The rapid development of disruptive technologies has attracted the attention of major countries in the world in recent years, and the mining and research on the texts of disruptive technology policies of these countries can reveal the key layout, focus areas, and development pattern of each countries disruptive technology. This article first crawls the texts of disruptive technologies from the science and technology policy websites of major countries. Then, the text is segmented by Spacy, the segment result is filtered by a word list to construct an applicable TF*IDF matrix, and finally the matrix weights are optimized with manually collected domain core words and important words. After these, extraction and statistics of technical entity are performed according to a specified word list. Through comprehensive analysis, it can be found that the keyword hotspots of the experimental texts are focused on artificial intelligence, information security, new energy, etc. The key areas of specific disruptive technologies are artificial intelligence, air and space, and new generation communication technologies. The result reflects the current situation and policy focus of disruptive technology development in these countries.
• Computing methodologies • Artificial intelligence • Natural language processing • Information extraction disruptive technology, science and technology policy, keywords extraction, technology entity extraction Science and technology field are changing rapidly, especially under the wave of big data and artificial intelligence. Some countries have introduced many S&T policies to promote the development of S&T. 2. Unlike academic papers, S&T policy texts do not carry keywords, so keywords need to be extracted from a large number of long texts of S&T policies. The text of S&T policy will involve both S&T projects, innovation mechanisms, transformation of results, industrial support, S&T rewards, S&T management and configuration, but also other specific S&T content, which are the research frontier, high-tech, industry common technology, disruptive technologies, and so on.
Disruptive technology was proposed by Harvard University
Professor Christensen in 1997 [
It is generally believed that disruptive technologies are strategic innovative technologies that open up new technological tracks based on new principles, combinations and applications of S&T, and produce an overall or fundamental replacement for traditional or mainstream technologies. Disruptive technologies have strong application capabilities, can enhance the scientific and technological competitiveness of enterprises and even countries, promote the renewal of scientific and technological products, improve social production efficiency, and are expected to have farreaching impact in many fields. Disruptive technology policies can stimulate technological innovation and provide corresponding support and guarantee, so it is necessary to study the mining of disruptive technology policies text.
Named entity extraction (NER) is a hot topic in the field of natural language processing (NLP), which specifically includes unsupervised or supervised recognition of specialized domain vocabulary such as names of people, places, time, products, and organization names, and textual keyword extraction on this basis is also an important application. Named entity extraction is usually the first step in intelligent information retrieval, relationship extraction, and even knowledge graph construction. Domainspecific entity extraction is also the key to explore the trends and hotspots in the field and to build dynamic knowledge graphs.
Keyword extraction is pivotal in NLP, and a few keywords can effectively reveal the main contents and topics of long documents.
The study of S&T policy texts has also been paid attention by
some scholars in recent years, Alan Porter proposed that Tech
Copyright 2021 for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).
Mining makes exploitation of text databases meaningful to those
who can gain from derived knowledge about emerging
technologies. [
At present, although there has been a lot of research on S&T policy, there is no specialist algorithm and thesaurus for disruptive technology policies, much less extracting keywords from these policies documents.
Keyword extraction algorithms mainly include unsupervised
methods and supervised methods. Unsupervised algorithms are
based on statistical features of the text, such as the TF*IDF
algorithm based on the word frequency of the document collection,
and Campos, Ricardo et al. had proposed the YAKE algorithm,
which makes full use of the basic features of the text, it uses 5
features: “Position of Word in the text”,” Word frequency”,” Term
Relatedness to Context” and “Term Different Sentence”. [
Deep learning, as an emerging supervised approach, provides
new ideas in keyword extraction. The basic approach is to vectorize
the candidate words by pre-training the model with word
embedding and then calculate the similarity to the document. Based
on this, the Embed Rank algorithm was proposed by
BennaniSmires et al. it uses both Sen2Vec and Doc2Vec, then a score will
be calculated by the Maximal Marginal Relevance (MMR) formula
which take into account similarity between text and phrase as well
as diversity of the keywords set.[
Although deep learning helps improve extraction precision, training such models often requires large amounts of annotated data, which is expensive to gather. In the case of keyword research in policy texts, especially in S&T policy texts, the data are mostly semi-structured long texts, and large-scale annotation training would be very difficult. In addition, for the traditional unsupervised algorithm, it will not have a high accuracy rate in a specific field, especially in our focus on S&T, and the extracted keywords are often not close to the topic of S&T.
