ABSTRACT. The outbreak of coronavirus disease 19 (COVID-19), the disease caused by the severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2), has shaken the world causing a global crisis in a completed unexpected way not seen in years. The rapid spread and its severity have incited scientists all over the world to investigate its causes, symptoms, treatments and effects, resulting in a huge number of publications and articles in just a few months. This overwhelming amount of information complicates access to proper investigations and facilitates the inclusion of non-relevant studies that can delay critical activities. Our goal is to determine the best way to categorize documents, determining which are the ones most relevant to different groups, such as policy-makers or biomedical community, to advance in their investigations, overcoming information overload. We have proposed five classes for a predefined COVID-related corpus (CORD-19), demonstrating that some of the articles included have no connection with the subject, and that the relevance of each paper is highly dependent on the specific area of study. Promising results were obtained making use of a simple model that combines word embeddings, topic modeling, and a Support Vector Classifier.

ABSTRACT. Technology plays a key role for firms in maintaining a competitive position in a fierce market. Disruptive technology, as a highly discontinuous type of innovation, has the potential to change firms’ technological competition status. To help firms adjust technology innovation strategies timely to respond to this innovation opportunity brought by disruptive technologies, early identification of technologies that have the potential to become disruptive is essential. Considering that previous related research mainly used qualitative methods, neglecting the use of objective information, we propose a framework based on patent data, with the help of text mining techniques to identify potential disruptive technologies in biomedical field from an alternative perspective. In this research, we firstly collected patent documents by defining a technological problem based on a market demand. Secondly, technological solutions solving the above defined problem are mined by extracting SAO structures from patent documents, and then clustering algorithms is introduced to cluster these solutions into groups as our technological candidates. Furthermore, we evaluate these candidates’ potential of replacing mainstream technology using two steps: to evaluate replacement potential quantitatively by measuring scientific attention, technological emerging score as well as market potential, and to compare these candidates’ technical advantages using text mining techniques. Finally, we use cancer treatment as a case to verify the feasibility of this framework. The results demonstrate that the technological candidate related immunotherapy shows the highest potential to become disruptive.

ABSTRACT. Transcription factor (TF) is a general term for a variety of proteins or genes that regulate gene expression, which controls the activity of a gene by determining whether the gene’s DNA is transcribed into RNA. Alzheimer's Disease (AD) is a kind of neurodegenerative diseases which are associated with abnormal gene expression. However, the direct relations from literatures between TF and AD are very weak. This study aims to mining the latent relations between TF and AD by knowledge graph and semantic path analytics based on Literature-based discovery (LBD). Firstly, Subject-Predicate-Object (SPO) triples related are retrieved to construct a domain KG. Then, semantic paths with predications are extracted from KG using path traversal algorithm. After that we mine these semantic paths from four respects to reveal latent relations between TF and AD: key concepts identification, paths strength measurement and ranking, paths clustering, and emerging intermediate concepts prediction. Finally, the latent relations will be visualized and interpreted. The study is in process and more details could be presented on the conference. It will help inspire research ideas and make new discoveries to scientists.

ABSTRACT. There has been an increase in the number published documents in the recent years making the process of text classification a vital tool for effective use of the information. Manual procedures for text classification which were fruitful for a handful of documents, lacked credibility with the growing number of documents. Text mining techniques facilitate assigning text strings to categories rendering the process of classification fast, accurate, and hence reliable. This paper classifies chemistry documents using machine learning and statistical methods. The procedure of text classification has been described in chronological order like data preparation followed by processing, transformation, and application of classification techniques culminating in the validation of the results. Keywords

ABSTRACT. This paper addresses the expertise-based solution for rising stars finding, which means identifying scholars with potential to become tomorrow’s experts in different domain expertise subjects, they play a vital role in driving future advances of their respective expertise. Current rising stars finding approaches used performance-based solution relying on using indicators of productivity and impact, while there are more realistic needs for knowing the expertise-related information. We propose a novel NMF-based hierarchical topic model approach for in-depth measuring of expertise, that can provided by extracting domain expertise topics and domain topical hierarchies, these are applied to rising stars find, for example, a measure of expertise potential based on topical hierarchies, that can provide valuable information for evaluating whether a scholar has in depth knowledge of an expertise topic. Overall, finding rising stars task, including both expertise knowledge and individual performance potential analysis, with respect to the expertise knowledge analysis, we estimate the domain expertise relevance score and domain expertise knowledge level of each person. With respect to the individual performance analysis, we estimate the time-weighted experience score of each person. Expertise relevance score, expertise knowledge level and experience score measure different aspects of scholar’s expertise, and can be combined to be used to rank and evaluate scholars for rising stars finding. Through an empirical analysis, we demonstrate rising stars for different domain expertise topics and different expertise hierarchies, it allows the seekers have a rising stars directory which can identify other rising stars who associated with each rising star’s expertise topic.

