. When the author topic model is directly used in the corpus database of each time slice for topic extraction, it is independent in each time slice. The model parameters do not have time dependence, and do not take into account the influence of topic changes over time, which increases the uncertainty of topic headings when assigning topics. Blei proposed the DTM model based on the LDA (Latent Dirichlet Allocation) model ^[2] to achieve the extraction of time series topics. However, the DTM model is based on the content modeling of the data set, not the publication model. The topics contained in each journal and its evolution over time cannot meet the needs of journal topic research.

Therefore, there is a lack of an effective means in the prior art to solve the topic evolution behavior detection based on publication time series.

SUMMARY OF THE INVENTION

The purpose of the present invention is to provide a method for detecting the evolution behavior of research hotspots based on the KL distance similarity measure in view of the defects of the prior art. Topic Model), and use this to extract temporal topics for publications, and then combine the topic similarity of KL distance to measure topic evolution, so as to detect the evolution behaviors of topic continuation, new birth, split, fusion, and demise.

A method for detecting evolution behavior of research hotspots based on KL distance similarity measure, including the following steps:

Step 1: Obtain publication documents, and build a subject word corpus with temporal attributes based on the publication time of the publication documents;

Wherein, the time slice is divided by the publication time of the publication document, the subject word corpus is composed of the data sets on each time slice, and the data set on each time slice is composed of the document feature vector of the publication document published at the matching time;

In the formula, C _t is the data set on the time slice t, ( _wi , _ji ) is the document feature vector of the publication document i, _wi is the feature word set of the publication document i, and _ji is the publication to which the publication document i belongs. , c _i is the ith feature word in the feature word set, n ₁ is the number of publication documents on time slice t, n ₂ is the number of feature words on publication document i;

Among them, the feature words of the publication document are obtained after the content of the publication document is processed by word segmentation;

Step 2: Construct the topic model of time series publications based on the publication theme and time series;

Wherein, each time slice in the time series publication topic model corresponds to a publication topic model, and the Dirichlet prior parameter α of the publication-topic distribution θ in the publication topic model of the next time slice in the two adjacent time slices, the topic - The Dirichlet prior parameter β of the word distribution φ is associated with the two Dirichlet prior parameters α, β of the previous time slice;

Step 3: Based on the publication topic model on each time slice in the time series publication topic model, subject extraction is performed on the data set on the matching time slice in turn to obtain the publication-topic distribution and topic-word distribution on each time slice;

Step 4: Obtain the topic and topic-word distribution of the publication to be tested on each time slice, and calculate the KL distance between any two topics of the same publication to be tested on adjacent time slices based on the topic-word distribution, and then snapshot the publication based on the topic The evolution model of research hotspots is used to obtain the evolution behavior of each topic in the publication to be tested;

Among them, the hotspot evolution model of the topic snapshot journals includes five types of evolution behavior detection rules: theme continuation, new life, demise, split and fusion, and each type of evolution behavior detection law is based on the similarity of topics on adjacent time slices and the identification of evolution behavior characteristics. , the evolution behavior characteristics are related to similarity, and the similarity of two topics is measured by KL distance.

On the one hand, the present invention proposes a topic snapshot publication research hotspot evolution model, which combines the KL distance to measure the similarity between two topics of the same publication to be tested on two adjacent time slices, and covers the continuation and new life in the topic evolution. , split, merge, and disappear behavior detection rules, realize the detection of the evolution behavior of time series topics in the publications to be tested. Among them, the characteristics of various evolution behaviors are as follows: ① Continuation behavior: the theme of the current time slice remains in the next time slice, so the theme of the current time slice is only very similar to a theme of the next time slice, and is not similar to other themes; ②New behavior: The topic in the current time slice is not connected to the topic in the previous time slice, so the topic of the current time slice is not similar to all the topics in the previous time slice; ③Split behavior: The topic of the current time slice is generated In order to split, multiple topics are generated, so the topic of the current time slice is similar to two or more themes in the next time slice; ④ Fusion behavior: multiple topics are merged into one topic, so the topic of the current time slice is the same as the previous one. Two or more themes in one time slice are similar; ⑤Dead behavior: the theme in the current time slice is not connected with the theme in the next time slice, so the theme of the current time slice and all the themes in the next time slice are not similar. Based on the characteristics of various evolution behaviors and the KL value used to measure similarity, the invention can deduce the evolution behavior of topics on the publication to be tested.

On the other hand, the present invention constructs a chronological publication subject model based on the subjectivity and temporality of the publication, which takes into account the influence of the subject changes over time, and adopts the method of parameter transmission to construct the relationship between the subject models of the publications on adjacent time slices. , which reduces the uncertainty of subject headings when assigning topics, making the model less confusing; at the same time, the topic model of time series publications is modeled for publications in the literature dataset, because the subjectivity of the subject areas represented by the publications is higher than that of the authors. Compared with the conventional author topic models ATM and DTM models, the topic model of the time series publication of the present invention is more in line with the needs of the present invention to study the topic evolution of the publication.

