CHAPTER 10

Trend 6—Natural Language Processing

Introduction

What Is Natural Language Processing?

Natural language processing (NLP) is a subset of artificial intelligence that is focused on programming computers so they can (a) understand the human language from a variety of inputs such as e-mails, instant messaging, written communications, and verbal which can then be processed by other systems as well as (b) producing natural language outputs again across various sources such as e-mails, written communication and verbal.

The benefits of NLP are widespread, namely:

NLP should not be confused with Neuro-Linguistic programming. They both use the abbreviation NLP.

The Three Main Types of NLP

There are various streams to NLP, but for this book they have been grouped into three main areas:

Speech Recognition

This involves converting human speech into some sort of computerized coding so business rules can be run against it.

For example, a customer telephones a call center, and they are met with an NLP agent. This agent will “speak” to the customer regarding what they want to do (such as checking a balance or making a cash transfer) and then interpret the request and then process the request as necessary. A second example is where an NLP program will review telephone recordings against a set of phrases (called lexicons) to determine whether any fraudulent activity may have taken place. If a match is found then it does not mean that there has been an offense but it will allow a warning to be raised so it can be investigated by the relevant human team.

There is a further development in this area which at the time of writing (winter 2021) is still fairly new and has not been rolled out by many firms. This relates to NLP agents being able to recognize the caller by speech recognition. Therefore, when a customer calls a contact center then the NLP software will authenticate the customer by their voice as opposed to using passwords, PINs, mother’s maiden names, and so on. The obvious advantage is the whole process is quicker, easier, and more customer-friendly but there are some disadvantages. The technology is fairly new and it can struggle with people with similar voices and people who have an illness that impacts their voice. Also, there are unclear legal implications of using voice recognition to identify a caller. Finally, people may feel uncomfortable that they have identified themselves by entering a password or PIN.

Natural Language Understanding

This is developing software that will allow the system to read text or, in effect, allow computers to “read” albeit in some limited manner. The NLP program will receive in some written text (say from an e-mail, chat message, or document) which it will then review against a set of phrases (called lexicon) to try and understand what is being communicated.

For example, the millions of messages within a chat channel could be “read” instantly to determine whether any criminal activity is taking place. The NLP program will look for lexicons along the lines of “fix the price,” “make sure legal do not know,” and so on. As with Speech recognition above, irrelevant items could be identified because the system will review text literally. Therefore, a human team must review the outputs before any action is progressed.

Natural Language Generation

This is in the way the complete opposite of the above two. The NLP software will take the output of some type of process and produce natural language output. This could be the form of verbal communication (such as part of a call center), e-mail, instant message, written communication, or anything else. The outputs could consist of standard phrases (such as “Thank you [CALLER] for contacting us”) or more complex constructed sentences that are dependent on circumstances.

How Does NLP Work?

Speech Recognition and Natural Language Understanding Work in a Similar Manner

Both these are very similar in concept. They both receive in some sort of human language message (either verbally or written) which needs to be translated so the computer system can “understand” it and process the request accordingly.

The process can be summarized with the three sequential stages below:

The first stage is to try and understand the message. This will involve receiving the message and performing the following activity:

Once the message has been understood and cleaned up then it is possible to try and determine the actual request is. This is unfortunately hard to do due to the wording of the request, understanding the context of the message, errors in the request, and so on. However, the message will be compared against a set of lexicons to try and determine what the request is. If the request cannot be determined positively then there needs some exception process to review them by a human.

However, if the request is understood then it can be actioned. For example, process the bank cash transfer or provide an investment valuation.

Natural Language Generation

This area works in the opposite to the above where the computer generates a natural language which is then communicated to a human.

The process can be summarized as follows:

Initially, the system will produce some sort of output from a normal process. This will typically include some sort of confirmation statement (e.g., confirmation that a bank transfer has been completed successfully) plus some communication-specific data (account details, amounts, etc.)

This output is then converted into a string of words that will form the communication. This will consist of a set of sentences but checks will need to be made to ensure they link together correctly, have a natural flow, they are grammatically correct, and all technical terms are used correctly. Also, it may be possible to translate the message into different languages.

Finally, the communication is issued. This could be via the computer “speaking” over a phone, in a letter, on an e-mail, on an instant message, and so on.

Uses With Financial Services

There are a variety of uses for NLP with Financial Services.

Improved Customer Servicing

One of the most common uses of NLP is to improve customer service across contact centers, websites, and deal with customers’ written correspondence (such as e-mails or letters).

