Artificial Intelligence and the Asian Legal Market
By Bill Novomisle, In-Gear Legalytics
New York LegalTech 2018 closed on Feb 1, 2018 and without a doubt, the highlight and dominant theme of the conference was artificial intelligence in the law. Although there has been extensive coverage of AI in legal application in North America and Europe, significantly less attention has been given to the unique challenges and opportunities for this technology in Asia. Within the Asian market there is a lot of confusion regarding what AI actually can do, how it works, and what opportunities may exist for adopting AI within legal departments and law firms in this region.
Perhaps the greatest source of confusion comes from asking the wrong question. Too often, users ask “how can I use AI in my legal practice/department?”. A potential purchaser of legal technology should not start with the technology they are hoping to use and then try to back-in to a use case for the technology. Rather, they must understand the problem they are hoping to solve first and foremost. Once a use case has been identified, including a clear vision of what “success” looks like and the role that technology can play in the path to success, then technology can be evaluated to see if it is the correct fix for the problem at hand. Nevertheless, the topic of artificial intelligence comes up regularly as a starting point (rather than end-point) for a discussion about how the practice of corporate law is changing. Despite the tremendous amount of talk surrounding AI, many lawyers do not actually understand how this technology works nor are they aware of specific pitfalls that apply across numerous Asian jurisdictions to the adoption of AI-powered tools.
This article attempts to address this challenge. In section 1, the primary forms of legal technology are described, including the basics of how they work, in non-technical language. Section I also will explore some of the unique challenges that legal technology faces in the Asian market from a functional (e.g. language) standpoint. Finally, in Section 2, an overview of near-term opportunities for AI powered legal technology are explored. For savvy purchasers of legal technology, now is an excellent time to consider leveraging this powerful technology to drive both efficiency and efficacy in their firms and legal departments.
Purchasing legal technology is a significant undertaking, both in terms of time and money. It is my hope that readers will gain a better appreciation of how the tools they are considering adopting actually work as well as sophistication as consumers of legal technology. Input from legal leadership, practitioners at all levels, and IT resources, are all necessary to the successful selection and implementation of legal technology. Experience in this area is invaluable and irreplaceable. Resources are available that have experience in selecting and implementing these tools that also bring practical experience in delivering legal services to drive business strategies and legal operations. For this reason, at a minimum, a potential purchaser is advised to speak to others who have already implemented the technology under consideration and ensure they have adequate internal resources to fully manage the project’s implementation.
1. The Three Basic Forms of AI Currently Being Applied to Legal Work
One source of confusion when discussing AI comes from the different types of technologies that all fall under the broader umbrella of AI. As an introduction to some of the types of technology that can be described as AI, the three forms of AI that have been meaningfully applied in legal practice are explained below. In practice, these three forms are not mutually exclusive, and indeed most AI tools incorporate at least two of the below forms in practice. Nevertheless, by dividing the technology into these three fictitious categories, a basic understanding starts to emerge regarding how AI works in the law.
- Deep Learning
Perhaps the aspect of AI that causes the greatest amount of excitement and fear is the idea that this technology has the ability to “learn.” Rather than a static system that takes predictable inputs to produce consistent outputs, AI has the ability to get better over time. As the AI algorithm encounters more and more situations it has the ability to adapt itself to those previously unexplored inputs. Most lawyers however, do not understand what deep learning is or how it works.
The first concept to understand with deep learning is the idea of “layers” of learning. Layers refer to the complexity of the relationship between variables. To use a mathematical example, imagine that you are trying to teach a single-layer deep learning algorithm to recognize the relationship y=2x. You throw a bunch of numbers at the machine, and then “teach” the algorithm by telling it when the rule is satisfied (for example, when y=4 and x=2). You also would have to teach the machine when the rule is not satisfied (for example, when y=10 and x=1). Eventually, after being trained with enough examples, the machine would identify the relationship between x and y and would be able to both predict situations when the rule was satisfied (for example, if you told the machine that x=7.5 it could predict that y=15) and situations when the rule would not be satisfied (for example, it would predict that x=7.5 and y=14 would fail the rule).