Further preferably, the evolution model of the research hotspots of the subject snapshot journals includes the following detection rules:

a: The KL distance between topic i on time slice t and the previous topic on the adjacent next time slice t+1 is less than the similarity threshold, and the KL distance from the remaining topics on the adjacent next time slice t+1 is greater than or equal to the similarity threshold, topic i remains in the next time slice t+1:

b: When the KL distance between topic i on time slice t and each topic on the adjacent previous time slice t-1 is greater than the similarity threshold, topic i on time slice t is a new topic:

c: When the KL distance between topic i on time slice t and each topic on adjacent next time slice t+1 is greater than the similarity threshold, topic i on time slice t does not continue in the next time slice t+1 , subject i dies:

d: When the KL distance between topic i on time slice t and at least two topics on adjacent next time slice t+1 is smaller than the similarity threshold, topic i on time slice t is split in the next time slice t+1 For multiple topics:

e: When the KL distance between topic i on time slice t and at least two topics on the adjacent previous time slice t-1 is less than the similarity threshold, topic i on time slice t is determined from the previous time slice t-1. Combining multiple themes.

Further preferably, the detection formula of each detection law in the evolution model of the subject snapshot publication research hotspot is as follows:

The detection formula of the continuous evolution behavior in the a-law is:

In the formula, are the KL distances between the topic i on the t time slice and the topic j on the t+1 time slice, and the KL distance between the topic i on the t time slice and the topic k on the t+1 time slice, are the topic-word distributions of topic i on t time slice, topic j on t+1 time slice, and topic k on t+1 time slice, T ^t+1 is the set of topics on t+1 time slice, and threshold_A is the similarity threshold;

The detection formula for the evolution behavior of the new theme in the b-law is:

In the formula, is the KL distance between the topic j on the t-1 time slice and the topic i on the t time slice, and T ^t-1 is the set of topics on the t-1 time slice;

The detection formula of the extinction evolution behavior in the c-law is:

The detection formula of split evolution behavior in d-law is:

The detection formula of fusion evolution behavior in e-law is:

Further preferably, the KL distance calculation formula of the two topics is as follows:

In the formula, is the KL distance between topic j on time slice t-1 and topic i on time slice t, represent the topic-word distribution of topic j on time slice t-1 and topic i on time slice t, respectively, respectively The word probability of the topic word x under the topic-word distribution, X represents the topic word set of topic j on the t-1 time slice, and x represents any topic word in X.

It should be understood that when calculating the KL distances on other two adjacent time slices, the above formula is also used, and this formula is a general formula. It should be noted that, in the formula, the subject word x does not exist in the subject word set of the subject i in the t time slice, so φ _i ^t (x) is taken as a preset small value, such as 0.001.

Further preferably, the similarity threshold is 0.4.

Further preferably, in the publication topic model on the adjacent time slice in step 2, the Dirichlet prior parameter α of the publication-topic distribution θ and the Dirichlet prior parameter β of the topic-word distribution φ are related to each other as follows:

β _t |β _t-1 ～N(β _t-1 ,σ ² I)

α _t |α _t-1 ～N(α _t-1 ,δ ² I)

In the formula, β _t and β _t-1 are the Dirichlet prior parameters of the topic-word distribution in the publication topic model on time slice t and time slice t-1, respectively, and α _t and α _t-1 are time slices, respectively. t. The Dirichlet prior parameters of the publication-topic distribution in the publication-topic model on the time slice t-1, N(β _t-1 ,σ ² I) and N(α _t-1 ,δ ² I) are both Normal distribution, σ ² I and δ ² I represent the variance of the corresponding random variable;

β _t |β _t-1 ～N(β _t-1 ,σ ² I) indicates that the prior parameter β _t of the topic-word distribution in the time slice t is affected by the topic-word distribution in the previous time slice t-1. The influence of the prior parameter β _t-1 and satisfy the N(β _t-1 ,σ ² I) distribution, α _t |α _t-1 ～N(α _t-1 ,δ ² I) represents the publications under the time slice t - The Dirichlet prior parameter α _t of the topic distribution is influenced by the publications under the previous time slice t-1 - The Dirichlet prior parameter α _t-1 of the topic distribution satisfies N(α _t-1 ,δ ² I) Distribution.

The present invention takes into account that academic journals are published periodically over time, and the evolution of their topics is gradual, and the adjacent time slices are connected by means of parameter transmission, that is, adjacent time slices are passed through Dirichlet's prior parameters α and β. parameter connection. Since the values of Dirichlet's prior parameters α and β will affect the formation of topics and change the distribution of words in the topic, the present invention uses the two parameters α and β to convert the publication-topic distributions θ and The influence of topic-word distribution φ is transmitted to the parameters of the topic model of the adjacent next time slice, which reduces the uncertainty of topic words when assigning topics, and makes the model less perplexed.

