A Morphology-Driven Method for Measuring Technology Complementarity: Empirical Study Involving Alzheimer's Disease

Makri, Hitt, and Lane (2009) define the concept of technology complementarity based on patent classification, that is, the degree of attention of two patents to different narrow technology fields in their common general technology field. According to this point of view, the standard to measure technology complementarity is the number of patents belonging to the same category but not to the same subcategory.

This method is more theoretical and computationally simple than the technology complementarity measurement methods based on industry. The formula provided by Makri et al. for a technology complementarity measure is as follows: (1) complementarity(A,T)=Overlap All Patent SubcategoriesTotal Patents A & T−Overlap All Patent ClassesTotal Patent A & T*Total Acquirer Patents In Commen SubcategoriesTotal Acquirer Patents. \matrix{{complementarity\,\left( {{\rm{A}},{\rm{T}}} \right){\rm{ = }}{{Overlap\,All\,Patent\,Subcategories} \over {Total\,Patents\,A\& \,T}}} \hfill \cr { - {{Overlap\,All\,Patent\,Classes} \over {Total\,Patent\,A\& \,T}}*{{Total\,Acquirer\,Patents\,In\,Commen\,Subcategories} \over {Total\,Acquirer\,Patents}}.} \hfill \cr }

In formula 1, A represents the acquirer and T represents the acquired party. Complementarity (A,T) measures the technology complementarity of T to A, which is related to the size of the range of classes and subcategories. Based on the technology complementarity measure method provided by Makri, Shang (2016) offered an improved and more detailed measure of technology complementarity in the area of technology mergers and acquisitions. The refined formula is expressed as follows: (2) complementarity=∑ NIPC6 B Involves−NIPC6 both A & B InvolesNIPC6 in NIPC4*PN of IPC4 in APN of A. complementarity\, = \,\sum {{{NIPC6\,B\,\,Involves - NIPC6\,\,both\,A\& B\,Involes} \over {NIPC6\,\,in\,NIPC4}}*{{PN\,of\,IPC4\,in\,A} \over {PN\,of\,A}}} .

In formula 2, A represents the acquirer and B represents the potential acquisition target. IPC4 represents a subclass within the IPC classification and IPC6 represents a large group in the IPC classification. NIPC6 B Involves, NIPC6 both A&B Involves and NIPC6 in NIPC4 represent the number of IPC large groups involved in B in the IPC subclass, the common number of IPC large groups involved in both A and B in the IPC subclass and the number of IPC large groups in the IPC subclass respectively. PN of IPC4 in A and PN of A indicate the number of patents in the IPC subclass and the total number of patents involved in A, respectively.

An alternative method of measuring technology complementarity provided by Dong (2018) avoids these obstacles. His measurement formula is expressed as follows: (3) complementarity=∑IPC4i=1nNIPC6 B Involves in IPC4i−NIPC6 both A & B Involves in IPC4iNIPC6 in IPC4i=1−NIPC6 A Involves in IPC4i=1*PN of IPC4i PN. complementarity\, = \sum\nolimits_{IPC_4^{i = 1}}^n {{{NIPC6\,B\,Involves\,in\,\,IPC_4^i - NIPC6\,both\,A\,\& \,B\,Involves\,in\,\,IPC_4^i} \over {NIPC6\,in\,\,IPC_4^{i = 1} - NIPC6\,\,A\,Involves\,\,in\,\,IPC_4^{i = 1}}}*{{PN\,of\,IPC_4^i\,} \over {PN}}.}

Like before, in formula 3 IPC4 represents a subclass in the IPC classification and IPC6 represents a large group in the IPC classification. IPC6 B Involves in IPC4i IPC_4^i and NIPC6 both A&B Involves in IPC4i IPC_4^i indicate indicate the number of IPC large groups of B in subclass IPC4i IPC_4^i and the common number of IPC large groups involved in A and B in subclass IPC4i IPC_4^i , respectively. NIPC6 in IPC4i IPC_4^i and NIPC6 A Involves in IPC4i=1 IPC_4^{i = 1} indicate the number of IPC large groups in subclass IPC4i IPC_4^i and the number of IPC large groups of A in subclass IPC4i IPC_4^i . PN of IPC4i IPC_4^i represents the number of patents in the IPC4i IPC_4^i and PN is the total number of patents in the target technology field. PN of IPC4i/PN IPC_4^i/PN therefore indicates the weight of the subclass IPC4i IPC_4^i in the target technology area.

