This article is written by Poulomi Sen, from Rajiv Gandhi School of Intellectual Property Law, IIT Kharagpur.
Table of Contents
Abstract
Artificial Intelligence, a discipline of Computer Science, converges with Intellectual Property in several ways. AI system’s unique ability to exploit large volumes of data, identify several complex patterns, predict and compute the most optimal and ideal result, accurately and efficiently consuming very less time proves to be very beneficial in this fast paced world. The study focuses on how Artificial Intelligence can simplify the time consuming and laborious task of patent search by substituting the traditional manual techniques. The biggest drawback of using manual techniques for patent search is the escalation in data processing errors, which potentially could enhance the inventor’s risk of losing an asset. As per the estimation of WIPO, a quarter of patent information is erroneous; therefore the risks involved are quite evident. The automation of searching large databases for prior patent research, which is required to be done prior to the filing of a patent application in order to comply with the criterion of “novelty”, will result in the elimination of complexities and errors involved in the manual techniques of patent search. Thus AI can simplify the herculean task of patent search with more preciseness, efficiency and accuracy.
Keywords: Patent, Semantic based search, Artificial intelligence, Machine Learning, Neural Network, Prior art, Patent Search, Boolean operators, classification
Introduction
Artificial Intelligence is a set of machine learning algorithms fed to AI systems such that it demonstrates traits associated with a human mind and programmed to simulate human intelligence to achieve a specific goal, without or with minimum human intervention. Artificial Intelligence has wide ranging applications in various sectors such as healthcare industry, financial industry, Banking and e-commerce, entertainment and gaming. At present AI is widely accepted in the domain of Intellectual Property for assisting in management and administration of Intellectual Property systems and tools. Prior Art Patent Search is one of the most crucial steps that should be done prior to filing a patent application in order to comply with the requirements and criteria of patentability. The manual process of patent search can be a tedious and a laborious task as the IP professionals have to meticulously input data and search huge databases in order to conduct a prior art search. The adoption of artificial intelligence for patent search and patent related discovery will automate the time-consuming and burdensome process which will potentially amplify the efficiency, accuracy and quality.
Fundamental concepts of artificial intelligence
Artificial intelligence, a branch of Computer Science, is based on the principle that human intelligence can be structured in a manner that a machine can replicate to execute simple as well as complex tasks. Some of the programmable functions of artificial intelligence include learning, reasoning, and perception, planning, problem solving, and decision making. AI Algorithm, unlike other algorithms, simultaneously takes a combination of both the inputs and the outputs so that the system can learn the data and display desired output according to the given input.
Types of AI
Subsets of AI
AI systems are actuated with the help of algorithms, using:
- Machine Learning
- Deep learning/ Artificial Neural Network
Techniques of AI
There are basically four techniques that are used to make a system artificially intelligent:[i]
- Machine Learning
- Natural Language Processing
- Automation and Robotics
- Machine Vision
Machine Learning
The term “Machine Learning” gained popularity due to Arthur Samuel, an American pioneer in ML. Arthur Lee Samuel in 1959 defined Machine Learning as:[ii]
“A Field of study that gives computers the ability to learn without being explicitly programmed.”
In a nutshell, ML algorithms are used to feed computer data to AI systems using statistical techniques so that AI systems get progressively better at performing a task and producing the desired output without any explicit set of program:[iii]
Figure 1: Machine Learning: Making machines learn from data[iv]
Machine Learning can be done in the following ways:[v]
- Supervised Learning: The correct output corresponding to an input is well defined for the algorithm from the inception.
- Unsupervised Learning: The output to be predicted is not defined. Hence the output prediction is based solely on the algorithm’s discretion to characterize similar data and link it accordingly, as it is fed with no answers.
- Reinforcement Learning: It boosts up the productivity and accuracy as with every correct prediction of output, the algorithm is awarded with a reward.
- Ensemble Learning: Multiple learning algorithms and base models are used in order to enhance the predictive-performance of a system.
Deep learning/ Artificial Neural Network
Just like the interconnected-neurons which forms the basic working unit of our brain and as a result of which our brain has its structural and functional properties, artificial neural networks are those sine qua non units that tend to make computers work like a human brain by making them “learn” so that they can take decisions and produce output in a human-like manner.