ABSTRACT. Keywords are consensus expressions of a concept formed by researchers in a specific field. They are the words that express the main topics of your research article. The emergence of new keywords performs better in representing the innovative state of scientific research in a specific discipline. Therefore, identifying emerging keywords can monitor the evolution of research topics and find valuable researches early. In previous studies, the bibliometric method has been widely adopted into the pioneering works of identifying emerging topic or technology. Novelty and growth are the earliest adopted and most basic characteristics of an emerging topic. Time feature is the essential element of measuring novelty, and it is an objective physical variable that is not affected by external factors. However, it is inadequate for different keywords to characterize novelty only considering the time factor, because different keywords may have personalized novelty at the same time point. Therefore, we proposed an approach that can measure the dynamic novelty of the individual keyword.

ABSTRACT. Citation count is always used to evaluate the scholar influence of a paper as more citations probably mean more endorsements received. Citation count is based on two questionable assumptions: the equal contribution assumption (i.e., each citation contributes equally to the citing paper) and the positive endorsement assumption (i.e., each citation is viewed as an endorsement from the citing paper to the cited work). Obviously, neither of these assumptions hold true. In this study, citation count are merged with two components of citation content analysis–purpose, which is the reason for the citation, and polarity, which reflects the author’s attitude toward the cited work – to provide a new convincing perspective for highly cited paper’s influence. Polarity is divided into three categories – positive, negative and neutral – and purpose into six categories – critical, comparative, used, substantiating, foundational and neutral. The full text of 100 highly cited papers and their citing papers are downloaded from Pubmed Central, then the citation contexts in citing papers are extracted and input to CNN (+Word2Vec) to identify and classify citation polarity and purpose with 89% accuracy. The result shows citations’ distributions across polarities and purposes in different periods for each paper as well as how domain, paper type (e.g. method, application, theory and review) and publish time influence distributions. Citation count, polarity distribution and purpose distribution are combined to evaluate and validate the paper’s influence based on paper type, domain and publish time.

ABSTRACT. This study compares the use of convolutional neural networks for the classification of scientific publications with a graph embedding approach that adds citation data. The underlying assumption is that publications are citing prior work that is highly relevant and that is, up to a certain extent, can be attributed to the same field as this cited work. Text classification models based on convolutional neural networks do not have access to these additional information sources but graph embedding do. Their predictions or classification models can take features, labels and properties of the neighborhood into account. This study applies both approaches on a set of publications and in order to classify them to 9 different categories within non-internal medicine. The results show similar results in obtained accuracy but much shorter training time for the graph embedding

ABSTRACT. As a critical approach for patent analysis, technology/ function matrix is capable of developing an in-depth understanding of patent documents. However, so far it is mainly obtained by expert reading, which is laborious and inefficient. To address this limitation, a novel method is proposed in which two deep learning models, BiLSTM-CRF and BiGRU-HAN, are respectively used for entity identification and semantic relation extraction. Then, two structures are generated which include technology ontology and technology-function network. Finally, by transforming each entity to its upper-level one in technology ontology (i.e., concept generalization), one can obtain technology-function networks with different granularities via network shrinking, and transform them into matrix form. For an empirical study, TFH-Annotated Dataset is applied for models training. Then, 1705 patent abstracts pertaining to thin film magnetic head in hard disk drive field are collected to generate technology/ function matrix. The result verifies the validity and feasibility of the proposed method. There are still some problems left in our method, among which the main problem is how to deal with ambiguous entities. By ambiguity, we mean that an entity mention with different references in different contexts, or multiple entity mentions refer to the same entity. In near future, entity linking techniques will be incorporated into our method for better Technology/Function Matrix generation.

ABSTRACT. In an open innovation era, many firms are struggling to find proper collaborators for successful performance of R&D projects. Thus, this research aims at proposing a new approach to exploring R&D collaborators by applying a link prediction approach based on doc2vec algorithm. The doc2vec-based link prediction can forecast future links between documents by reflecting the context of documents. If a pair of firms that has no connection with each other in a patent network has many potential links, they can collaborate to complement their R&D competences. For this, first, we collected citation information and textual information of patents, and drew a patent network. Then, a collaboration network of firms was generated by aggregating the nodes of patent documents that each firm possesses. Second, we investigated missing links between patents, and converted the textual contents of patent documents into vectors, based on doc2vec algorithm. We calculated the similarity between patents in missing links of a patent network with the document vectors, and potential links of patent documents could be predicted based on the cut-off value of the similarity. Then, the results of link prediction in the patent network are transformed into forecasting potential links in the collaboration network. Third, we compared the proposed approach with the Adamic–Adar technique, one of the traditional link prediction techniques. In this research, we applied the approach to automobile technology field to validate its applicability. We found that the proposed approach shows better prediction performance in exploring relevant collaborators than traditional techniques.

