
Collaborative knowledge exchange promotes
innovation
Tomoya Moria,b,1, Jonathan Newtona, and Shosei Sakaguchic
aInstitute of Economic Research, Kyoto University. Yoshida-Honmachi, Sakyo-Ku, Kyoto, Kyoto 606-8501, Japan.; bResearch Institute of Economy, Trade and Industry, 11th
floor, Annex, Ministry of Economy, Trade and Industry 1-3-1, Kasumigaseki Chiyoda-ku, Tokyo 100-8901, Japan.;
c
Faculty of Economics, The University of Tokyo, 7-3-1, Hongo,
Bunkyo-ku, Tokyo 113-0033, Japan.
This manuscript was compiled on November 7, 2022
Considering collaborative patent development, we provide micro-
level evidence for innovation through exchanges of differentiated
knowledge. Knowledge embodied in a patent is proxied by word
pairs appearing in its abstract, while novelty is measured by the fre-
quency with which these word pairs have appeared in past patents.
Inventors are assumed to possess the knowledge associated with
patents in which they have previously participated. We find that col-
laboration by inventors with more mutually differentiated knowledge
sets is likely to result in patents with higher novelty.
collaboration |knowledge |differentiation |novelty |innovation
H
umans are a collaborative species (
1
–
3
) and bring this
collaborative nature with them to the workplace and
to the laboratory. Collaboration within teams has been in-
creasingly important within academic research (
4
) and patent
development (
5
,
6
), and has long been an essential part of
other creative endeavors such as the production of Broadway
musicals (
7
). A necessary condition for successful collaboration
is successful coordination on roles and tasks (
8
,
9
), which can
be complicated when the characteristics of one’s collaborative
partners are imperfectly understood (
10
). Hence, collabora-
tion requires shared knowledge among collaborators. However,
for collaboration to be creative, it also requires differentiated
knowledge (
11
). There is, for example, evidence that the suc-
cessful production of Broadway musicals involves teams that
comprise both people who have previously collaborated and
people who have not previously collaborated (7,12).
Using Japanese patent data and adopting a measure of
novelty based on word combinations that appeared in the
patent (
13
), we find a strong positive relationship between
collaborators’ mutual knowledge differentiation and the novelty
of their output.
Results
Wp
represents the set of distinct word pairs in patent
p
’s
abstract. (Hereafter, a bold capital letter expresses a set, and
the corresponding italic letter its cardinality.) The novelty
of a word pair at a given point in time is measured by the
likelihood of its appearance in patents filed in the past (
13
).
Specifically, the novelty
nwt
of word pair
w
at time
t
is the
ratio of (i) the sum of
Wp
=
|Wp|
over all patents
p
filed at
dates up to and including
t
, to (ii) the number of these patents
that include word pair
w
. We measure patent novelty by the
average novelty of word pairs in its abstract,
1
WpPw∈Wpnwtp
,
where tpis the patent’s filing time.
We consider collaborative aspects of patent development
by focusing on the productivity per inventor pair, following
(
11
).
Hp
is the set of all inventors who participated in patent
p
, while
Mp≡ {
(
i
,
j
) :
i
,
j
,
∈Hp
,
i6
=
j}
is the set of pairs
of such inventors. The average pairwise-contribution to the
patent’s novelty is given by
np=1
MpWpX
w∈Wp
nwtp.
Denoting by
Git
the set of patents inventor
i
participated in
at time
t
, define
i
’s knowledge at
t
by
Kit
=
∪τ <t ∪p∈Giτ Wp
and its novelty by kit =Pw∈Kit nwt.
Inventor pair
{i
,
j}
has total knowledge
Kijt
=
Kit ∪Kjt
,
with novelty
kijt
=
Pw∈Kijt nwt
, and inventor
i
’s differenti-
ated knowledge relative to
j
is
KD
ijt
=
Kit\Kjt
, with novelty
kD
ijt =Pw∈KD
ijt
nwt.
Knowledge differentiation between
{i
,
j}
is evaluated by the
geometric mean of their respective differentiated-knowledge
shares in the union of their knowledge,
sijt =qkD
ijtkD
jit/kijt ∈[0, 0.5] .
Their average in patent p,
sp=1
MpX
(i,j)∈Mp
sijtp
measures knowledge differentiation in
p
. We focus on patents
with
sp>
0, since
sp
= 0 implies no knowledge exchange as
inventors can be indexed so that K1⊆K2· · · ⊆ KHp.
We estimate the effect of spon npby the model:
np=β0+β1sp+· · · +βmsm
p
+γ¯
Kp+δMp+fp+ϕp+τp+εp,[1]
controlling for average knowledge size
¯
Kp≡1
HpPi∈HpKitp
,
inventor-pair count
Mp
(reflecting the costs/benefits of coor-
dination and task specialization), and fixed effects,
fp
,
ϕp
,
and
τp
, for firms, classes of International Patent Classification
(IPC), and years, respectively. εpis a stochastic error.
The estimated conditional expectation and quantiles of
np
indicate a positive association between
sp
and
np
, except for
a range of small
sp
, while the observed
sp
are spread over the
entire feasible range, (0, 0.5] (Fig. 1).
To see the robustness of the result, we consider citation
count of a patent as an alternative measure of output. Let
¯cp
be
the citation count of patent
p
within five years of application,
excluding self-citations, where the self-citations include those
Author contributions:Conceptualization: T.M., J.N.; Funding acquisition: T.M; Theoretical analysis:
J.N., T.M.; Empirical analysis: S.S., T.M.; Writing; J.N., T.M.
The authors declare no conflict of interest.
1To whom correspondence should be addressed. E-mail: mori@kier.kyoto-u.ac.jp
1–3
arXiv:2210.01392v5 [econ.GN] 4 Nov 2022