Neural network is a big mathematical equation that constitutes three hierarchical layers:[vi]
- Input layer: Consists of initial data set which is fed for the NN.
- Hidden layer: Intermediate layer where all problems are solved by virtue of computation and calculation in order to predict and compute the output. At times more than one hidden layer is required for solving problems.
- Output Layer: Displays the desired output corresponding to the input.
Figure 2: Basic structure of Artificial Neural Network[vii]
Boosting AI algorithm’s Performance
AI systems can be made to work more efficiently and accurately if the machine is incorporated with:[viii]
- Big Data: More the data, better the performance.
- Large Neural Networks: Training a very large neural network, results in a more efficient performance.
Figure 3: Performance Comparison of Deep learning-based algorithms Vs Traditional algorithms[ix]
Traditional Prior Art Patent Search Methods
Prior to the filing of a patent application, a prior art search should be carried out in order to meet the prerequisites of patentability i.e. the criterion of “novelty” and “inventiveness”. Prior Art search is a due diligence exercise that defines the scope of protection in patent claims. The rationale behind conducting a prior art patent search is to:
- Avoid filing patents and prosecution proceedings, in case the patent to be filed is not novel or obvious, thereby saving plenty of money.
- Ensure a patent holder that they will not get indulged in patent litigation if a patent is granted.
- Enlighten the applicant regarding prior art related to the same subject matter and give useful insights into their competitor’s products and corporate strategies so that they can file a set of claims in the patent application complying with the patentability requirements.
- Enable the inventor to take an informed and well calculated decision regarding the feasibility of filing a patent application.
Some of the traditional Prior Art Search strategies are:[x]
- Keyword Search
- Citation Search
- Name Search
- Classification Search
- Boolean Search
- Positional Search
- Search by Dates (e.g., priority date, application date, publication date, grant date)
- Search by Patent reference or identification numbers (application number, publication number, patent number)
- Search by Names of applicants/assignees or inventors
Though there are numerous strategies to search prior arts, searching should be done using techniques which produce optimum results. It has to be verified what strategies will help the IP professional to strike a balance between;[xi]
- The recall of the search: total number of hits that are relevant.
- The precision of the search: the proportion of hits which are relevant.
Drawbacks of conventional prior art search techniques
In this fast paced world, where millions of patent documents are already published world-wide, accessing such a huge database manually for prior art patent search is a herculean task as well as a time-consuming one. Further, after searching, analyzing and scrutinizing the vast number of patent publications, extracting the relevant piece of work out of the voluminous patent and non-patent literature, available in multiple formats, languages and sources, is an extremely laborious and tedious task.
Following issues are usually involved when traditional prior art search techniques are employed:
- Data processing errors
- Errors due to language pitfalls
- Errors due to faulty syntax
- Classification error
- Affects accuracy and quality
- Affects efficiency
- Large number of false positives and false negatives
- Time-consuming process
- Exorbitant costs involved
Automating the process of prior art search by employing AI
To overcome some of the drawbacks of manual prior art patent search and to enhance the quality of search results, the entire process can be automated incorporating ML and AI technologies. AI is highly proficient at processing voluminous sets of data with great precision and accuracy, saving plenty of time. With the help of AL, one can penetrate into otherwise impenetrable and inaccessible volumes of data and databases. Patent Research done manually usually takes days and months whereas the same task can be accomplished within a couple of hours, that too with high accuracy and efficiency, by deploying AI in an effective way. This is possible with the help of Machine Learning that enables semantic based prior art search which takes a researcher’s intent into consideration and exposes relevant documents instead of simply comparing the keyword entered with the patent pre-existing patent documents, thus eliminating the risk of false positives and false negatives. A patent applicant can himself search prior art and obtain a search report without consulting an expert in IP, if they avail the benefit of AI based semantic search technology which facilitates one to enter even a paragraph in the search query for the reason that the neural network algorithms has the capacity to accept large block of texts. The AI engine automatically identifies the most relevant documents by deriving the context and associating meanings according to the matter in the search query. Semantic engines work by focusing on the idea rather than the expression; hence it delivers the most apt output corresponding to the input.[xii]
Technical process of deploying AI techniques for prior art patent search