ABSTRACT. The estimation of similarity between patents is a critical issue for algorithmic data-driven patent analysis. Currently, there are three main ways to measure patent similarity: IPC code analysis, citation analysis, and keyword analysis, they represent three types of relationships between patents. Still, with the rapid development of deep learning and natural language technology, there are more options for obtaining patent similarity. In this study, we attempt to give the patent’s IPC labels, citation relationships, or patent’s textual information to neural network models for performing the patent similarity task. The neural models used in this study are: the Node2vec model for learning patents’ co-citation relationships; Doc2vec model for learning 3 million patents’ textual representations; Pre-trained BERT model for learning 30 thousand patents’ IPC information by fine-tuning, also learning patents’ textual representations by post-training. 7 test datasets (4 auto-constructed sets and 3 manually constructed sets) with ground- truth were created to complete this verification task. The motivation behind the research is to find a suitable representation for patents, to support subsequent patent analysis and applications for R&D decision making and technology forecasting.

ABSTRACT. For public policy science scholars, evaluating how effective a given policy and regulation is, and profoundly assessing whether it contributes to social betterment is essential. In the previous studies on S&T policy evaluation, scholars often focus too much on economic values, while ignoring its deep social value led by assessing the equity of distributional consequences and policy concern on the views and interests of stakeholders. In this research, we take China’s S&T policy as a representative case study and utilize the policy citation data in academic papers as clues to quantitatively track how scientific research response to S&T policy. We initially selected 18 of the most influential S&T policy documents from 1978-2019 as object files for data retrieval. Then, we collected literature data from China National Knowledge Infrastructure (CNKI) by using the title of a targeted policy text to conduct a full-text search from 1978-2019 and finally obtained 14323 relevant records after data cleaning. The results show that the first peak for researchers reacting to a given policy document is about one year after the policy being issued.  And as the stakeholders of S&T policy, researchers' attention and discussion on policy documents are characterized by timeliness, permanence, comprehensiveness and profoundness. Universities and scientific research institutions affiliating to government agencies are positively responding to policy documents. And the main topics they focus are S&T plan, comprehensive regulations, S&T system reform and intellectual property. Moreover, the intentions for researchers citing policy documents in their papers are different, which can be divided into various kinds by observing papers' relevance to the content of policy texts. 

ABSTRACT. Since the Sustainable Development Goals (SDGs) have been introduced by the UN, many initiatives have proposed attempts to link R&D outputs (e.g. scientific publications), to goals and targets. From the policy perspective, results from these mapping exercises can provide better understanding of what research contributes to achieving the SDGs. From the funding perspective, such connection could show the potential impacts of research. Most of these approaches are based on keywords, or ML models trained on text: SDG related terms are searched for in publications (e.g., titles and abstracts). There are at least two limitations with such approaches. First, the language used in the SDG description relates to the political (societal) goal, while language in a publication relates to the scientific goal. Also, by searching for direct links (using language), a significant amount of research (e.g. basic research) may not be picked up since most “policy” language is found in applied research. Second, research is a collaborative and cumulative effort. Collaboration and controversies feed scientific progress. Linking individual publications to SDGs would ignore the value of the community working on a certain topic. We developed an approach which links research communities, i.e. publication clusters on a topic, to SDGs. Communities may be linked to a SDG more or less strongly, reflecting the idea that there are no strict boundaries for relevant R&D. Our approach also provides a platform to evaluate other approaches and contribute to a more open, robust definition and understanding of what R&D should be linked to an SDG.