Contract center telephone systems have been enhanced to allow customer telephone requests, e-mails, instant messages, written requests, and other requests to be processed by NLP. The system will “understand” the request and use a set of rules to either process the request or direct it to the correct area for processing.

In addition, the website has been enhanced with NLP functionality. This has allowed websites to support intelligent form filling or even auto-completion. It has also allowed better website searching because NLP will allow a better understanding of what the client is looking for and present the links they need. Finally, it has allowed virtual assistants (or Chabot’s) to be installed which will allow customers to interact directly with the website which will “listen” and “respond” using NLP supported by a set of business rules and knowledge base to answer any questions or queries.

The end-result is that these improvements provide customer benefits because all requests are processed quickly and effectively. They also provide benefits to the firm because they do not need to employ large and expensive teams to process these.

Internal Efficiencies

NLP also allows internal efficiencies and cost savings.

All firms have clients and products across different jurisdictions and countries which means firms need to work with multiple spoken languages. Therefore, as NLP allows input and output communication to be processed automatically then it can also cope with different languages. This will save the need and cost of employing a team of translators.

In a similar vein, all firms have a vast amount of documents of which some could be many decades old. For example, firms will have mountains of client and supplier contracts often in paper or scanned format. Some of these contracts will have out-of-date clauses (e.g., around liability or indemnities) which could put the firm at risk. Therefore, NLP can be used to “read” these contracts and assess the suitability of all the clauses. Anything of concern can be assessed.

Fraud Management

NLP is being used to assess communications to determine whether any criminal and/or fraudulent behavior is taking place.

Inbound and outbound communications (e-mail, telephone recordings, trading messages, etc.) will be “listened to” by an NLP system against a list of known criminal and/or fraudulent phrases (or lexicons). Once the phrases are understood then they will be processed by a rules engine to determine whether they are fraudulent, criminal, or not. Anything suspect will be reported to the relevant teams. (These rules engines will typically be developed using Machine Learning techniques (see Chapter 7).)

A second example is where an NLP program will review telephone recordings against a set of phrases (called lexicons) to determine whether any fraudulent activity may have taken place. If a match is found then it does not mean that there has been an offense but it will allow a warning to be raised so it can be investigated by the relevant human team.

Product, Marketing, and Public Relations

Most firms receive a large amount of content. This cover market reports, newspaper, published articles, videos, interviews, trade journals plus many more. NLP allows this content to the “read” quickly to determine whether any new trends are emerging or whether the firm is being mentioned (especially if the firm is being mentioned poorly). This will allow the firm to take appropriate action.

Better Data Collection

One of the good side effects of NLP is that when an inbound communication is “read” and “understood,” it is converted in a standardized format. This then provides a great source of data that can be used (by Machine Learning technologies (see Chapter 7)) to understand customer behavior, market trends, and so on. This will allow the firm to implement customer-servicing improvements, implement efficiency and cost-saving changes as well as many other possibilities.

Challenges of NLP

Ensure There Is a Clear Business Reason to Implement NLP

There are many stories of firms (not just in the Financial Services) implementing new technology for the sake of the technology as opposed to implementing technology to meet some type of business or strategic need. (In the same way as any major change) this means before embarking on an NLP implementation then a firm must have a clear business reason to implement the technology. Therefore, please refer to Appendix B which provides a checklist on the activities required when creating a Business Case.

Implement or Roll Out NLP in a Slow Phased Approach

One of the issues with NLP is that any issues can be very visible to customers and the outside world. For example, if a Chabot is not working or giving strange results then clients will see this immediately. This means that NLP must be implemented correctly.

Therefore, it would be sensible to structure the implementation plan around a phased approach. The initial phase would contain the implementation of the actual core platform but then there should be phases each containing a prioritized set of NLP. The prioritization should focus on the rules that are (a) easier to define and code and (b) the processes that offer the most business benefit. This approach will hopefully ensure the project delivers some immediate benefits which will create goodwill, momentum, and confidence. It will also help with financial payback and allow the firm to learn about NLP which could provide useful lessons for the future.

It is also important to understand what is needed to support the implementation. This could involve staffing such as senior management, operational staff, in-house technology staff, NLP platform vendor staff, technology staff who support the systems being integrated with, legal staff, compliance staff plus others. In addition, if the rollout of NLP involves staff losing their jobs (because their roles are being replaced by the new NLP technology) then these staff may need to be involved in the project. If so then this will require careful and tactful management.

Finally, it may also be necessary to ensure there is desk space for the team, test environments, shared document folders, e-mail lists, and so on.