Further preferably, the number of topics of the time series publication topic model described in step 2 and the Dirichlet prior parameter α of the publication-topic distribution θ in the publication topic model on the first time slice, the Dirichlet prior of the topic-word distribution φ. The test parameter β is a preset value.

Further preferably, the number of topics in the topic model of the time series publication is 50.

Further preferably, in the publication topic model on the first time slice, the Dirichlet prior parameter α of the publication-topic distribution θ is 1, and the Dirichlet prior parameter β of the topic-word distribution φ is 0.01.

beneficial effect

1. The present invention proposes a brand-new topic snapshot publication research hotspot evolution model, which combines the KL distance to measure the similarity between two topics of the same publication to be tested on two adjacent time slices. By detecting the continuation, new birth, split, fusion, and demise of the topic evolution in the journal, fine-grained analysis of the evolution of research hotspots in publications is realized, which fills the gap in the existing technology to solve the effective detection method of the topic evolution behavior based on the time series of publications. Among them, the detection rules of the evolution behaviors of continuation, new birth, split, fusion and extinction in the subject snapshot publication research hotspot evolution model provided by the present invention are deduced based on the similarity between subjects of the same publication on adjacent time slices, which accurately reflects The evolution of the theme.

2. The topic model of time series publications is constructed based on the topicality and time series of the publications, which combines the characteristics of the JTM and DTM models of the publication topic model, that is, on the one hand, the influence of the topic changes over time is taken into account, and the method of parameter transfer is used to construct the The relationship between the topic models of the publications on the adjacent time slices reduces the uncertainty of the topic words when assigning topics, making the model less perplexing, overcoming the influence of the topic model over time, which is not considered in the topic model of a separate publication, and increasing the number of topics. Defects of the uncertainty of words in assigning topics. The topic model of the time series publication of the present invention connects adjacent time slices through the Dirichlet a priori parameters α and β. Since the values of the Dirichlet prior parameters α and β will affect the formation of the topic, changing the words in the topic Therefore, the present invention transfers the influence of the publication-topic distribution θ and the topic-word distribution φ in the previous time slice to the topic model parameters of the adjacent next time slice through the two parameters α and β. On the other hand, the topic model of the time series publications of the present invention is based on the modeling of the publications in the document data set. Since the subject area of the subject area represented by the publication is more thematic than the subject area represented by the author, although the DTM model takes into account the subject matter of the subject area The time change is affected by the previous topic, but it only models the content of the dataset without considering the publication, which cannot meet the evolution needs of the publication topic model. Therefore, compared with the independent publication topic model JTM and the existing DTM model, the time series publication topic model of the present invention is more in line with the needs of the present invention to study the topic evolution of publications, and the publication-topic distribution and topic-word distribution on each time slice can be obtained by using it. distribution, which lays the foundation for the detection of the subject evolution of subsequent publications.

3. Through experimental verification, the time series publication subject model provided by the present invention has better performance in terms of perplexity and running time, and the perplexity degree of the time series publication subject model is lower than that of the author's subject model ATM and DTM models, and the operation of the time series publication subject model. The time is close to that of the DTM model and is shorter than that of the ATM.

Description of drawings

Fig. 1 is a kind of schematic flow chart of a research hotspot evolution behavior detection method based on KL distance similarity measure provided by the present invention;

Fig. 2 is the schematic diagram of the publication theme model provided by the invention;

Fig. 3 is the schematic diagram of the time series publication subject model provided by the present invention;

4 is a schematic diagram of the subject evolution behavior in the subject snapshot publication research hotspot evolution model provided by the present invention;

5 is a schematic diagram of the evolution of the theme under the publication ID: 003 provided by the present invention between 2010 and 2016;

FIG. 6 is a schematic diagram of the perplexity comparison of the ATM model, the DTM model and the TS-JTM model provided by the present invention.

Detailed ways

The present invention will be further described below with reference to the embodiments.

Since the research hotspots in the academic field are mainly reflected in academic journals, how to analyze the evolution behavior of the topics in the data sets of academic journals is of great significance for researchers to understand the development trajectory of disciplinary research hotspots and grasp the development laws of research hotspots. As shown in FIG. 1 , the present invention provides a method for detecting the evolution behavior of research hotspots based on the KL distance similarity measure based on the need, including the following steps:

Step 1: Literature information preprocessing. Firstly, the publication documents are obtained from the public document information database and preprocessed, and then a subject word corpus with time attribute is constructed based on the publication time of the publication documents.

The preprocessing process is as follows: extract the title, abstract, keywords, publication name and publication time of the publication document, and then format it, use the word segmentation tool to divide the abstract and document title into phrases, delete stop words, and combine the remaining phrases with key words. Words make up the characteristic words of the document. In other feasible embodiments, the characteristic words of documents may also be derived only from abstracts, or from abstracts, keywords; or from abstracts, document titles, etc., which are not specifically limited in the present invention.

