top of page

Claira’s Natural Language Understanding (NLU) vs. General Purpose Natural Language Processing (NLP)

Overview

A financial service firm can be represented as a summation of its legal agreements. Fully understanding a firm’s contractual obligations in its transactional exchanges of equity, debt and cash - including a firm’s roles, responsibilities and regulatory obligations - is critical to the success of any business. Most organizations capture less than 25% of the information contained in their contractual agreements - generally extracting only the “happy path”, i.e. when things go according to plan. Unfortunately, over 75% of their contractual information remains trapped in their legal documents but represents the “unhappy path”, i.e. when things don’t go according to plan. Given the increased pace of business and regulatory change, firms have repeatedly found themselves traveling this “unhappy path” without having a complete understanding of their contractual agreements. Hence, they are required to either engage external consulting resources, external counsel or rely on significant number of internal resources to manually sort through the contractual terms and conditions to find the necessary insight to safely manage through the “unhappy path.” Data trapped in financial agreements have significant business value if it can be extracted, structured and ultimately understood in a human-readable way. Doing so would enable it to be operationalized for greater process efficiency, improved risk management and to assist in easing the burden of regulatory changes such as Initial Margin (IM), Qualified Financial Contracts (QFC) and IBOR Transition. By structuring and understanding this unstructured data, organizations would have the information and insight necessary for more informed decision making, risk management and faster time to action. Technology enabling simple extraction and structuring of documents can be found in many general-purpose, first-generation Natural Language Processing (NLP) solutions available on the market today. However, achieving the next level of human understanding requires capabilities only found in the rapidly maturing field of Natural Language Understanding (NLU) solutions. Claira is a next-generation NLU solution that extracts, structures and understands data trapped in your financial agreements. Claira provides financial service firms with greater insights across a virtually limitless set of scenarios, such as:

● IBOR / LIBOR Transition ● Zero Floors & Negative Rates ● Contractual Risks & Scenarios ● Reps and Warranties ● Prime Brokerage Client Analysis ● Roles & Responsibilities ● Initial Margin (IM) ● Alpha Generation ● Deep Data Extraction ● ISDA/CSA Scenarios ● Credit Downgrades/Defaults ● Other Legal/Reg Risk Issues

Industry Challenge When general-purpose, first-generation Natural Language Processing (NLP) solutions on the market today are applied to financial contracts (e.g. a derivatives contract), at most we can identify the economic terms and other information such as Party A, Party B. The diagram is a simplification identified by ISDA [1]

This doesn’t work where the answer is simply ‘it depends’. Over 75% of a financial agreement is based on conditions. The conditional logic is critical. What it does not contemplate is the other aspects of the trade - we call that the “unhappy path.” The facts do change based on the circumstances. ISDA identified potential events that are required to understand.

In order to truly understand how a given provision works, and thereby a whole given document, it is necessary to understand the actual language contained within. However, legal language is well-known to have a distinct use and style, known as “legalese”, which is characterized largely by its obscurity to readers who are not domain experts. It therefore presents a particular challenge for traditional, general purpose NLP solutions because of how noticeably it diverges from typical English. Consider the following provision, an entirely unexceptional definition of an “Additional Termination Event” clause found in an ISDA Master Agreement. “It shall constitute an Additional Termination Event, with respect to Party A, if at any time and for any reason, it ceases to own beneficially and of record (directly or indirectly) at least fifty-one percent (51%) (determined on a fully diluted basis) of the capital stock of each class of Party B.” While it takes 52 words in legalese, when rendered into standard English, it takes only 23: “if Party A does not own 51% of the total stock of Party B, then an Additional Termination Event occurs concerning Party A”. The difficulty in overcoming the obscurity of legalese has motivated a number of attempts to solve the problem:

One-off Manual Approach - Teams of lawyers and/or consultants manually searching, tagging and remediating. One-off manual remediation is expensive & doesn’t scale.

Semi-Automated Approach - Electronic document management and workflow assisting in manual searching, tagging and remediating.

First-Generation, General Purpose AI - Statistical-based NLP or pattern matching requires the user to invest a significant amount of time, effort and expense to help train the “AI” NLP model using a corpus of 1,000’s of documents per doc type. The process is referred to as “style-sheeting” where the user tells the machine not only where to look but whether it got the correct answer or not. If not, then you repeat the process until you get an accurate result. Each new doc type requires retraining… a laborious process for relatively mediocre accuracy and a less-than intelligent outcome. Some models apply this approach to language but can detect only reliable patterns of relations between the elements of a sentence. The broad term for these approaches is shallow semantic parsing, which comprises such methods as question answering, named-entity recognition and role labeling. Even where shallow semantic parsing succeeds, the most it can generally offer are answers to specific, preferably short factual questions. The kinds of queries that provide genuine value to businesses are of a much more nuanced kind, which require a much more nuanced approach.

Claira's Next-Gen Solution

Claira applies next-generation, specialized AI to understand the language of financial services. Claira’s deep language understanding is tailored to financial legalese, coming equipped with abilities specifically prepared for the ways in which financial legalese is written. The comprehensive and conservative approach in creating a high quality parsed output along with full transparency of results enables businesses to utilize this data in assessing risk scenarios, automating downstream processes with conditional logic and reducing time to negotiating agreements based on historical benchmarks. Some sentences in legalese are too complicated to parse and may indeed have been drafted incorrectly with incomplete sentences, ambiguity or circular references. No solution, human or machine, is capable of properly understanding such incorrect drafting language. In order to handle such cases, Claira is able to provide full transparency to its findings - including a complete picture of the difficulty of the task and its expectation of success, enabling full visibility into the process of understanding. Claira’s comprehensive and conservative approach in creating a high quality parsed output along with intuitive workflow, escalation, collaboration and data visualization enables businesses to utilize this data to gain significant insights for numerous business-driven scenarios, including:

● IBOR / LIBOR Transition ● Zero Floors & Negative Rates ● Contractual Risks & Scenarios ● Reps and Warranties ● Prime Brokerage Client Analysis ● Roles & Responsibilities ● Initial Margin (IM) ● Alpha Generation ● Deep Data Extraction ● ISDA/CSA Scenarios ● Credit Downgrades/Defaults ● Other Legal/Reg Risk Issues


Claira's Benefits


More detailed data, stored automatically: Claira quickly parses the whole document by default (even the sentences beyond its complexity range). Any downstream queries are performed on data already analyzed and stored.

Low sample size for training: Claira does not depend on training data on a use-case basis, since it relies on analyzing input at the linguistic level, rather than on abstract pattern matching or supervised machine-learning techniques.

High confidence in the result: The combination of Claira’s monitoring model and complexity score is extremely accurate at determining when the parse has been successful and when it has not, so the data can be trusted. User thresholds can be set to accommodate to whatever level of responsibility to give to Claira versus human solutions.

Flexibility across many different documents: Claira is designed to understand financial legalese, which is widely used across many different types of agreements (ISDAs, CSAs, ACAs, MRAs, FRNs, et al.). Even considerable variations on the style of such legalese will still be captured by Claira.

[1] ISDA Legal Guidelines for Smart Derivatives Contracts; https://www.isda.org/a/MhgME/Legal-Guidelines-for-Smart-Derivatives-Contracts-Introduction.pdf

bottom of page