ML and AI techniques automate the process of patent related discovery by replacing conventional searching methods like Boolean search, positional search, keyword search etc. with AI enhanced semantic search, thereby, retrieving documents consisting of similar concepts or logics by the usage of ordinary and accustomed language. The quality and accuracy of search results gets increased in this process as the complete text of the patent application is compared with all the relevant existing publications in the database. Therefore much more appropriate information is extracted as it considers and includes the context of keywords used within the respective documents. This addresses the drawbacks associated with the keyword based patent search that often delivers inaccurate and inappropriate results for the reason that it does not take synonyms or abstract terms related to the given keyword or query word into account. The search engines such as ESPACENET, PATENTSCOPE, TOTAL PATENT, PATSNAP patent search, Google Patent Search, USPTO Web Patent databases etc. fail to aptly deliver relevant documents owing to the fact that only the specific keyword is taken into account and not the entire text of the patent application and further it is impossible for humans to go through the entire text of all the relevant patent documents.[xiii]
A Deep Belief Neural Network, which is essentially a semantic model comprising concepts and topics in the form of mathematical vectors, extracts logics and meanings from patents. Neural Network which is a subset of Machine Learning enables a computer to scrutinize large volume of data and further derive meaning from the data examined.[xiv]
Process of assigning vector to each document in order to exhibit outputs based on concepts:
- A dataset of related patents is obtained from the patent database.
- Patent publications are categorized using a system which has emerged from Neural Network analysis of the indexed literature.
- Great stress and emphasis is laid on the various categories and concepts in the entire piece of data.
- Patent texts and documents are assigned with numerical vectors according to its corresponding category.
- Based on these vectors, full text similarities between two patent documents are computed, making the process automatic and reliable.
In a nutshell, AI automates the process of prior art patent search by comparing the full text of the patent application with the pre-existing patents in the database. Subsequently, a similarity score is obtained based on which one can categorize and rank patents according to its resemblance with the patent application. The patents exceeding a certain threshold can be suggested as a prior art.[xv]
Since patent search deploying AI techniques is based on semantic search, the artificial neural network algorithms encode concepts and logics embodied in a document into highly comparable semantic vectors, regardless of the terminology used for patent searching. This helps in overcoming the barrier of language pitfalls involved in manual keyword search as AI displays research beyond the terminology used, based on the codes associated with the concepts attached to the inserted keyword.[xvi] The most relevant and domain-specific documents can be obtained using Semantic based algorithms as they lay utmost emphasis on digging out all the closely connected documents by identifying and understanding the words that an user has typed and subsequently, linking it with the concepts attached to the keywords. To enhance the accuracy and to obtain results with less noise or junk and fewer false positives, it is advisable to incorporate a language model and a literature knowledge graph.[xvii]
A slight shift in the role of IP professionals
IP professionals and experts no longer have to learn all about complicated search terms, keywords, and Boolean operators in order to get results related to patent history of a subject matter. They simply have to feed the details of the invention into the AI engine to get an insight regarding the prerequisites of patentability. With the help of AI engine, they obtain results regarding:[xviii]
- Novelty – AI provides proof of overlap with the prior art and novel elements.
- Visualization- Competitor protection in the technology of interest.
The entire process of patent search is accomplished within a very short time span employing the AI and ML technologies, which is not possible by traditional patent research methods. Earlier, clients had to solely rely on IP professionals and experts to provide them with a Search Report, as they themselves had no idea about accessing patent data. AI has solved the issue making it possible for anyone to have access to the prior art patent data without any hustle.
Advantages of automating the process of prior art patent search
Automated prior art search techniques using Machine Learning and Neural Network language model addresses the drawbacks associated with the manual searching methods and grants the applicant with the following benefits:
- Increases quality and precision of the search results.
- Takes researcher’s intent into consideration.
- Gives an insight about competitor’s strategy.
- Reduces the amount of false positives and false negatives.
- Increases efficiency and accuracy.
- Reduces manual work.
- Saves plenty of time required for conducting the search.
- Minimizes costs.