This perspective includes constructing the industry chain and computing complementarity. The construction of the industry chain includes determining the industry field according to the research content, analyzing the technology chain in the industry field, and identifying all the key technologies used in the industry chain. For example, Wu et al. (2018) constructed the industry chain of the power battery, then identified the target enterprises for cooperation on this basis and optimized the cooperation area; then matched the established industry chain and blank area to realize the potential R&D partner identification under the principle of technology complementarity. Zhang, Xiao, and Li (2014) built a complementary technology tree for industry technology according to the unique complementary structure of upstream and downstream industries in an industry chain, and used text mining to judge the position of a patent in the complementary tree to measure technology complementarity.

The approach from the perspective of industry is based on industry codes. If the industry codes of the acquirer and the object of the acquisition are not the same, their M&A complementarity is judged according to whether their industry codes are relatively close. The method is simple, but the industry code classification is broad and guidance regarding classification is not clear. As a result, calculations cannot be performed accurately, which restricts the researchers to perform only qualitative analysis. For example, Xu et al. (2009) calculated technology complementarity by judging the proximity of industry codes, and then used those calculations to judge the degree of technology complementarity for M&A.

The first step in the research framework begins with downloading patent data from the Derwent Database. The Derwent database is a comprehensive global patent database covering all technical fields (Madani, Daim, & Weng, 2017; Sampaio et al., 2018). The data in the Derwent database comes from 48 patent issuing agencies worldwide (Oppenheim, 1982). The knowledge value of Derwent is the result of a complete classification, abstraction and index editing process. It rewrites the original title and abstract to reveal the actual invention and highlight the main uses and advantages of the technology, making the content clear and easy to understand (Wolter, 2012). Based on the advantages of Derwent database, this article selects the Derwent database as the data source database. The patent data are then reprocessed and converted into a format that SemRep can recognize. The original SAO result is directly extracted by Natural language processing tool SemRep based on UMLS and then technology-related SAO is directly screened through preprocessing.

The Unified Medical Language System (UMLS) is a database system for biomedical research developed in 1986 by the National Library of Medicine (NLM) (Alonso & Contreras, 2016). The Metathesaurus, Semantic Network, Specialist Lexicon and lexical tools are composed using the UMLS. The Metathesaurus contains information on biomedical and health-related concepts, various names, and their relationships (Pivovarov & Elhadad, 2012). The Semantic Network consists of two parts: a catalogue of semantic types and a second catalogue of semantic relationships (Bodenreider & McCray, 2003). The Semantic Network has a wide range of semantic categories that allow for the categorization of various terms in multiple domains, including 133 semantic types and 54 semantic relationships (McCray, Burgun, & Bodenreider, 2001). With the exception of “is a” relationship, the other 53 semantic relationships are classified into five categories: physically related to, spatially related to, temporally related to, functionally related to, and conceptually related to (Long et al., 2019). The semantic relationships of functionally related to and conceptually related to are linked to technology. The SAO of semantic relationships are functionally related to and conceptually related to cleaned and acquired.

The SAO structure is composed of a subject (S, noun phrase), an action (A, verb phrase) and an object (O, noun phrase) (Choi et al., 2012). For our purposes, the subject represents the technological means and the action and object (AO) combination is the functional concept and represents the problem solved (Moehrle et al., 2005). In this paper, S and AO are extracted from the SAO semantic structure, and an S similarity matrix and AO similarity matrix are calculated based on the UMLS. Then data analysis software is used to perform a cluster analysis. Through literature research and industry background analysis, the key technology issues and methods are then identified according to the clustering results.

The process relies on measuring the semantic similarity between two concepts. This paper adopts the method of similarity measure proposed by Lin, which is defined as follows, and can be used for measuring the semantic similarity of S (Resnik, 1999). (4) sim(d1,d2)=2×depth(lcs(d1,d2))depth(d1)+depth(d). {\rm{sim}}\left( {{d_1},{d_2}} \right) = {{2 \times depth\,\left( {lcs\left( {{d_1},{d_2}} \right)} \right)} \over {depth\,\left( {{d_1}} \right) + depth\,\left( d \right)}}.

lcs(d₁, d₂) represents the common nodes of d₁ and d₂ to the root of the concept hierarchy, Depth(d₁) and depth(d₂) respectively represent the depth from the concept nodes d₁ and d₂ to the root of the concept hierarchy (i.e. the information amounts of the synonym sets d₁ and d₂. 0 ≤ sim(d₁, d₂) ≤ 1, where 1 indicates that the concepts d₁ and d₂ are the same).