Build a Robust Infrastructure to Support NLP

A robust infrastructure must be built to support NLP. This will cover (a) defining the end state, (b) implementing the new NLP technology and ensuring it integrates with the existing back end systems, (c) ensuring there are suitable controls and governance in place, and (d) the firms have sufficiently skilled people to support the infrastructure.

Defining the End State

NLP cannot be used for all inbound communications. Because essentially NLP uses a set of rules to determine what is being “said” then it will best with requests that are simple and easily codified. For example, asking for a balance, asking for a product document, requesting a transfer, and so on. NLP will not work well with complex requests such as a process the death of a customer, closing an account, or a very complicated complaint.

Although NLP can be generally used for most outbound communication on the assumption that the NLP systems can generate the grammatically correct outputs in the required languages.

Therefore, the firms must think clearly about the processes or functions that they would like to include within NLP. The best functions are those that can be codified easily and offer benefits to either (a) the firm in terms of risk management or efficiencies or (b) the client in terms of service improvements.

A Technology Infrastructure to Support NLP Will Need to Be Implemented

While it is fair to say that implementing NLP is not as complex and challenging as some of the other technologies in this book (say Big Data or Cloud Computing), it is still challenging.

For the inbound communications then the following will be required:

The NLP front-end and rules engine is likely to be provided by an external supplier. It is possible to develop this technology in-house but it is simpler to purchase a package and then configure it. Therefore, the supplier must be selected carefully because once NLP is live and being used then this supplier will become a strategic supplier. Appendix A provides a list of the activities and areas to be covered when selecting a supplier.

Furthermore, the infrastructure (as in servers, networks, etc.) that will be supporting the NLP technology will need to be secure and robust. Remember that NLP is very visible to the external world and customers, and if problems are encountered then these will be seen by everyone. For example, if the contact center telephone NLP functionality is not working or unavailable then customers will notice immediately. This means that the infrastructure has sufficient resilience and BCP arrangements.

Therefore, some type of formal technology platform vendor selection process must be followed covering the points listed below:

Integration with Back-End Systems Will Be Challenging and Should Not Be Under-Estimated

While NLP is a very useful technology (especially if it is implemented correctly), it cannot work in isolation. When NLP “reads” in communication, it needs to pass some sort of message to another system for processing. Likewise, when NLP sends out a communication, it will need to be told what to say by other systems. Therefore, NLP cannot be implemented on its own and it will need to be integrated with existing systems.

Unfortunately, all firms often use legacy back-end systems built on older technology and supported by a combination of suppliers and in-house so implementing. This means that any integration would be complex and costly. Therefore, it will need careful design and planning from the start.

Ensure There Is Sufficient Governance and Control in Place around the NLP Platform

As well as the implementation of the new NLP technology, new processes, and governance frameworks must be designed and implemented to control the new technology. This will cover four main areas.

Ensuring the Firm Has Sufficiently Skilled Staff to Support NLP once the Project Completes

A dedicated project team will have been formed to implement the initial NLP project. This will have consisted of senior management to provide oversight and steer plus many “on the ground” people who would perform the developments, integrations, and so on required. This group of people would have been sourced from in-house staff, contractors, and possibly staff from the platform vendor.

Therefore, firms must develop the necessary skills to be able to support the NLP platform once the project closes, because otherwise they will be reliant on costly external contractors and platform vendor staff.

This means that senior management will need to be educated on understanding NLP and its benefits at a general level.

Also, more junior staff will need to be trained on NLP, the specific platform selected, the integration with back-end systems as well as the suite of bots developed and rolled out as part of the project. This training can be done by training existing staff but it may also be necessary to recruit new permanent staff with the required skillsets.

The Majority of Issues Relate to Understanding Inbound Communications

During implementation, the majority of problems will be trying to fully understand inbound communications. For example:

Therefore, any NLP rules developed will need to be able to cope with the above.

There Are Still Some Challenges With Natural Language Generation

There are also some issues with outbound communications (although they are less challenging than inbound communication). For example:

It Is Important to Ensure Lexicons and Vocabulary Is Kept Up-to-Date

Firms must ensure that NLP lexicons (for understanding inbound communications) and vocabulary (for outbound communications) are constantly maintained to ensure they meet the needs of the system. There are several ways that this can be performed.

NLP Still Needs Some Social Acceptance by Customers and the General Public

There is an issue or challenge with the social acceptance of NLP.

