Patent analysis for technology development of artificial intelligence: a country-level comparative study.

Author:Tseng, Chun-Yao

As humans pursue better life quality and competition strength, knowledge economies are attracting considerable attention not only from countries, but also from businesses in general. 'Knowledge economies' is an important concept that addresses how to maintain learning experience and improve learning efficiency by using computer systems, which is also a core concept of Artificial Intelligence (AI). That is to say, AI plays a key role in knowledge economies, because it can be used to develop systems that think like humans, act like humans, think rationally, and act rationally (Russell & Norvig, 2010).

AI can be usefully grouped around five basic families, each with distinctive evolutionary prospects: Expert AI, Autonomous robots, Cognitive prostheses, AI theory and algorithms, and Turing Test AI (David, 2006). Russell and Norvig (2010) introduced AI using the following topics: (1) Intelligent agents; (2) problem-solving; (3) knowledge, reasoning, and planning; (4) uncertain knowledge and reasoning; (5) learning; and (6) communicating, perceiving, and acting. The field's historical trajectory from a focus on isolated and sophisticated mental faculties, to a focus on the commonsense knowledge needed for everyday tasks, and more recently to a focus on the construction of complete, situated, and embodied agents, is a trajectory form a priori conceptions of intelligence toward theories derived from the natural forms of intelligence that we observe around us. The logical extension of this trend is to model not only the products of natural evolution, but also its process (Spector, 2006).

Patent information plays an important role in the era of knowledge-based economies. Patents can encourage innovation and economic growth under certain conditions, and hamper it under others. The impact of patents on innovation and economic performance is so complex that a fine-tuned patent system is crucial to ensure maximum benefit for a country's firms and its overall economy. Consequently, this study employs patents as a major measure to elucidate technological innovation. Although several studies focus on the technical issues of AI, to the best of our knowledge there is not study addressing the technology development of AI based on patent analysis.

Restated, this study investigates three main issues related to the technology development of AI. First, the aggregate technology development of AI is examined, and different sub-technological fields of AI are compared. Second, this study employs measures of patent quantity and patent quality to demonstrate the technology development of AI among different countries. Finally, we investigate the technological position of different countries in different sub-technological fields of AI.


Typically, studies have examined patent count as a measure of technology. Patents granted are assigned a higher value than mere patent applications, because patents are only granted when they contain technological innovations exceeding a certain level of newness (Sherry & Teece, 2004). In this study we selected patents granted from the United States patent and trademark office (USPTO), as the most reliable proxy of measuring technology development. Because USPTO provides full text descriptions for patents issued from 1976 to the present, we restricted the research period from 1976 to 2010. Patent examiners assign to each patent application one main technology code, and one (or more) secondary technology codes. The two most widespread systems of patent classification are US patent classification (UPC) and international patent classification (IPC). This study adopts the UPC patent classification system. The class 706 is entitled 'Data Processing: AI'. The USPTO defines the class 706 is a generic class for AI type computers and digital data processing systems and corresponding data processing methods and products for emulation of intelligence and including systems for reasoning with uncertainty, adaptive systems, machine learning systems, and artificial neural networks. Class 706 is divided into 62 sub-classes in (United States Patent and Trademark Office, 2008), and their relative IPC sub-classes are shown as Table 1.

In this study, we divide the field of AI into four sub-classes: problem reasoning and solving, machine learning, network structure, and knowledge processing systems. This study adopts this classification, and assigns patents to these four sub-technological fields of AI, as follows:

(1) Problem reasoning and solving. Subject matter comprising a specific circuit arrangement for performing approximate reasoning where truth values and quantifiers are represented by possibility distributions. This sub-technology field includes sub-classes 1-11 of class 706 in USPTO.

(2) Machine learning: Subject matter wherein a system has the capability to automatically add to its current integrated collection of facts and relationships. This subclass includes induction, deduction, applications involving learning (i.e., data mining and knowledge discovery), and statistical learning techniques. This sub-technology field includes sub-classes 12-26 of class 706 in USPTO.

(3) Network structure. Subject matter wherein the system contains construction details of processors or their interconnections. This subtechnology field includes sub-classes 27-44 of class 706 in USPTO.

(4) Knowledge processing system: Subject matter wherein a system comprises specific domain data that is integrated as a collection of facts and relationships (i.e., knowledge representation), and applies a reasoning technique. This sub-technology field includes sub-classes 45-62 of class 706 in USPTO.


The unit of analysis in this study is based on the country level, and this study focuses on the top 10 countries in AI technologies. To compare the technology development in AI of different countries, this study employs patent quantity and patent quality measures to demonstrate different technology strengths of AI among countries. The measure of patent quantity is based on patent count data, whereas patent quality is measured by citation data (Bergek & Bruzelius, 2010; Bloom & Reenen, 2002; Guellec & van Pottelsberghe de la Potterie, 2001; Schoenecker & Swanson, 2002; Tseng, Hsieha, Penga, & Chu, 2011; Tseng & Wu, 2007). The analysis model used in this study is shown in Figure 1. Citation data have been utilized to measure inter-organizational knowledge flow (Hu & Jaffe, 2003; MacGarvie, 2005; Tseng, 2009). Backward citations have long been recognized as a rich and potential source for the innovation and technological evolution, which in turn brings value to a company. This study uses citation data as a proxy for technology flow. The number of backward citations in one country has been empirically used as a predictor of technology input from other countries. Conversely, the number of forward citations is used as a measure of technology output from a country.


Table 2 presents the number of AI patents issued during the period 1976-2010. The total number of AI patents around the world before 1998 was low, and it incrementally increased after 2003, peaking in 2010. The total number of AI patents over the period 1976-2010 totaled 5,228. The number of AI patents issued in 2010 was more than 50 times those issued in 1997. The aggregate technological innovation of AI has made great progress each year. Moreover, the yearly growth is evident by the increasing trend of the number of AI patent issued each year (Table 2).

Figure 2 and Table 2 illustrate the comparison between the four sub-technological fields of AI during the period 1976-2010. Over the period 1976-2010, the sub-technological field of Knowledge Processing Systems had the highest patent count (2,695), Network Structure the lowest (516), and Machine Learning the second highest (2,474). The patent count in the field of Problem Reasoning and Solving was 758; third highest among the four sub-technological fields. Before 1998, all four sub-technological fields of AI have low patent counts. The patent counts began increasing incrementally after 1997, peaking in 2010. The patent count...

To continue reading