It is fairly clear that a single-layer deep learning algorithm has no practical worth in legal practice. Consider a document or diligence exercise that was focused on a transaction called Project Prometheus. A basic search can simply scan through electronically stored information (ESI) and quickly and easily match the search term “Project Prometheus” to all documents containing those words. However, if a practitioner wanted to use a single-layer deep learning algorithm to perform that same search, she would have to manually train the algorithm by reviewing several dozen (or perhaps several hundred) documents and identify “positive” hits that contained the term “Project Prometheus” as well as negative hits that lacked the key phrase. Given that non-AI search technology can do that same task near-instantly and with no human intervention, single layer deep learning AI is obviously the wrong tool for the job. This is why single-layer deep learning does not exist as a practical matter.
As deep learning AI adds layers however, its ability to recognize and adapt to patterns that have more complex relationships increases and begins to approximate human decision making more closely. Consider the following rule: Positive documents only include Project Prometheus documents that address total deal value, financing arrangements, debt-for-equity swaps, or transferring key personnel, but only if those personnel were director level or higher. Also, positive documents also include Project Prometheus documents that address transfer of intellectual property including trademarks, trade secrets, and patents numbered 0123456789 and 9876543210. This hypothetical rule (that readers will hopefully identify as somewhat more realistic) now involves at least 14 characteristics (aka variables). And within these variables, some will likely include additional non-enumerated concepts (consider “financing arrangements”). Furthermore, other variables are independent of each other (total deal value or debt-for-equity swaps) and others still are dependent on each other (“transferring key personnel only if … director level or higher). An AI driven deep-learning tool with the appropriate number of layers (it is not entirely clear how many would be required here, but it is likely around 25-35), given the proper amount of training, would be able to massively and radically improve on the speed of a human reviewer. However, because of the investment in time that is needed to train the system, the efficiency savings are only realized when seeking out “positive” hits in a data set with several hundreds of thousand documents. There, a diligence review could be completed in a matter of hours rather than in a matter of weeks or months that a manual review would require.
Note that this result is dependent on two factors. The first, the number of layers that the deep-learning technology employs is a key limit of the capabilities of technology. The state of the art is now such that most highly sophisticated algorithms being employed by the leaders in legal technology have the ability to handle thousands of layers. Nevertheless, this technology’s limitation is important for a potential buyer of AI technology to keep in mind, particularly with the large number of start-up software offerings that are flooding the AI legal tech market. Buyers should not necessarily flock to the most complex deep-learning algorithms, but should carefully consider the complexity of the problems they hope AI will help them address (now and in the future) and make sure that the solution is appropriately sized for the problem. Often, determining whether the solution is “right-sized” will require expertise in both law and technology – which may not always be available within a single individual, or even an organization.
The second factor identified above is that the algorithm is given the appropriate amount of training. This critical step is easily overlooked by buyers of legal tech that are eager for an “out of the box” solution. If the AI solution you are considering is based upon deep learning, the tool you receive out of the box will only deliver results once a significant human investment has been made into training the tool. Understanding this point shows the first place that adoption of AI based legal technology tools across Asia deviates from markets in North America and Europe. The leading AI applications in the legal sector have now all been operating for several years. Kira has been around for seven years. RAVN is approaching its eighth birthday. Same for Seal Software and IBM’s Watson. Luminance began over five years ago. Recognizing their need for massive amounts of high quality data, each of them has necessarily sought out partnerships with law firms (or were borne from law firms themselves) in order to train their algorithms. Now, with a significant track record and amount of data that has passed through their machines, they are able to claim for their products “out of the box” ability to do certain high-volume tasks, such as analyze contracts in a data room for the due diligence portion of an M&A transaction. Today’s purchaser is able to take advantage of the years of training that the machine already has experienced.
However, the buyer of legal technology is cautioned here to make sure they check their tool of choice well before purchasing. The adoption and sale of AI tools in Asia has been fairly limited thus far. As experienced practitioners are well aware, local formatting and style of legal documents in Asian jurisdictions is often unique and particular to its jurisdiction – even if the language of the document is solely English. However, if the use case that the AI tool is designed to assist with is based on outbound investment, or other cross-border documents that are likely to have a heavy bias in favor of North American, UK, or European drafting conventions, or otherwise conforms to a universal standard (such as ISDA forms or some NDAs) it is quite possible that these tools will be able to work quite effectively without much additional training. If, however, the use case under consideration is deeply localized, some skepticism is warranted. Make sure to insist on a test case demonstration performed at your offices, using your actual data. Take the time after the demo to independently verify the results of the machine’s work. You may find the accuracy is closer to 80%, which may be good enough with the understanding that you will be undertaking training yourself to bring the machine up to higher levels of accuracy. However, you will understand both the work necessary to improve the tool and the best way to implement the tool by doing your appropriate diligence.