To address these issues, we combine automatic word separation and technology domain word lists to achieve simple and fast entity extraction in the technology domain, including the extraction of named entities, and based on this, we achieve keyword extraction by the weighted optimized TF*IDF algorithm, and also perform simple statistical analysis of disruptive technology entities. Our approach combines supervised and unsupervised algorithms, and since it focuses only on the technology domain, it is relatively easy to build our corpus, and the keyword extraction based on the technology domain word list itself focuses on the thematic and semantic relationships between keywords and documents. 2
By researching major global S&T policy websites, we determined the search strategy and key websites, followed by automatic crawling of S&T policies using python to form a database of disruptive technology policy texts. Afterwards, we judged the initial data for disruptive technology relevance and screened out 1005 policy texts with high relevance. For these experimental texts, we performed automatic keyword extraction and specific disruptive technology extraction, respectively. For keyword extraction, we performed automatic word segmentation by self- designed automatic word segmentation based on spacy, then determined the TF*IDF matrix of candidate words and the overall text collection by screening the domain keyword list, and finally optimized the weights based on the self- designed core words and technology. The final results and rankings are calculated based on the selfdesigned core words and technology related word list for weight optimization.
The main large websites crawled are the UK government website, the EU publications website, the US Center for Strategic and International Studies, etc.
In terms of relevance evaluation, considering that the occurrence of keywords is not limited to fixed pairings, for example, disruptive and technology may not appear next to each other, but if the keywords are split and then matched statistically, the accuracy rate will be reduced, so we took a compromise approach, that is, to detect whether keyword pairs appear in a sentence.
In general, the keywords in the title and abstract of an article are relatively more important, so the content of the article is divided into three parts according to the position: title, abstract and body, and the corresponding weights are 5,3,1. In addition, only considering the word frequency will ignore the influence of the length of the article on the score, so a balancing factor of the average article length is added to the formula.
Here is the formula for calculating the correlation score: The meaning of each of these symbols is as follows.
N： Total number of keywords in a document b：A free constant that specifies how much the document length affects the score k：Free constants to specify the upper limit of the impact of a single word on the rating ld：Length of the document lavg： Average document length
The total number of texts crawled was over 10,000, and 1005 of them were identified as experimental texts after relevance screening. 2.2 The specific process of keyword extraction 2.2.1 Words segmentation and entity extraction. Specific process: document reading and sentence slicing, using Spacy natural language processing tools, specifically using Spacy's dependent syntactic analysis to determine the predicate of each sentence, that is, the root of a sentence, as one of the basis for slicing, using Spacy's deactivation table to mark the deactivation words, and then mark some specific lexical words, specifically including'ADV','AUX','CONJ','INTJ','NUM','PRON','SYM','SCO NJ','PREP'. The above words are used as cut nodes to split the words and get the preliminary splitting results.
2.2.2 Disruptive technology lexicon construction and weighted TF*IDF calculation. The first step is the identification of core terms, key terms and domain word lists. Considering that the topics of the texts we collected are all about technology policy and disruptive technologies, when extracting keywords we need to prioritize the words that directly match this topic, so we first defined 27 core terms as follows: disruptive innovation, radical innovation, innovation, disruptive technology, incremental innovation, open innovation, new product development, business model, absorptive capacity, technological innovation, developing technology, advanced technology, integrated technology, future technology, promising technology, next generation technology, evolving technology, radical technology, Next Big Thing, radical technology, breakthrough technology, game changer, gaming changing technology, emerging technology, revolutionary technology, transformative technology.