After the feature word set of each document is obtained, the time slice is divided according to the publication time of the document, and the feature words of the document belonging to the same time slice and the publication information described in the document form the data set of the time slice. The datasets for each time slice constitute the subject word corpus.

For example, obtain scientific and technological literature information from CNKI public literature resource database to construct a subject word corpus. From 2010 to 2016, 6487 article abstracts and their corresponding publication names and publication time were selected as experimental data. Divide all the literature information into 7 time-slice datasets by year, and then use the Chinese word segmentation system NLPIR of the Chinese Academy of Sciences to segment and remove stop words for each paper abstract to form a set of subject words for each literature. Among them, (wi _i , j _i ) is used to represent the document feature vector of document i. where w _i represents the set of feature words in document i, and _ji represents the publications published by document i. The dataset C _t consisting of n ₁ documents in time slice t can be expressed as

Step 2: Construct a Time Series Journal Topic Model (TS-JTM) based on the journal topicality and time series.

The model of the Time Series Journal Topic Model (TS-JTM) in each time slice is the journal topic model, as shown in Figure 2. α and β in the model represent the Dirichlet prior parameters of the publication-topic distribution θ and the topic-word distribution φ, respectively, K represents the total number of publications, and T represents the number of topics. The core idea of the publication topic model is: the publication J to which an article belongs selects a topic z from its corresponding topic distribution θ, and randomly generates a word w according to the probability distribution φ of this topic on words. Repeat this process until every word in this article is generated.

In the time series publication topic model (TS-JTM) of the present invention, there is an associated relationship between the publication topic models on adjacent time slices. As with the DTM model, as shown in Figure 3, adjacent time slices are connected by the Dirichlet prior parameters α and β, where the values of the Dirichlet prior parameters α and β will affect the formation of the topic, changing the words in the topic. distributed. The calculation formula of parameters between adjacent time slices is as follows:

β _t |β _t-1 ～N(β _t-1 ,σ ² I) (1)

α _t |α _t-1 ～N(α _t-1 ,δ ² I) (2)

φ _t ~Dir(β _t ) (3)

θ _t ~Dir(α _t ) (4)

Among them, formula 1 indicates that the prior parameter β t of the topic-word distribution under the time slice t is affected by the prior parameter β _{t -} 1 of the topic-word distribution under the previous time slice t-1 and satisfies N(β _{t -1} ,σ ² I) distribution, β _t and β _t-1 satisfy the first-order Markov process; in the same way, formula 2 indicates that the Dirichlet prior parameter α _t of the publication-topic distribution under the time slice t is subject to The influence of the Dirichlet prior parameter α _t-1 of the publication-topic distribution under a time slice t-1 and satisfy the N(α _t-1 ,δ ² I) distribution. Equation (3) and Equation (4) indicate that the parameters β _t and α _t are the Dirichlet prior parameters of the topic-word distribution φ _t and the publication-topic θ _t in the model, respectively. The values of the Dirichlet prior parameters α _t and β _t affect the publication-topic distribution and the topic-word distribution.

Based on the model architecture of the above-mentioned time series publication topic model, set the number of topics in the model and the values of Dirichlet's prior parameters β ₁ and α ₁ on the first time slice, and then perform topic extraction on the data set on the first time slice, namely The publication-topic distribution θ ₁ and the topic-word φ ₁ distribution on the first time slice can be obtained; then the Dirichlet prior parameters β ₁ and α of the first time slice can be obtained by using formula (1) and formula (2). ₁ Calculate the new β ₁ ' and α ₁ ', and pass the new parameters β ₁ ' and α ₁ ' to the second time slice as the initial values of the model hyperparameters in the second time slice, and then for the first time slice Subject extraction is performed on the datasets on the two time slices, and this process is repeated continuously to obtain the publication-topic distribution and topic-word distribution on each time slice. That is, the Dirichlet prior parameters α _t and β _t in other time slices are calculated according to α _t-1 and β _t-1 in the previous time slice, respectively. Among them, the process of using the publication topic model on the time slice to perform topic extraction on the data set on the matching time slice to obtain the publication-topic distribution and the topic-word distribution is a prior art implementation process, which is not described in detail in the present invention, only briefly described illustrate.

The Gibbs sampling (Gibbs Sampling) method is used to infer the parameters of the publication topic distribution θ and the topic word distribution φ in the time series publication topic model. For each word, the journal and topic are sampled according to Equation 5. The right side of Equation 5 is p(topic|journal)·p(word|topic), that is, the probability that the journal selects the topic and the topic selects the word. Since there are T topics and K journals, the physical meaning of the formula is to sample in these K×T paths.