Finally, in the context of discussing how to train an AI system, two additional concepts are worth understanding: precision v. recall. Imagine that you are looking for something specific within a large set of documents. After you run your search, the overall dataset is divided into positive and negative hits. Precision is a measure of the positive hits that were identified correctly. For example, if the search returned 100 positive hits and after manual review it was concluded that 98 were correctly identified as positive, this search had a 98% level of precision. Recall, in contrast, is a measure of the positive hits that were correctly identified versus the positive hits that were missed by the search. Returning to the same example, if there were actually 300 positive hits in the overall dataset, but the search used only pulled out 98 true positives (out of the 100 returned), this was a search with merely 33% recall. In other words, there were 202 (67%) false negatives in this hypothetical, which would pose a severe efficacy problem for most users.
Remember that in order to train an algorithm, the machine needs to see examples of both positive and negative hits. There is a significant debate in the legal deep learning community as to the best way to train an algorithm, and whether the appropriate approach is to use a large static sample set, completely reviewed by a human, or if it is better to do several iterative smaller training sets, with the algorithm being allowed to run in between training sets, as a way to boost precision as quickly as possible. Rather than there being a single best approach, it is more important to understand the particular use case the technology is being applied to and also to understand the specific risks that are posed by such use case. In short, certain risks are exacerbated by poor precision, while others are worsened by poor recall. If AI technology is purchased and it is determined that training of some kind is needed in order to maximize the utility of the tool, it is imperative that the client has carefully thought through the types of risks it is concerned with so that they can work with their IT department, consultant and/or software engineer to devise the most appropriate training protocol for their system.
- Natural Language Processing
As its name implies, Natural Language Processing (NLP) is a form of AI that is concerned with teaching machines to “understand” and even converse in natural human language. NLP is the technology that underlies smartphone and smart-home assistants (such as Siri and Alexa) as well as digital chat bot assistants found on many websites. NLP itself utilizes deep learning. In order for an NLP system to function effectively, the technology must be able to tackle four interrelated and overlapping aspects of linguistics: syntax, semantics, discourse, and speech. While anyone who has used Siri or Alexa will recognize that these systems are remarkably effective, particularly in spoken English, to apply these tools to Asian languages requires material adaption of both syntax and semantics (Syntax is the breaking down of language, semantics is the aggregation of language into meaning).
Syntax breaks the language being analyzed down into smaller components with the goal of having the machine understand the individual units of the language. Perhaps the greatest challenge for NLP applications in Asia lies here. Linguists refer to morphology as the study of how words are formed and their relationships to other words in the same language. Morphology within English (all dialects) is fairly straightforward and while there are some ambiguities, the challenges are well known and have been effectively addressed. In comparison, Mandarin (to use the most commonly spoken language in East Asia), poses several challenges for NLP that are being worked on, and have been addressed with varying degrees of success. For example, Mandarin Chinese incorporates many out-of-vocabulary words such as proper names that pose a significant challenge for machines because unlike English and other Indo-European languages, there are no spaces between words/characters. For example, “Canada” is translated as加拿大 which are the characters for “to add”, “to hold” and “big” respectively. This is borrowed from Cantonese, where the sound of the word – Ganádà – is a close approximation of the English word. Mandarin has in turn adopted the same characters for the name of the country, resulting in a word that is pronounced “Jianádà”. This type of out-of-vocabulary word poses translation challenges for NLP systems.
Also, native Chinese “words” may be a single character, two characters, or three depending on the context. For example, 小笼包 (xia˘o lóng bao) is commonly translated as the adjective-noun phrase “steamed dumpling”, but is composed of the adjective “small” (小), the noun “bamboo steamer basket” (笼) and the verb “to wrap” (包). The morphology of this term would require the machine to recognize that these three characters were acting together to form a new noun, rather than acting on each other to describe a small bamboo basket that had been wrapped. Because morphology lies at the root of many of the functions of NLP processing, accuracy in Mandarin and Cantonese morphology remains a significant challenge for NLP developers. Similar morphological challenges are also posed by other Asian languages such as Korean and Japanese.