Customers can be nervous with both (a) speaking and communicating to a computer or (b) a computer speaking or communicating back to them. Therefore, they will often stay on hold until they are passed to a human operator. This means firms still need humans in place.

Secondly, staff internally can be uncomfortable if they feel their activities are being monitored. In effect, some sort of “big brother” oversight. This means that firms need to be very clear about what monitoring is in place and why this monitoring is required.

Future Challenges

Table 10.1 Future challenges of NLP

Area	Details
Increased regulations	The impact is neutral at the moment. No specific NLP-related regulations have been implemented so far (apart from the existing regulations around ensuring operating models are robust, customers are looked after, etc.). However, if NLP becomes more popular and/or there are many major issues then the regulators may look to implement focused rules around NLP.
Changing nature of clients	The impact on clients is positive. NLP should allow customer servicing to be improved and to be run cheaper which in turn should allow for clients to be charged lower fees. However, there is a slight downside because some customers will be unhappy or uncomfortable “speaking” with a computer. Therefore, firms need to still ensure they provide a human to speak to if needed.
Evolution of products	The impact on products is positive. NLP allows products to be offered and serviced better and cheaper. Although the issues noted above regarding customers not wanting to “speak” with a computer will need addressing.
Lack of trust	The impact could be negative if not managed carefully. Customers can be nervous with both (a) speaking and communicating to a computer or (b) a computer speaking or communicating back to them. Therefore, they will often stay on hold until they are passed to a human operator. This means firms still need humans in place. Secondly, staff internally can be uncomfortable if they feel their activities are being monitored. In effect, some sort of “big brother” oversight. This means that firms need to be very clear about what monitoring is in place and why this monitoring is required.
Accurate data	This area is impacted adversely. NLP is very reliant on accurate data for processing inbound communications, building rules, processing outbound communications, and so on. If there are gaps, errors, and so on in the data then it will impact how well NLP can operate. Therefore, firms will need to implement more processes and rules around ensuring all NLP-related data is accurate.
Poor operating and technology models	This area is impacted adversely. While NLP does offer real business benefits, it does require another complex and business-critical component to be added into already existing complex operating models, thus increasing complexity and risk further.
profitability/Cost drivers	This area is impacted favorably albeit in the medium to longer term. While NLP will improve efficiency as well as reduce costs and risk (e.g., around frauds), it does require a sizeable implementation fee which will take time to payback.
Changing nature of the workforce	This area is neutral. While some staff will lose jobs because their roles have been replaced by NLP processes, there is the opportunity to learn new skills to support NLP going forward.
New competition and replacements	The impact here is neutral. Although NLP may be able to allow new competitors or entrants to move in the business slightly quicker and cheaper.
Risk profile	There is a neutral impact on risk levels. While NLP will reduce risks around customer servicing, operational efficiency, and internal risks (around fraud), it is counter-balance with the additional risks to the operating model and the need for correct data.

Case Study

This case study relates to a global Asset Manager who used NLP techniques to analyze all their external contracts.

This manager had a large number of legal contracts. Some of them were recent and others were several decades old. Also, the actual contracts were spread over various formats (such as MS-Word, PDF, pictures, scans, paper copies, and microfiche).

The issue was that the manager did not have a clear understanding of all the clauses across the contracts. For example, what were the manager’s commitments? what were the suppliers’ commitments? what are the termination clauses? what liabilities are in place? plus others. Therefore the manager had no real understanding of their legal risk.

The manager decided to use NLP to try and address this problem, namely:

While a large manual or human effort was required, the NLP rules dramatically reduced the timeline and cost required for the initial assessment from several months down to a couple of weeks. This allowed the firm to focus its legal teams on the high-value work of actually changing the contracts.

Summary

NLP is a useful technology that provides benefits around customer servicing, operational efficiency, internal controls, reviewing documents, and providing standardized client data.

However, the complexities and challenges of implementing NLP must not be under-estimated which means firms will need a clear business reason for implementation. This should be supported by designing a robust NLP infrastructure covering the actual technology but also the required processes, skills, and oversight required. To further reduce the risk profile then NLP should be rolled out on a gradual basis.

Finally, there is an issue or challenge with the social acceptance of NLP. Firstly, customers can be nervous with both (a) speaking and communicating to a computer or (b) a computer speaking or communicating back to them. Therefore, they will often stay on hold until they are passed to a human operator. This means firms still need humans in place. Secondly, staff internally can be uncomfortable if they feel their activities are being monitored. In effect, some sort of “big brother” oversight. This means that firms need to be very clear about what monitoring is in place and why this monitoring is required.