On the basis of the semantic similarity between two concepts, Park et al. (2013) proposed two similarity measures for sentences X₁ and X₂, the sentences X₁ and X₂ can be expected to correspond to d₁ and d₂ in formula 4, and each word in the sentence X₁ and X₂ is expected to have the same weight.

AO can be recognized as one sentence, the semantic similarity of AO is composed of two concepts of A and O, based on formula 1 and 4. In medical cases, the classification of A is concentrated and the similarity between A and A is very low, the weights of A and O are considered equally important, and the semantic similarity of AO calculation formula is expressed as follows: (5) sim(AOα,AOβ)=12×sim(Oα,Oβ)+12×sim(Aα,Aβ). {\rm{sim}}\left( {{\rm{A}}{{\rm{O}}_\alpha },{\rm{A}}{{\rm{O}}_\beta }} \right) = {1 \over 2} \times {\rm{sim}}\left( {{{\rm{O}}_\alpha },{{\rm{O}}_\beta }} \right) + {1 \over 2} \times {\rm{sim}}\left( {{{\rm{A}}_\alpha },{{\rm{A}}_\beta }} \right).

According to the results semantic similarity of AO and S, conducted the similarity Matrix of AO and AO, similarity Matrix of S and S, The Gephi mapping technique requires similarity Matrix as input, then using the drawing technology principle to form the cluster map. The cluster of AO represents “key technical issue”, and the cluster of S represents “key technical method”. The drawing technology principle is as follows: (6) W(X1,…,Xn)=∑i<jSij‖ Xi−Xi ‖2 W\left( {{\boldsymbol{X}_1}, \ldots ,{\boldsymbol{X}_{n}}} \right) = {\sum\nolimits_{i < j} {{S_{ij}}\left\| {{\boldsymbol{X}_\boldsymbol{i}} - {\boldsymbol{X}_\boldsymbol{j}}} \right\|} ^2} where the vector X_i = (x_i1, x_i2) represents the position of item i in the 2D map. where ||■|| represents the Euclidean norm. Minimization of the objective function performed under constraints. (7) 2n(n−1)∑i<j‖ Xi−Xj ‖=1 {2 \over {n\left( {n - 1} \right)}}\sum\nolimits_{i < j} {\left\| {{\boldsymbol{X}_\boldsymbol{i}} - {\boldsymbol{X}_\boldsymbol{j}}} \right\|} = 1

The constrained optimization problem of (6) minimization based on (7) is solved numerically in two steps. The constrained optimization problem is first transformed into an unconstrained optimization problem. The latter problem is then solved using a so-called majorization algorithm. The majorization algorithm is a variant of the SMACOF algorithm described in the multidimensional scaling literature. To increase the likelihood of finding a globally optimal solution, the majorization algorithm can be run multiple times, each time using a different randomly generated initial solution.

The Swiss astronomer Zwicky first coined and used morphology analysis (MA) in 1942 to develop jet and rocket propulsion systems (Zwicky, 1969). The basic idea of morphological analysis is to divide the subject into several dimensions, which describe the subject as comprehensively and as in a detailed manner as possible (Wissema, 1976), as shown in table 1. The subject is qualitatively decomposed into descriptive attributes and levels to explain the characteristics of the subject (Glenn & Gordon, 2009). The different attributes and levels constitute a series of possible choices for morphological analysis. Morphological analysis therefore offers a non-quantitative modeling approach for constructing and analyzing technical, organizational, and social issues by breaking down topics into several fundamental dimensions (Pidd, 1997).

However, the qualitative aspects of morphological analysis rely on expert intuition in the analysis process, so it cannot provide a quantifiable objective method to define attributes and levels (Wissema, 1976). In order to overcome this limitation, MA in different instances has been combined with text mining, F-term and joint analysis to minimize the intuitive dependence of experts. For example, Yoon and Park (2004) improved MA's usefulness in conducting technology predictions through joint analysis and bibliometric analysis of patents. Xu and Leng (2012) further developed MA by employing information technology to engage in patent text mining. The basic parameters of the morphological matrix are defined as factors in factor analysis, first by using the patent keyword matrix, clustering and factor scores, and then employing patent citations, patent registration year, and keyword frequency as influencing factors to evaluate morphological structure.

In this paper, we use the notion of MA to create a technology morphological matrix based on patent text mining. Through cluster analysis, text mining, literature research and expert consultation, we identify key technology issues and methods in the field. These issues and methods will be decomposed into several dimensions, with the core elements of the technology morphological matrix defined in terms of SAO.