Conclusion
Prior Art Patent Search is a sine qua non step and should be conducted even before filing a patent application with the aid of a patent attorney in order to meet the prerequisites of patentability. While conducting patent search, patent attorneys and experts mostly rely on keyword based search engines which often deliver erroneous results consisting of a large number of false positives and false negatives. While false positives further add on to the burden of the patent examiners as they have to exclude the irrelevant documents, false negatives may lead to the fallacious granting of patent. To eliminate the hurdles associated with manual prior art search techniques employing AI techniques for prior art patent search is extremely necessary. With the help of AI based semantic search engines, the entire text of the patent application can be compared with the pre-existing patent documents in the database. The semantic based model lays more emphasis on the idea and key concepts rather than the expression used in the search query. As a consequence, more precise and apt outputs are extracted, avoiding irrelevant patents. Artificial Intelligence and Machine Learning techniques prove to be very advantageous for conducting patent search as it decreases the manual work load, increases accuracy and efficiency of the search results and saves plenty of time by taking researcher’s intent into consideration.
References
[i] Artificial Intelligent Techniques, available at, https://www.educba.com/artificial-intelligence-techniques/
[ii] Machine Learning, available at, http://www.contrib.andrew.cmu.edu/~mndarwis/ML.html
[iii] What are the three types of AI, by Serena Reece, available at, https://codebots.com/artificial-intelligence/the-3-types-of-ai-is-the-third-even-possible
[iv] Artificial Intelligence algorithm, available at, https://www.edureka.co/blog/artificial-intelligence-algorithms/
[v] Artificial Intelligence algorithm, available at, https://www.edureka.co/blog/artificial-intelligence-algorithms/
[vi] Everything you need to know about Neural Network and Backpropagation, available at, https://towardsdatascience.com/everything-you-need-to-know-about-neural-networks-and-backpropagation-machine-learning-made-easy-e5285bc2be3a
[vii] Developing Artificial Neural Network and Multiple Linear Regression Models to Predict the Ultimate Load Carrying Capacity of Reactive Powder Concrete Columns, available at, https://www.researchgate.net/publication/317671554_Developing_Artificial_Neural_Network_and_Multiple_Linear_Regression_Models_to_Predict_the_Ultimate_Load_Carrying_Capacity_of_Reactive_Powder_Concrete_Columns?enrichId=rgreq-8d44512bbe23b5c9b8216c6559225553-XXX&enrichSource=Y292ZXJQYWdlOzMxNzY3MTU1NDtBUzo1MDcwNzk2NTU5NDgyODhAMTQ5NzkwODYxOTk5Mw%3D%3D&el=1_x_3&_esc=publicationCoverPdf
[viii] Machine Learning Yearning, available at, https://www.deeplearning.ai/content/uploads/2018/09/Ng-MLY01-12.pdf
[ix] Deep Learning for anomaly detection: A survey, available at, https://arxiv.org/pdf/1901.03407.pdf
[x] WIPO guide to use Patent Information, available at, https://www.wipo.int/edocs/pubdocs/en/wipo_pub_l434_3.pdf
[xi] e learning on Patent information, WIPO
[xii] How Semantic Searching unlocks AI, https://ip.com/blog/semantic-searching-unlocks-ai/
[xiii] Automating the search for a patent’s prior art with a full text similarity search, Lea Helmers, Franziska Horn, Franziska Biegler, Tim Oppermann, Klaus-Robert Müller, available at, https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0212103
[xiv] AI meets IP: Artificial Intelligence solutions, available at, https://ip.com/blog/ai-meets-ip-artificial-intelligence-solutions/
[xv] Automating the search for a patent’s prior art with a full text similarity search, Lea Helmers, Franziska Horn, Franziska Biegler, Tim Oppermann, Klaus-Robert Müller, available at, https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0212103
[xvi] AI meets IP: Artificial Intelligence solutions, available at, https://ip.com/blog/ai-meets-ip-artificial-intelligence-solutions/
[xvii] Why AI is crucial for patent searching and mining, available at, https://www.iam-media.com/why-ai-crucial-patent-searching-and-mining
[xviii] 8 ways in which Artificial Intelligence is being used by IP departments of Companies, available at, https://ttconsultants.com/blog/8-ways-in-whh-artificial-intelligence-is-being-used-by-ip-departments-of-companies/
LawSikho has created a telegram group for exchanging legal knowledge, referrals and various opportunities. You can click on this link and join:
