The paper discloses a new approach to emerging technologies identification, which strongly relies on capacity of big data analysis, namely text mining augmented by syntactic analysis techniques. The opportunities of the new big-data-augmented methodology are shown in comparison to existing results, both globally and in Russia. The integrated ontology of currently emerging technologies in A&F sector is introduced. The directions and possible criteria of further enhancement and refinement of proposed methodology are contemplated.
Technology identification and mapping exercises for effective science and technology
(S&T) and innovation policies shaping become less feasible without modern data
science techniques application. This happens due to the explosive growth of diversity and
quantity of available S&T information, drawbacks of human-performed analytics, as
well as overextended periods of foresight studies and budget limitations. The attempts
to solve the problem include tech mining [
For the purposes of emerging technologies (new technologies that might have a significant impact on the economic activity) identification, we see text mining / semantic analysis tools as the most appropriate, as identification of new man-made phenomena of known nature (technologies in this case) can be reduced to identification of new syntactic constructions signifying them. The fact that man-made artifacts tend to be explicitly named, described and discussed with the use of written language makes the problem well-posed.
To demonstrate text-mining-augmented techniques applied to technology
identification and mapping we consider the case of the agriculture and food (A&F sector). Our
choice is driven by the fact that large proportion of global challenges are directly related
to A&F sector [
Methodology The main hypothesis in this paper is that "emerging technology" as a signifying syntactic construction has not lost its semantic utility despite the hype around this concept. The analysis is based on the ample material of the two-year A&F sector foresight study, and relies on the capabilities of the Text Mining System of the National Research University Higher School of Economics (NRU HSE). Composition of data sources of the system include stratified random sample of summaries and metadata of top cited research papers and international patents, as well as newsfeeds items from tops of global news portals with science and technology flavor, analytical and forecast reports, declarations, proceedings and other documents in PDF format (all acquired through open access sources). At the time of the study, the system featured more than 12 million documents, several hundred million sentences, of which up to 3 million documents were at least partially relevant to A&F sector and adjoining sectors, such as biotechnology and bioenergy, more than 150 million terms - object signifiers (among which technologies are presented).
In this paper, we present one of possible approaches of technology identification, namely cascade identification of words being governors within terms. The method allows identifying unigrams – universal signifiers of semantic field of "techologicality", i.e. words that radically increase the probability of an n-gram containing them to be a name of certain technology. Examples of such words are technology, method, system, platform, model, tool, layer, enzyme and others. Extraction of all object-signifying words allows getting hundreds of thousands of terms – candidates for being names of technologies (for instance, DNA sequencing technology, or recirculating aquaculture system, etc.). These lists are filtered with the use of author-built machine learning algorithms dealing with "information-richness" of terms, their monopolism and specificity and other attributes.
Then, analysis of dynamics of presence intensity in the discourse during the last
years is conducted for the candidate technology-signifying terms. The relative
frequency of terms is calculated using the following formula [
, – the amount of sentences in the i-th document,
In order to calculate the dynamics of the candidate technology-signifying terms, we adapted the formula of average annual growth rate (AGR): where – the amount of years, for which the collection of documents is available, (2) A&F technologies identification results include the list of 181 items. A random sample is provided below: aeration technologies cultured meat technolhorticultural technoloagricultural conserva
dairy technologies ogies nologies gies
DNA micro array techgies
integrated soil fertility management technologies
LEISA technologies nologies algal biofuel technolofeed probiotics meat processing techbioconversion technolfertilisation technolosmart irrigation
Technologies were distinguished by dynamics of intensity of their presence in the discourse during the last years. It can be visualized as trend maps: 2-dimensional plots with one axis representing the popularity of a term and the other showing the year-byyear dynamics of the normalized popularity (relative frequency of use). For trend map of technologies in agriculture on media resources see Fig.1. The upper-right quadrant consists of the strongest topics shaping the future agenda of the sector, they are popular and gaining traction: in media they are exemplified by CRISPR technologies, agroforestry and aquaponic technologies, precision agriculture and microalgae technologies etc. The lower-right quadrant contains the so-called "weak signals": they are highly trending but underrepresented in discourse yet. They can contain the emerging technologies. This group presented by smart irrigation technologies, molecular breeding and zinc-finger nucleases technologies etc. Among the popular topics losing their significance are fertilisation, pruning, antifouling technologies and many more. The method applied in this study yields most results with data obtained on research papers and especially patents abstracts, which contain less low-informative terms than general reports of international organizations, and discuss technologies in more concrete terms. Within this approach, any individual terms are filtered out without an anchor term within. One of the consequences is that branded, trademarked and other proprietary technologies are almost not present in the output (with some exceptions, such as Round-Up pesticide, which name has gone almost denominative in the GMO application discourse, so that "roundup technology" were mentioned in the texts analyzed). The next steps of filtering the obtained lists of technologies may include building the semantic map that demonstrates dynamic classification, trend maps based on other sources of data, as well as hype maps that show difference in normalized popularity of topics in different data sources (e.g. media vs patents).
The limitations of such approach is high dependence on the marker terms. Some technologies may never co-occur with “technologicality” terms meaning the algorithm will miss them. In order to overcome this obstacle, future studies will concentrate on two main points: searching terms that are relatively more specific to patent literature compared to other sources of data (such as scientific publications, media news, analytical reports) as potentially technical terms, as well as using identified technology terms as a sample for machine learning based on word embeddings. In other words, the main hypothesis for the future studies is that terms that are semantically highly similar to technology terms (based on word2vec, GloVe or other approaches) are also likely to be candidates for being names of technologies.