The construction process of the technology morphological matrix is as follows:

Identify the key technology issues (also called dimensions or parameters). Assuming that there are k technology issues in the given field TQ₁, TQ₂……TQ_k;

Identify the key technology methods corresponding to k technology issues. It is assumed that there are f technology methods TM₁, TM₂……TM_f for k technology issues.

Establish the S and AO corresponding to key technology issues and methods and put the SAO structure into the technology morphological matrix for different institutions. The schematic diagram of the technology morphological matrix is shown in table 2.

The technology morphological matrixes of key technology issues and methods are then used to improve the accuracy and effectiveness of measuring technology complementarity. As shown in formula 8, we use an improved technology complementarity measure (based on formula 3) to calculate the technology complementarity between different institutions. Assuming that there are k technology issues in the given field TQ₁, TQ₂……TQ_k and f technology methods TM₁, TM₂…… TM_f for k technology issues, the technology morphological matrix of each organization is a matrix of f*k.

For example, consider Institutions IA and IB. The technology complementarity of IB for IA is calculated as shown in formula 8. (8) Complementarity(IA←IB)=∑k=1kTM IB Involves in TQi−TM both IA&IB Involves in TQiTM in TQi−TM IA Involves in TQi×TSAO of TQi TSAO. Complementarity\,\left( {IA \leftarrow IB} \right)\, = \sum\nolimits_{k = 1}^k {{{TM\,IB\,Involves\,in\,\,T{Q_i} - TM\,both\,IA\,\& \,IB\,Involves\,in\,\,T{Q_i}} \over {TM\,in\,\,T{Q_i} - TM\,IA\,Involves\,\,in\,T{Q_i}}} \times {{TSAO\,of\,T{Q_i}\,} \over {TSAO}}.}

In formula 8, complementarity (IA ← IB) indicates technology complementarity of IB for IA (where 0 ≪ complementarity (IA ← IB) ≪ 1). TM stands for technology methods, TQ_i stands for technology issue i, and TM IB Involves in TQ_i represents the number of technology methods in the technology issue TQ_i for the institution IB. TM both IA&IB Involves in TQ_i represents the number of technology issues in the technology issue TQ_i just both belongs to the Institutions IA and IB. TM in TQ_i represents the number of technology methods in the technology issue TQ_i, TM IA Involves in TQ_i represents the number of technology methods in the technology issue TQ_i for the institution IA. TSAO of TQ_i means the number of SAO corresponding to technology issue TQ_i in the technology morphology matrix for all research institutions in the target technology field while TSAO represents the total number of SAO in the technology morphology matrix for all research institutions in the target technology field. (TSAO of TQ_i)/TSAO characterizes the weight of the SAO for the technology issue TQ_i in the target technology field.

Based on the methodology outlined in section 3.1, 75 synonyms from the Metathesaurus for Alzheimer's disease were applied as the research terms for the time span of 2000 to 2018 (see table 3). 48,268 patents were identified as containing one or more search terms as a result.

In this paper, we selected the 539 patentees with unique patentee codes and whose number of patents numbered more than ten. The total number of patents for the 539 patentees was 31,998. We then used SemRep to extract the SAO structure of materials in the Medline database on PubMed. The 31,998 patents were converted into a format that SemRep could recognize. For an illustration of SemRep, consider the two semantic predications extracted from the input sentence in example (1). Arguments of the predications (subject and object) are represented as Concept Unique Identifier (CUI): Concept Name (Semantic Type). (1) MRI revealed a lacunar infarction in the left internal capsule.

Based on the methodology described in section 3.2, S and AO were extracted from the SAO semantic structures and a semantic similarity matrix was constructed based on the UMLS similarity calculation method for S and AO. Then perform cluster analysis on S and AO to find out the key technical issues and methods.

A total of 8,850 unique concepts were extracted from 122,769 valid SAO structures with a total of 6,410 concepts corresponding to S. According to equation 4, the calculation results obtained of the similarity among the S concepts are shown in Table 6, and all the S concepts obtained are listed in the appendix. In addition, a total of 8,850 unique concepts were extracted from 122,769 valid SAO structures with a total of 9,527 concepts corresponding to AO. According to equation 5, the calculation results obtained of the similarity among the AO concepts are shown in Table 7, and all the AO concepts obtained are listed in the appendix. It should be noted that among the AO results cases were eliminated where the A in the AO was the same but the similarity between O was 0 (because there were as many as 140,799 instances where this was this case. For example, in the case of C0001721 AFFECTS C0596991 myelination and C0001721 AFFECTS C0036690 Septicemia, even though A is the same any similarity is undercut by the similarity between O being 0).