In the formula, z _i =j, x _i =k here represents that the i-th word in a document is assigned to the j-th topic (Topic) and the k-th publication. W _i =m represents that the i-th word is the m-th word in the dictionary. Z _-i , X _-i represent the topic and publication assignments of the remaining words except the i-th word. represents the total number of words m has been assigned to topic j before this assignment, represents the total number of publications k assigned to topic j so far. N is the total number of words in the dictionary, which consists of all the different feature words in the dataset. Formula (1) only needs to record two matrices in the parameter estimation of the model, one is the topic-word (word by topic) count matrix N×T, the other is the journal-topic (Journal by topic) count matrix K×T, and then According to these two count matrices, the calculation formulas of topic-word distribution φ and publication-topic distribution θ are calculated as formula (6) and formula (7), respectively.

In the formula, φ _mj represents the probability that topic j uses word m, θ _kj represents the probability that publication k selects topic j, m' represents any word that has been assigned to topic j, and j' represents any word that has been assigned to journal k. Theme of.

Step 3: Based on the publication topic model on each time slice in the time series publication topic model, subject extraction is performed on the data set on the matching time slice in turn to obtain the publication-topic distribution and the topic-word distribution on each time slice.

Based on step 2, the architecture of the topic model of time series publications is constructed. In this embodiment, the number of topics of the topic model of time series publications, the Dirichlet prior parameter α of the publication-topic distribution θ, and the Dirichlet prior parameter of the topic-word distribution φ are set. The initial value of β, and then subject extraction is performed on the data sets on each time slice in turn to obtain the publication-topic distribution and topic-word distribution on each time slice. The process is to use the TS-JTM model for topic extraction, that is, 1.1, 1.2, and 1.3 are executed cyclically for each time slice t;

1.1 Use the TS-JTM model in the time slice t to extract topics from the dataset to obtain the topic set T ^t and topic-word distribution;

1.2 Add the subject set T ^t to the set TC of time series subjects;

1.3 Update the model TS-JTM with the parameters α _t and β _t of the current time slice model.

It should be understood that updating the TS-JTM model is updating the model parameters α and β on the next time slice in the topic model of the time series publication.

Step 4: Obtain the topic and topic-word distribution of the publication to be tested on each time slice, and calculate the KL distance between any two topics of the same publication to be tested on adjacent time slices based on the topic-word distribution, and then snapshot the publication based on the topic The evolution model of research hotspots is used to obtain the evolution behavior of each topic in the publication to be tested.

As shown in FIG. 4 , it represents the subject snapshots of three adjacent time slices issued by a certain publication, and the dotted lines between the time slices represent the relationship between the subjects. The subject snapshot publication hotspot evolution model proposed by the present invention includes the behavioral characteristics between subjects, which are as follows: ① a one-to-one relationship indicates that the subject of the current time slice is a continuation of the subject of the previous time slice; ② the next time If the theme in the film is not connected with the theme in the previous time slice, it indicates that there is a new theme; ③ a one-to-many relationship indicates that the theme of the previous time slice has been split and multiple themes have been generated; ④ a many-to-one relationship indicates that many 5. When the theme in the previous time slice is not connected with the theme in the next time slice, it means that the theme has died.

In order to measure the similarity between two topics, the present invention adopts the KL distance. KL (Kullback-Leibler Divergence) distance was proposed by Solomon Kullback and Richard Leibler ^[3] , also called relative entropy (Relative Entropy), which is often used to measure the similarity between two probability distributions, and KL distance can be used to measure adjacent time The similarity between any two themes in the film. The following formula 8 is the calculation formula of the KL distance, wherein, and Represent two probability distributions, respectively, when the two probability distributions are exactly the same, the value of the KL distance is 0.

The present invention uses the KL distance to measure the similarity between two topics distributed on two adjacent time slices, and establishes the correspondence between the adjacent time slice topics, and the probability distribution in the formula corresponds to the topic-word distribution of the topic .

Based on the evolution behavior in the aforementioned 1-5, the subject snapshot publication research hotspot evolution model of the present invention includes the following detection rules:

a: The KL distance between topic i on time slice t and the previous topic on the adjacent next time slice t+1 is less than the similarity threshold, and the KL distance from the remaining topics on the adjacent next time slice t+1 is greater than When equal to or equal to the similarity threshold, topic i remains in the next time slice t+1.

b: When the KL distance between topic i on time slice t and each topic on the adjacent previous time slice t-1 is greater than the similarity threshold, topic i on time slice t is a new topic.

c: When the KL distance between topic i on time slice t and each topic on adjacent next time slice t+1 is greater than the similarity threshold, topic i on time slice t does not continue in the next time slice t+1 , the subject i dies.

d: When the KL distance between topic i on time slice t and at least two topics on adjacent next time slice t+1 is smaller than the similarity threshold, topic i on time slice t is split in the next time slice t+1 for multiple themes.

e: When the KL distance between topic i on time slice t and at least two topics on the adjacent previous time slice t-1 is less than the similarity threshold, topic i on time slice t is determined from the previous time slice t-1. Combining multiple themes.