ABSTRACT. International technology diffusion thus has become a popular topic in the literature on economics and technology policy, yet research gaps still remain. Interesting phenomenon in AI area indicate that frequent and massive international technology diffusion does not benefit all countries at the same level, but intensifies the trend of polarization, and the gap between leading countries and other countries is widening with each passing day. By carrying out empirical analysis on the basis of patent citations data, tries to answer the question that what are the consequences that international AI technology diffusion draws upon national innovation process in leading countries and others respectively? The focus of this study is the internationalization degree of technology diffusion, which aims to describe the relationship between domestic technology diffusion and cross-border technology diffusion. It is measure by the ratio of the number of citations by applicants from other countries to the total number of citations. Our empirical results show that the structure of AI technological innovation shows an obvious trend of polarization, where China and the United States are currently leading international AI technology diffusion process, while other countries have been drawn into the passive process of international technology diffusion, especially Japan. Besides, being passively involved in the international technology diffusion process means that the technological innovation achievements of a country will be first identified and utilized by other countries, rather than spillover and enhancing the domestic technological innovation capability, which is contrary to the findings of existing research.

ABSTRACT. Our study shows that occupancy of brokerage positions in the regional R&D network is a crucial determinant of characterizing regional industrial identities. Particularly, the question of how structural position represents regional industrial dominance is dependent on the type of brokerage position occupied by technologies in regional R&D activity. In region-based government R&D, technologies in liaison positions are more influential for the relational dynamics in given heterogeneous R&D fields, whereas coordinator positions are influential in the linkage of technologies belonging to a homogenous group. The case study focuses on network topologies of technology fields in Daegu & Gyeongsangbuk-do Province, South Korea and illustrates how dynamics of brokerage positions influence regional industrial specializations. We study topologies of R&D networks by examining a matrix of the co-occurrence of the classification of technologies. We analyze 5,444 R&D projects which are performed in Daegu & Gyeongsangbuk-do Province, South Korea. Each project need to select maximum 3 technologies which represent topic of the project. In our study, those technology classification are used as nodes. In particular, the network method provides a systematic analytical tool to uncover the hidden structure of R&D relations representing regional industrial specializations and to monitor the effectiveness of regional-based R&D projects designed to foster R&D activities across different technology fields. The research question of the paper is accordingly: How dynamics of brokerage position affect regional R&D activities? The main question has implications for future program development efforts that would support the performance of regional-based R&D activities characterizing regional industrial dominance.

ABSTRACT. The aim of this presentation is to show a method for performing a strategic analysis of a company mobilizing different data sources and different techniques for Technology Intelligence. The strategic analysis we aim for is to be able to propose different scenarios of evolution of an innovation domain for a given company The method is tested on a use-case which is a German automotive supplier company, ElringKlinger, and focuses and their fuel-cell activity. We start by mobilizing a variety of datasources (patents, publications, European projects, annual reports, financial data) to study the recent evolution of the company. Some sources will be used as such, in other cases textual analysis is used to analyse a combination of data sources. The extracted information and computed indicators will be used to explain the different positions that the firm has taken. An analysis of the environment of the firm allows us to contextualize the different indicators. Using the conclusions we pull from these indicators we connect the different conclusions and build scenarios for the evolution of the company. Using this method, we are able to show how Technology Intelligence indicators and methods can be used to provide elements for the strategic analysis of a company that can also easily be completed with information provides by other departments inside the company (product and marketing for example).

ABSTRACT. The report, which takes a focus on Artificial Intelligence (AI), looks at the expertise held by Canadian researchers and institutions undertaken both domestically and abroad in this continuously evolving technology field. Artificial Intelligence (AI) is a technology area that has garnered significant interest in recent years, however measuring innovation pertaining to AI is a challenging task since the field involves a variety of different techniques that can be broadly applied across a wide array of industries. The findings reveal Canada ranks sixth globally, both in terms of the number of patented inventions assigned to Canadian researchers and to Canadian institutions. Canada’s rankings fall behind notable countries that file prolifically, namely China and the United States. The report divides the Canadian analysis in two main sections, presenting the Canadian patent landscape from the perspective of Canadian institutions and also of Canadian researchers. These two sections provide a detailed overview of filing activity looking at areas of specialization by AI sub-category, key players, geographical distribution across the country, and patent landscape maps. The details presented in these sections are useful to better understand the evolution and the current state of innovation in this technology field. The report showcases a number of complementary research activities undertaken including as section that derives from CIPO’s collaboration with Statistics Canada, as well as the latest metric developed at CIPO called the IP Concentration Index (IPCI).