A crawler program on Web of Science was used to obtain the search results of the advanced search formula “TS = (disruptive technology OR disruptive innovation)”, then we extracted all the keyword fields and performed a simple word separation process and lemmatization. Then the keyword frequency statistics were arranged in descending order. The keyword statistics were then used as the base keyword list, followed by manual identification to construct a keyword list of specific disruptive technologies. Finally, 7400 base keywords and more than 300 non-repetitive technologyspecific words was collected.
Since Web of Science integrates academic journals, invention patents, academic conferences, academic websites and various other high-quality information resources to provide academic information in multiple fields, it is accurate and reasonable to grasp the dynamics of disruptive technologies and keywords in academia through Web of Science.
Finally, we constructed a TF*IDF matrix of 1005*5246 based on the results of word separation to determine the words and word frequencies for each text. For the core words, we multiplied the TF*IDF results by 20 as weights, and the words appearing in the specific technical word list were multiplied by 10 as weights. Finally, the top ten words in weight for each document were filtered as candidate keywords.
Through the previous collection of more than 300 disruptive technology entity words, we used python's FastText tool to match these entity words to get the results of each text technology entity extraction. The process of the entity extraction part does not have much detail and focuses on the analysis of the results
In order to effectively measure keyword extraction algorithms, we will compare them with some mainstream unsupervised keyword extraction algorithms. These algorithms include: Yake, TextRank, KeyBert.
3.1.1 P/R/F-Score comparison without considering keyword order. they are domain-independent and more applicative. However, since our proposed algorithm is supervised algorithm based on domain word lists, it has some degree of advantage in terms of accuracy and recall. If these unsupervised algorithms are optimized in combination with domain word lists, they will be more effective.
We conducted statistical analysis on the keyword extraction results, selected the top 120 keywords in terms of word frequency, and generated the word cloud map by python program as follows
Figure 3 Classification statistics of hot technical words Observing the word cloud map, it can be seen that the experimental text has numerous disruptive technology hotspots, involving information security, artificial intelligence, Internet of Things, new energy and even military technology fields.
We performed a simple count of high frequency words from the results of disruptive technology entity extraction and categorized them by domain.
As Figure 5 shows, the hot high frequency words in the chart mainly belong to the fields of artificial intelligence, air and space technology and new generation communication technology, such as AI and Machine Learning in the field of artificial intelligence, aircraft and satellite in the field of air and space technology and internet, ICT and 5G in the new generation communication technology, which all reflect the technology hotspots in these fields .
As for the disruptive technology entity extraction part, the method is mainly depend on a manually designed words list, so it has great scope for improvement. There are existing excellent named entity recognition (NER) algorithms that can be utilized, for example the currently popular named entity recognition model of BILSTMCRF via BERT pre-training. And Emma Strubell et al. first used IDCNN for entity recognition, reducing the time complexity of the
In the keyword extraction part, considering the special characteristics of scientific and technical texts, i.e., the focus topic of the text is often the current development of a scientific and technical field or technology, we enlarge the weight of scientific and technical terms so that the keyword extraction procedure can be more sensitive to these field terms and often extract the less frequent but important scientific and technical field terms. The results of the automatic word screening are intuitive and measured well because of the use of a manually developed word list and the increased weight of technical terms in the calculation of TF*IDF.
However, considering the subjective nature of the manually developed word list, the fact that the content of the word list limits the determination of candidate keywords and phrases, and the fact that the field of S&T is developing rapidly and the nouns of S&T are changing rapidly, it is not possible to extract technical noun entities and candidate keywords by relying on the manual word list alone.
The keyword algorithm is insufficient: it does not have the ability to extract new words, and the extraction results are overly dependent on the word list, so it needs a great degree of optimization. Optimization direction: Later on, neural networks and deep learning algorithms can be used to extract a wider range of noun entities in the field of S&T in combination with the word list, and the ability to extract words outside the word list is enhanced.
In specific disruptive technology extraction part, the extraction in this paper relies entirely on the manually developed word list, and later we can also rely on deep learning algorithms for optimization. Through stronger autonomous extraction, statistical analysis, we will optimize discovery of hotspots in S&T fields and make preparation for domain knowledge extraction.
This work was supported in part by The National Social Science Fundation of China (Number: 17BTQ066).