In summary, the detection formulas corresponding to the a-e detection laws are as follows:

in, Indicates the evolution behavior status of the i-th topic in the t-th time slice, and Threshold_A is the similarity threshold. After repeated experiments, when threshold_A is set to 0.4, it can reasonably reflect the evolution behavior of the theme. In other feasible embodiments, other values may be adopted.

For the processing of the publication to be tested on each time slice, the following procedures 2.2.1 and 2.2.2 are executed respectively:

2.2.1 Extract the subject set T ^t of the current time slice and the subject sets T ^t-1 and T ^t+1 of the two time slices adjacent to the current time slice from the set TC, and obtain the publications to be tested in the set T ^t Topics in ^-1 ,T ^t ,T ^t+1 ;

2.2.2 According to formula 9, the evolution behavior of each topic of the publication to be tested on the current time slice is detected.

In order to more clearly describe the solution of the present invention, a number of examples will be provided below.

1. Changes in subject headings over time

As shown in Table 1 below, for the publication with ID: 003 in the dataset, in the topic distribution in 2010, topic number 2 is a topic related to the field of face recognition. The topic-word distribution of topic number 2 from 2010 to 2016 is shown in Table 2, which shows the 10 topic words with the highest probability under this topic in each year. It can be seen from the table that the core vocabulary in the topic of "face recognition" has not changed significantly over time, such as "image", "feature", "face recognition" and other popular topics related to face recognition The vocabulary is always in the theme. But the "genetic algorithm" that appeared in 2013 and the "deep learning" that appeared in 2015 are the application of some new methods in the field of "face recognition". From 2010 to 2016, the KL values of two time slice subjects adjacent to each other were 0.20, 0.26, 0.23, 0.17, 0.21, 0.19, and these KL distances were all smaller than the similarity threshold threshold_A. During this period, "Face Recognition" "The KL distance between the topic and other topics in the next time slice is greater than the similarity threshold threshold_A, so the topic "Face Recognition" continued from 2010 to 2016.

Table 1 The topic-word distribution table of "Face Recognition"

2. The evolution of publication themes over time

For the convenience of description, we will use their English abbreviations for topics in subsequent articles. Table 2 shows the top 10 topics with the highest probability in the three topics of "Neural Network (NN)", "Deep Learning (DL)" and "Speech Recognition (SR)" in each year from 2010 to 2016. As can be seen from Table 2, the core words such as "neural network", "neuron" and "feature" in the topic NN remain basically unchanged, and the distribution of edge words such as "sample" and "particle swarm" in different time slices big change. The topic NN in 2013 and the topic DL in 2014 have the same words in the top 10 topic words as "training", "classification", "performance", "feature", "neuron", due to the similarity of word distributions. The KL value between these two topics is small, 0.27, which is less than the similarity threshold threshold_A, the KL value of the topic NN in 2013 and all topics in the next time slice are 0.55, 0.27, 0.21, 0.69, 1.84, 1.16, 0.92, 1.53, the topics corresponding to the minimum values are DL and NN respectively, and the rest are larger than the similarity threshold, so the topic DL is generated by the topic NN split.

Table 2 The word distribution of three topics including "speech recognition" in 2010-2016

3. Analysis of the evolution of the publication theme

The evolution of topics under publication ID: 003 from 2010 to 2016 is shown in Figure 5. Since the same topic forms different numbers through clustering in different time slices, the same topic in the figure is represented by an English abbreviation. As can be seen from Figure 5, the KL values of all topic distributions in 2015 and topic SR in 2016 are 0.74, 0.46, 0.23, 0.16, 0.81, 0.95, 1.37, respectively, and the two topics less than the similarity threshold are NN respectively and SR, the remaining KL values are larger than the similarity threshold, indicating that the topic NN is integrated into the topic SR, and the KL distances between the 2014 "aircraft" topic and all the topics in 2015 are 1.72, 1.46, 1.25, 1.07, 1.20, 0.83, 1.59, respectively. , the minimum value of KL is 0.83, which is greater than the similarity threshold, so the "aircraft" topic died out in 2015; the KL values of all topics in 2010 and the "cloud computing" topic in 2011 were 1.16, 0.75, 1.37, 2.32, 1.51, the minimum value of KL is 0.75, which is greater than the similarity threshold, and the subject is a freshman; similarly, the subject of "target tracking" has been in a continuous state, and the new subject in 2013 is "entity recognition".

Model performance verification

In order to verify the model performance of the Time Series Journal Topic Model (TS-JTM) proposed by the present invention, the present invention adopts the perplexity index. The following formula 10 is the calculation formula of Perplexity, where D _test represents the test set, which is a collection of M documents, p(W _d ) represents the probability of the word in the document being selected, N _d represents the number of words in the document d, W _d =(w _1d , w _2d ,...,w _id ,...,w _nd ) represents the word vector form in document d. The smaller the Perplexity value, the better the performance of the model.