The S and AO similarity matrix files were then imported into the social network software programs UCINET and Gephi for clustering and visual analysis. The clustering diagram of S and AO is shown in Appendix 3 and Appendix 4 respectively. According to the semantic similarity matrix of S and AO based on UMLS in 4.2, 425 clusters for S and 36 clusters for AO were obtained, corresponding to 425 key technology methods and 36 key technology issues in the field.

Based on the methodology described in section 3.3, 36 key technology issues were decomposed in this field into 36 dimensions, and 425 key technology methods were decomposed in this field into 425 dimensions. The basic elements of the morphology matrix were defined in terms of SAO to construct the technology morphology matrix. The construction process of the technology morphology matrix for each institution is as follows:

Identify the 36 key technology issues (i.e. Nephritis, Autoimmune Diseases; Neurobehavioral Manifestations, Memory Disorders, Learning Disorders; etc.).

Identify the 435 key technology methods (i.e. Acetic Acid-Related; Acids, Reagents, Solvents; Purines, nucleobase, pyrimidine; etc.); and

Establish the S and AO corresponding to key technology issues and methods and list the corresponding SAO structure in the technology morphology matrix for each institution. The technology morphology matrix for MERI-C is shown in table 8.

Based on the methodology outlined in section 3.4, we then constructed the matrix between 539 patentees and their corresponding 425 S-classes and the matrix between 36 AO classes and their corresponding 425 S-classes. Then we used equation 8 to calculate the technology complementarity between institutions based on SAO with the results presented in table 9. In the calculation results, the value of each cell represents the technology complementarity between the institution in the corresponding row and the institution in the corresponding column. For example, the value of cell (3,2) indicates technology complementarity of ABBI-C for AAKS-C is 0.0535.

Traditional IPC complementarity measurement methods must be used to generate results for comparison with a new morphology-driven method. The IPC method uses formula 3 to calculate the technology complementarity between institutions, with IPC-6 representing the big group of the IPC classification number and IPC-4 representing the big class.

We used VantagePoint software to construct the matrix between 539 patentees and 309 IPC-4 classification numbers and the matrix between the 309 IPC-4 classification numbers and 1,461 IPC-6 classification numbers. The results are shown in table 10.

The repetition rate indicates the proportion of the number of similarity values that appear multiple times to the total number of similarity values. The lower the repetition rate of calculation results between R & D institutions the higher the degree of distinction, indicating a better method by which to distinguish the technology complementarity between different R & D institutions. Similarly, the lower number of times the value 0 results from calculations between R & D institutions the higher the degree of fineness, indicating more nuance in a method's ability to recognize the degree of technology complementarity between different R & D institutions.

Figure 3 shows the findings of a statistical analysis comparing the calculation results (both the repetition rate and the number of value 0 results) between the technology complementarity measurement method based on SAO and the method based on IPC. From figure 3 we can see that the repetition rates of statistical results using the SAO and IPC technology complementarity measurement methods are 92% and 95%, respectively. The number of value 0 results based on the SAO and IPC methods are 713 and 1494, respectively.

Figure 3

The results show that the repetition rate and number of value 0 results using the SAO method are lower than when using the traditional IPC method. The SAO method is therefore an improvement over the traditional IPC as it generates higher degrees of distinction and fineness.

In addition to the comparative analysis via statistical results, experts from Institute of engineering medicine, Beijing University of Technology, Xuanwu Hospital of Capital Medical University, Institute of Medical Information were invited to re-rank the top 10 of MERI-C research institutions, as obtained by the two technology complementarity measurement methods. Then, the sum and average of the absolute value of differences between the rankings were calculated by technology complementarity methods with rankings based on experts. The smaller the sum and average value, the smaller the difference between the rankings based on technology complementarity method with rankings based on expert knowledge. The result was closer to the rankings based on experts, and the results were more reliable. As shown in table 11, the sum and average of the absolute value of differences between the rankings based on SAO technology complementarity method with rankings based on experts were 26 and 2.6, respectively; the sum and average of the absolute value of difference between the rankings based on IPC technology complementarity method with rankings based on experts were 48 and 4.8, The sum and average of the absolute value of the difference between the rankings based on the SAO technology complementarity method with rankings based on experts are smaller and much closer to the data presented by experts. The SAO method is therefore an improvement over the traditional IPC as it generates higher degrees of distinction and fineness. Therefore, the improved technology complementarity method based on SAO is more of a supplementary and refined framework for the traditional IPC method.