The current state of the technology is such that proper names, particularly those utilizing out-of-vocabulary words, remains a bit of a challenge that has not been entirely solved because the answer often lies in the cultural context between the speakers. However, the best technology in this area has developed well-functioning and elegant solutions to identify multi-character words and most of the other morphological/syntactical differences found in Asian languages. This has opened the door to some of the first NLP based AI tools targeted to multinational businesses operating in Asia (as discussed in Section II).
Semantics also poses a challenge in Asia. Semantics focuses on the relationship between the units of language and how they combine to form meaning. Semantics is a particular challenge to legal documents because the relationship between words is the entire key to the meaning of a law or regulation. Thus, a machine-error in the semantic construction of a legal document can be devastating. To continue with Mandarin as an example, consider this sentence from the PRC Securities Regulatory Commission amendments to the Securities Issuance and Underwriting Decision Law (2014):
This can be translated as “after offline and online investors receive the placement, they should pay the subscribed-to funds on time and in full.” However, a semantic mistake made by a machine could confuse dependency of the clauses and translate this law as requiring the funds to pay the investors once placement is completed. This type of error turns the intent of the law on its head! Uncritical or unsupervised reliance on a mistake prone system can lead to costly and embarrassing errors. Highly accurate semantic analysis of native language is absolutely essential before any reliance on AI driven technology can be justified. While there are scholarly articles regarding the use of language specifically in Chinese law there are still more opportunities for a commercial entity to provide a legal technology tool applying this work to an NLP based system for Mandarin. This problem is also true for the other primary Asian languages as well.
The last two aspects of NLP technology, discourse and speech, also have distinct challenges in Asian languages, however they pose challenges with comparable complexity to those found in other languages and are not particular warning flags for purchasers of legal technology in Asia. While NLP does include spoken language interaction with machines, for now, it is rare to see voice recognition technology currently being built into the tools on the market. The focus remains on written (typed) input and output. Significant strides have been made in Mandarin NLP and research continues at a rapid clip. While the majority of AI assisted tools for lawyers remain English only, it is only a matter of time before these challenges are adequately addressed and additional language capability is added to the region marketplace.
- Expert Systems
Expert systems are some of the longest established forms of AI technology. Indeed, expert systems have become so commonplace many people do not recognize (or acknowledge) that many well-established business software products are AI enhanced expert systems (e.g., SAP, Siebel, Oracle). An expert system emulates the decision-making skills of a human. The technology utilizes two components: a knowledge base and an inference engine.
The knowledge base simply represents what the machine is told are the facts regarding the world. These can be laws, fact patterns, or a combination of both. The inference engine is the automated reasoning system that takes the existing knowledge base and by applying its rules to the knowledge base, generate new knowledge. The system is highly adaptable, such that it can recognize when there is a change in the fact pattern and is able to immediately reapply the inference engine to determine any changes in output. The system can also handle hypothetical reasoning (e.g. multiple parallel possibilities) as well as assigning probabilities to outcomes rather than rigid certainty in order to evaluate more complex decisions (fuzzy logic). The most recent development in expert systems has been their application as Bayesian network (BN). Technical terminology aside, a BN expert system analyzes data patterns for causal or dependency relationships. This in turn can be used to create predictive models and decision support systems.
Expert systems are well established technologies, which have not undergone major advances since the 1990s – although incremental improvements to them continue on a regular basis. In addition, an expert system is not a type of interface, and therefore is agnostic to the input language that is used for its knowledge base and for the search query. For this reason, a buyer of an expert system has no regional-specific obstacles to overcome.
2. Opportunities for AI-Powered Legal Technology in Asia
Before focusing on the marketplace Asia, this discussion must shift from the types of technology employed to the type of solution the technology purports to offer. Most of these solutions have been covered in my earlier articles. These are:
- Legal research aids
- Litigation prediction/strategy support
- Preventative law (monitoring)
- Unstructured data analysis (due diligence, contract analysis, document review)
In the current state of the market, not all of these above use cases are likely immediate-term targets for AI-based tools (however, it is only a matter of time). Legal research tools targeted to private consumers has proven elusive for the most part. Within China, the China Supreme People’s Court started using an AI-enabled legal research tool called FaXin to search for precedent and identify analogous decisions to help guide judges. With a Mandarin-only interface, Legal Miner, is also offering AI-assisted legal research tools. Singapore saw the introduction of Intelllex to assist law students and junior lawyers with legal research tasks. Indian legal research tool CaseMine has achieved some success in offering both legal research and automation of basic tasks.