In order to measure the performance of the time series publication topic model (TS-JTM), three parameters of the model need to be set before the experiment. The value of the number of topics |T| gradually increases from 10, and the two Dirichlet hyperparameter values They are set as: α=50/|T|, β=0.01, and the two Dirichlet hyperparameter values of DTM and time series publication topic model in the first time slice are respectively set as α=50/|T|, β = 0.01, α and β in the remaining time slices are automatically obtained by the model. The comparison experiment results are shown in Figure 6. The horizontal axis represents the number of topics, and the vertical axis represents perplexity. We can see that with the change of the number of topics, the perplexity of TS-JTM is always the smallest, which indicates that TS-JTM The performance is the best. In addition, the Perplexity only decreases with the increase of the number of topics. When the number of topics is greater than 50, the Perplexity remains unchanged, which indicates that the number of topics for the TS-JTM model is set to 50 is reasonable. In this embodiment, the number of topics in the topic model of the time series publication is preferably 50, and the corresponding two Dirichlet hyperparameters are set as: α=50/|T|, β=0.01.

On the other hand, the present invention tests the running time of the Time Series Journal Topic Model (TS-JTM) on the dataset. Compared. Using these three models to process the same data separately, the running times of the three models are 23.8 minutes, 25.6 minutes, and 24.2 minutes, respectively. This shows that the running time of TS-JTM and DTM is close, and the running time of ATM is the longest. Combined with the model perplexity performance in Figure 4, it shows that the time series publication topic model (TS-JTM) not only has low perplexity, but also has good performance in running time.

To sum up, the subject evolution of academic journals reflects the development trend of research hotspots in the academic field. Because the thematic and temporal nature of a publication will affect the distribution and evolution of the publication's topic, there are evolutionary behaviors in the evolution of the topic, which complicates the identification of the evolution track of the publication's research hotspots. This paper proposes a topic model TS-JTM for temporal publications based on the subject nature of publications and the temporal nature of publications. TS-JTM is used to extract temporal hotspots from academic publications, and the performance of the model TS-JTM is verified through a perplexity comparison experiment. On this basis, a time-series-based topic snapshot publication research hotspot evolution model is established, and the KL distance is used to measure the similarity to detect the continuation, new birth, split, fusion, and demise behaviors of topic evolution in topic snapshots at adjacent moments. A fine-grained analysis of the evolution of research hotspots in publications.

It should be emphasized that the examples described in the present invention are illustrative rather than restrictive, so the present invention is not limited to the examples described in the specific implementation manner, and all the examples obtained by those skilled in the art according to the technical solutions of the present invention Other embodiments that do not depart from the spirit and scope of the present invention, whether modified or replaced, also belong to the protection scope of the present invention.

References are as follows:

[1] Rosen-Zvi M, Griffiths T, Steyvers M. The Author-Topic Model for Authors and Documents [C]. Proceedings of the 20th Conference on Uncertainty inArtificial Intelligence. 2004:487-494.

[2] Blei D M, Lafferty J D. Dynamic Topic Models [C]. Proceedings of the 23rd International Conference on Machine Learning, 2006: 113-120.

[3] David J.C. MacKay. Information Theory, Inference, and Learning Algorithms [M]. Cambridge University Press, 2003: 22-48.

Claims (9)

1. A KL distance similarity measurement-based method for detecting evolution behaviors of research hotspots is characterized by comprising the following steps: the method comprises the following steps:

step 1: acquiring publication documents, and constructing a subject term corpus with time attributes based on publication time of the publication documents;

dividing time slices by publication document publication time, wherein the subject term corpus is composed of data sets on each time slice, and the data set on each time slice is composed of document feature vectors of publication documents published at matching time;

in the formula, C_tIs a data set over a time slice t, (w)_i,j_i) Document feature vector, w, for publication document i_iSet of characteristic words, j, for publication document i_iAs a publication to which publication document i belongs, c_iIs the ith feature word in the feature word set, n₁Is the number of publication documents on time slice t, n₂The number of characteristic words on publication document i;

wherein, the characteristic words of the publication documents are obtained after the content of the publication documents is subjected to word segmentation processing;

step 2: constructing a time sequence publication theme model based on publication themes and time sequence;

each time slice in the time-series publication topic model corresponds to a publication topic model, and a dirichlet prior parameter α of publication-topic distribution theta and a dirichlet prior parameter β of topic-word distribution phi in the publication topic model of the next time slice in two adjacent time slices are associated with two dirichlet prior parameters α and β of the previous time slice;

and step 3: sequentially carrying out theme extraction on the data sets on the matched time slices based on a publication theme model on each time slice in the time sequence publication theme model to obtain publication-theme distribution and theme-word distribution on each time slice;

and 4, step 4: the method comprises the steps of obtaining the theme and theme-word distribution of a publication to be tested on each time slice, calculating the KL distance between any two themes of the same publication to be tested on adjacent time slices based on the theme-word distribution, and obtaining the evolution behavior of each theme in the publication to be tested based on a theme snapshot publication research hotspot evolution model;

the topic snapshot publication research hotspot evolution model comprises five types of evolution behavior detection rules, namely topic continuation, new generation, extinction, division and fusion, each type of evolution behavior detection rule is identified based on the similarity of topics on adjacent time slices and the evolution behavior characteristic, the evolution behavior characteristic is related to the similarity, and the similarity of the two topics is measured by adopting KL distance.