ABSTRACT. Nowadays, most organizations are facing the challenge of tracking the latest technological developments and identifying technology opportunities or threats of the competitive environment. In this context, intelligence analysis methods have been widely used, and lots of technology intelligence techniques have been embedded into general purpose tools to support the need for extracting valuable information from textual data. However, there was no single tool powerful and flexible enough to incorporate all the key elements (data retrieval, preprocessing, normalization, analysis, visualization, and interpretation) in analysis process or have only in limited form. Therefore, obtaining such intelligence awareness, especially from textual data, remains one difficulty. In this paper, addressing concerns of competitive technological intelligence and remedying the shortcomings mentioned before, ITGInsight has been developed. It presents four key features that are remarkable in respect to other software tools: (a) powerful data preprocessing module; (b) flexible user-defined analysis module; (c) gorgeous data visualization module; and (d) professional automatic interpretation module. Finally, an empirical study for synthetic biology is used to describe ITGInsight in deeper detail. The experiment results show that it is a powerful tool for generating effective competitive technological intelligence, such as profiling science & technology domain, mapping research front relationships, and discerning overall trends, providing more insightful information to intelligent consumers especially in non/few-expert supported environment.

ABSTRACT. Patents are an interesting source of technical information. Solutions are generally disclosed in patents prior to any other public source; however, the content and the language used in patents are difficult to mine due on the one hand to the different fields, suitable for text mining, where the information may be written. On the other hand, patents tend to cover as much domain as possible in the case of drugs in order to protect any potential market for the assignee. Here is proposed a method for sorting out the first problem and partially the second.

ABSTRACT. Technology knowledge networks are dynamic and self-organized, they can be predictable by identifying and mining existing links to delineate potential technological opportunities. However, with open innovation has draw much more attention, there are still technological boundaries and barriers, especially in industry. Knowledge spillover between different enterprises can be difficult. This problem has often been ignored in technology opportunity analysis using knowledge networks, since usually we describe technology opportunities as missing links or incomplete sub-networks without considering if there is resistance for them to conduct from the perspective of knowledge flow. In this paper, we present a technology opportunities analysis framework using link prediction and patent network constructed with Derwent Manual Code (MC) from a knowledge flow point of view. Organizational attribute has been introduced for node proximity weighting. And Alzheimer's disease related patents have been retrieved for empirical analysis.

ABSTRACT. Interdisciplinary research has demonstrated its significance in promoting knowledge production and assisting problem solving. Despite that research funding is one of the drivers of discipline integration, potential connections between funding and interdisciplinarity were seldom under consideration, let alone from the perspective of research topics. In this paper, we propose a methodology to investigate funding patterns of interdisciplinary research from both article and topic perspectives. We first feature funding agencies with their organizational functions, and define governmental, academic and industrial funding sources. Based on ESI research fields proposed by ISI, we then uncover collaboration and funding preference among agencies, and detect the interdisciplinarity of funded articles and topics by incorporating network analytic techniques, bibliometric methods, and topic modeling approach. To demonstrate the feasibility and effectiveness of our methodology, we extract 9980 funded articles from Web of Science with a time span of 2010 to 2019 to perform a case study of big data research. The empirical results of this study can provide decision support for agencies in exploring funding topics and updating funding strategies, thereby promoting interdisciplinarity in big data area.

ABSTRACT. Technology forecasting has long been of interest to research community. In this paper, we aim at predicting milestone technology at the embryo stage. The main contribution of this paper is, we explore temporal features of milestone technology by A/B testing. We have acknowledged dynamic characteristic of milestone technology is the pace of acceptance by scientific community and temporal features of citation graph. If a technology is cited at an accelerating speed, it is more likely to be a milestone technology. Based on observations of temporal features, we use a temporal embedding based methods to capture the characteristics of temporal citation network. The results showed the effectiveness of the temporal embedding based method, especially for identifying milestone technology at the early embryo stage.

ABSTRACT. Technology proves to be a focal driver of economic success, as technological change can influence the competitive environment to a considerable extent and in different directions. Particularly in the field of emerging technologies, novelty plays a crucial role besides other indicators. Novelty is not only characterised by scientific breakthroughs leading to new technological solutions; novelty may also be caused by utilising an existing technology for a new application, or a combination of both. While the terms of novelty are coined well, the question remains how to measure them. Recent approaches measure novelty based on patents only in a single dimension and do not differentiate between the two types of novelty. To achieve this additional requirement, we propose and develop a semantic based measurement. For this purpose, we proceed in five steps: generation of semantic technological oriented thesaurus, collection and pre-processing of patent data, calculation of novelty values for each bi-gram, generation of semantic anchor points and calculation of novelty value for each patent. Our research contributes to the understanding of technological oriented and application oriented novelty. From a managerial perspective, we enable experts to use a novel method to search for technologically and/or application oriented novel patents. Doing so, they are able to identify the driving institutions and inventors behind those patents. Furthermore, analysing the semantic anchor points from a content oriented perspective shows the changing themes of a technology’s development over time.