A tool engaged in litigation prediction or litigation support strategy is similar to a legal research tool, but applied to a specific set of disputed facts. Likewise, a litigation monitoring system, like NexLP or Intraspexion is meant to identify litigation threats by actively monitoring the data and communications of a corporate entity as it is being generated. To date, no such tool has been offered commercially in any Asian jurisdiction. However, due to problems discussed in earlier articles, as well as problems involved in interpreting multiple languages, dialects and idiosyncratic user variability, as well as a legal culture that is policy driven (rather than textual driven), these areas are not likely to see entrants in the near future and there is not a huge market demand for this type of tool across the region.
When it comes to unstructured data analysis, there are major advancements that have occurred in Asia and it is reasonable to expect that more will be forthcoming in 2018. Certain lawyers working for multinationals may find themselves equipped with additional AI-based tools due to a combination of head office decisions and mergers/JVs amongst legal tech providers. For example, users of the iManage document management system (DMS) will automatically have access to AI data extraction tools powered by the deep learning and NLP tool developed by RAVN, which iManage acquired last year. In a move that will likely further integrate the legal function within the greater organization, iManage has also introduced knowledge management, universal search, and data visualization tools (powered by legacy RAVN AI) to allow for information insights across departments. Not to be outdone, legacy users of NetDocuments will be able to take advantage of Kira AI tools to allow for analysis and extraction of matter, contract and transactional content. This will be particularly beneficial for M&A related work because of the automated entity and clause extraction tools provided by Kira.
For potential users that are not inheriting AI tools due to decisions made far away from their offices, there have been significant opportunities in the Asian market they may wish to consider taking on their own. Axiom has introduced its Contracts Intelligence Platform, which is targeted to all aspects of the M&A process (buy side, sell side, and post-merger integration) which offers both streamlined clause extraction to aid in diligence, as well as advanced data visualization to provide business intelligence regarding deal synergies and risks. It is well known that other companies have been examining similar technology (e.g. KorumLegal, Loom Analytics, Surukam) and more announcements may be forthcoming in 2018.
That said, there has been one potentially revolutionary announcement made in this space. Deloitte Legal has now introduced an NLP-driven tool to assist corporations with compliance across a myriad of Chinese regulations. This is particularly significant because Deloitte Legal has a long history of working with international leaders in legal AI (e.g. Seal and Kira) and has developed this new Mandarin based NLP itself. This has the potential to be the leading edge of a new wave of NLP products that expand well beyond the English language tools that currently dominate the market. This is not a traditional legal research tool in the sense that a law firm lawyer would use, but it does allow a corporate client a technology assisted solution for compliance with a myriad of local, regional and state-level regulations that are prone to change rapidly and with little notice.
Beyond that, the usual list of global players (IBM, Microsoft, Kira, Luminance, Elevate, Seal, LawGeex, ContractPod, Ayfie, among others) continue to be open to opportunities for sales and development across the region. While most of these players offer out-of-the-box solutions that can be applied to large data sets, they have not specifically targeted the Asian region (in the sense of setting up a permanent sales/support force) to date. These players can also help create bespoke technology solutions, ranging from chat-bots to business intelligence dashboards for corporate clients. Additionally, law firms would be wise to look at practice areas that are under pricing pressure from their clients to seek out opportunities to leverage an investment in technology for improved efficiency in their delivery of work product. Once again, this is why it is critical to understand the use case you wish to target as a purchaser before contacting vendors. This way, criteria for the tool can be established in advance of evaluation and careful consideration of the challenges set out in Section I of this article can be tested during the diligence phase of vendor selection.