2. The method of claim 1, wherein: the topic snapshot publication research hotspot evolution model comprises the following detection rules:

a: when the KL distance between the topic i on the time slice t and the topic on the next time slice t +1 is less than the similarity threshold value and the KL distances between the topic i on the next time slice t +1 and the remaining topics on the next time slice t +1 are greater than or equal to the similarity threshold value, the topic i keeps continuing in the next time slice t + 1:

b: when the KL distance between the topic i on the time slice t and each topic on the adjacent time slice t-1 is larger than the similarity threshold value, the topic i on the time slice t is a new topic:

c: when the KL distance between the theme i on the time slice t and each theme on the next adjacent time slice t +1 is larger than the similarity threshold, the theme i on the time slice t is not continued in the next time slice t +1, and the theme i disappears:

d: when KL distances between the theme i on the time slice t and at least two themes on the next adjacent time slice t +1 are both smaller than a similarity threshold value, the theme i on the time slice t is split into multiple themes in the next time slice t + 1:

e: and when the KL distances between the topic i on the time slice t and at least two topics on the adjacent last time slice t-1 are smaller than the similarity threshold value, fusing the topic i on the time slice t by a plurality of topics in the last time slice t-1.

3. The method of claim 2, wherein: the detection formula of each detection rule in the topic snapshot publication research hotspot evolution model is as follows:

the detection formula of the continuous evolution behavior in the rule a is as follows:

in the formula,KL distances between a subject i and a subject j on a t +1 time slice and between a subject i and a subject k on a t +1 time slice respectively,topic-word distributions of topic i, topic j, and topic k on T +1 time slices, T^t+1A topic set on t +1 time slice, and threshold _ A is a similarity threshold;

the detection formula of the evolution behavior of the new theme in the rule b is as follows:

in the formula,is KL distance between topic j and topic i on T-1 time slice^t-1Is a topic set on a t-1 time slice;

the detection formula of the death evolution behavior in the rule c is as follows:

the detection formula of the splitting evolution behavior in the rule d is as follows:

the detection formula of the fusion evolution behavior in the e rule is as follows:

4. the method of claim 1, wherein: the KL distance calculation formula for both topics is as follows:

in the formula,the KL distance of topic j to topic i on t-1 time slice,respectively representing the subject-word distribution of a subject j on a t-1 time slice and a subject i on a t time slice,are respectively asWord probability of a topic word X under topic-word distribution, X representing a topicX represents any subject word in X.

5. The method of claim 1, wherein: the similarity threshold is 0.4.

6. The method as claimed in claim 1, wherein the Dirichlet prior parameter α of the publication-topic distribution θ and the Dirichlet prior parameter β of the topic-word distribution φ in the publication topic model at the adjacent time slice in step 2 are related as follows:

β_t|β_t-1～N(β_t-1,σ²I)

α_t|α_t-1～N(α_t-1,δ²I)

in the formula, β_t、β_t-1Journal topic models on time slice t and time slice t-1 respectivelyDirichlet prior parameter of medium topic-word distribution, α_t、α_t-1Dirichlet prior parameter N (β) for publication-topic distribution in publication topic models at time slice t and time slice t-1, respectively_t-1,σ²I) And N (α)_t-1,δ²I) Are all normally distributed, σ²I and delta²I represents the variance of the corresponding random variable;

β_t|β_t-1～N(β_t-1,σ²I) prior parameter β representing topic-word distribution under time slice t_tSubject-word distribution prior parameter β for last time slice t-1_t-1And satisfies N (β)_t-1,σ²I) Distribution, α_t|α_t-1～N(α_t-1,δ²I) Prior parameter α representing publication-topic distribution under time slice t_tPrior parameter α of publication-topic distribution under last time slice t-1_t-1And satisfies N (α)_t-1,δ²I) And (4) distribution.

7. The method as claimed in claim 6, wherein the topic number of the time series topic model in step 2 and the Dirichlet prior parameter α of the topic-topic distribution θ and the Dirichlet prior parameter β of the topic-word distribution φ in the topic model at the first time slice are preset values.

8. The method of claim 7, wherein: the number of topics for the temporal publication topic model is 50.

9. The method as recited in claim 7, wherein the Dirichlet prior parameter α of the publication-topic distribution θ is 1 and the Dirichlet prior parameter β of the topic-word distribution φ is 0.01 in the publication topic model at the first time slice.