There is one form of AI that does not seem to get much attention in Asia, but can be used to address the need for efficiency in many routine, high-volume requests and tasks. This is the use of custom designed expert systems. Consider one well-established expert system provider, Neota Logic. The underlying expert-based system was always set up so that an individual could program the system without having any knowledge or ability to program computers. This means that complex, but routine questions (such as those involving labor law, advertising and marketing restrictions, non-disclosure agreements, routine renewals for licensing or intellectual property, periodic financial closings, etc.) can be programmed by a lawyer (or paralegal in some instances) to create an expert system that can then be utilized by non-lawyers. If a question becomes too complicated, a human lawyer can be injected into the workflow, and if not, questions or compliance can be handled without the intervention (and delay) of the legal department. But with the latest release, Neota now has the ability to integrate with the legacy technology that underlies most common existing legal applications in finance, matter management and workflow management. This has taken what was once an expert system tool and has added workflow management as well as dashboard analytics to its capabilities. For a corporate client looking to create a self-service tool for high-volume, routine legal questions, this has tremendous potential for efficiency gains while still maintaining visibility and control over the process. For a law firm looking to expand its offerings, this software tool offers the chance to program its knowledge into a system that can then be sold to an unlimited number of clients simultaneously without any drain on lawyer’s limited time.
The key for any potential purchaser of legal technology is to start with the problem – not the solution. For organizations that have been specifically mandated to start incorporating AI into the legal work they perform, there are current opportunities and there will likely be even more in the near future. None of the pitfalls described in this article are insurmountable, nor should a potential purchaser be deterred from taking the leap into the future of law. They should merely do so with eyes wide open, and with the most sophisticated understanding of their undertaking as possible. There is no time like the present to become part of the future.
 Despite this article’s focus on AI, there is much more happening in the Asian legal technology sector as a whole. From practice management solutions, to document management, to e-billing, to matter management, to blockchain, to e-signatures, and more, legal work is increasingly benefiting from automation and targeted software solutions. For the sake of brevity and focus, those topics are beyond the scope of this article.
 The consultancy co-founded by the author, In-Gear Legalytics Limited (IGL), is one such resource. Neither the author nor IGL is affiliated with any company mentioned herein nor do they offer any proprietary software solution to the public. The discussion of current technology offerings is made without endorsement and no prejudice is implied against any technology providers that have been omitted. Every provider offers different strengths and weaknesses. Because IGL is software agnostic, its services are based solely on client needs (operational and technical) and delivering the optimal solution for those needs.
 Not the proper technical term. The truly ambitious reader with a pre-existing technical and programming background is invited to take the udacity.com free, three-month, deep learning programming course taught by Google. More information available here.
 Training time at this point requires human labor for legal AI applications. Self-teaching AI systems that do not require humans, such as AlphaGo Zero, cannot yet handle the level of complexity legal tasks entail.
 At a macro level, this type of technology testing is analogous to hiring lawyers or law firms on the basis of their industry or transactional experience.
 For an introduction into how to start to address this training issue, see here. In general, systems that have poor precision but good recall require significant human capital (and sometimes financial capital) to improve. Systems with poor recall but good precision present risk tolerance issues for the client to grapple with.
 Modern NLP may more accurately be described as Statistical Natural Language Processing to distinguish itself from earliest forms of NLP which were decision tree based. However, these NLP systems are a combination of both self-teaching and manually taught. To avoid excessive complexity, the technical details of how this works is omitted. However, smartphone users may have noticed that predictive text and virtual assistants improve with use over time by a specific user. This is an example of how statistical NLP combines self-teaching with human input.
 Morphology is an aspect of the broader field of syntax.
 For the original text of the law, see, 证监会关于修改《证券发行与承销管理办法》的决定.
 See, e.g. Deborah CAO, Chinese Law: A Language Perspective (2004); 中西法律传统, v1 2001.
 Discourse refers to the relationship between phrases to give specific meaning. For example, an aspect of discourse analysis would focus on how to teach a machine that the phrase “side entrance” in the sentence “You can enter my house via the side entrance” refers to a structure that is part of the speaker’s home.
 Speech-to-text translation, which is significantly more challenging than text-to-speech translation.
 However, this may be beginning to change. Case.one (a US-focused practice management tool) has announced a partnership with Amazon’s Alexa voice system that allows US-based users to use voice command to call up relevant information.
 See Introduction to Chinese Natural Language Processing, Kam-Fai WONG, Wenjie LI, Ruifeng XU and Zheng-sheng ZHANG (2009)
 This is SQL technology. A SQL database is one of the most common forms of databases, and it is the technology that underlies most common existing legal applications in finance, matter management, and workflow management.