The finance industry has been an early adopter of AI. It is likely that the use of algorithms in trading and the fact that most large financial firms already have teams of software developers aided the transition into data science and AI applications in the industry.
The advent of other AI use-cases in finance, such as chatbots and conversational interfaces, has seemingly also driven up the interest in natural language processing (NLP). NLP could allow a company to garner insights by summarizing documents or gauging brand-related sentiment across the web.
We spoke with Peter Hoopes, VP WW Sales at Gamalon, Inc., who laid out some of the value that NLP might bring to customer service in finance.
In this article, we’ll start with an overview of natural language processing in finance, and then we’ll discuss two prominent applications for NLP across insurance firms, banks, and other financial institutions. These applications are as follows:
- NLP for Searching Customer Support Inquiries
- NLP for Analyzing Trader-Client Interactions
We also delve into one of the common challenges that businesses in finance face with NLP and machine learning projects: the “black box” of machine learning. Business leaders in finance will find that many AI and ML products available commercially are essentially black boxes, where it is not really possible to inspect or understand how the algorithm is accomplishing what it is accomplishing. We also discuss how financial businesses can adapt to the black box challenge.
NLP for Searching Customer Support Inquiries
Businesses in the finance sector have historically collected vast amounts of data about customers, financial transactions, and markets, in many cases due to regulations. This includes customer interaction records for incoming calls, email, text messages or social media chat transcripts.
Large financial institutions have millions of customer service tickets coming in from customers across the globe. Each of these tickets could be relevant for one or more internal departments within the firm. Additionally, these customer service requests come in through a variety of communication channels; for example, customers could be calling in or filling out a customer service form on a website.
Financial institutions collect data in the form of free-form text in customer support tickets and call transcripts in which customers describe their issue. The sheer scale of the incoming requests makes it difficult to read through each ticket manually and take action.
NLP and machine learning software could help with automatically having a computer analyze these customer service messages and categorize them, and ideally, predict a next best action.
For example, let’s say JPMorgan receives large volumes of customer service requests from customers all over the world. NLP software could, in theory, help a bank sift through all the customer messages and automatically identify the top customer concerns.
The software might find, for example, that customers have been comparing a certain product to another brand often in the last month. The software could then route the tickets to the product or marketing teams.
The challenging aspects in extracting genuine insights about customer satisfaction might lie in developing NLP algorithms that understand context when reading customer messages.
For instance, if a financial institution wants to understand how customers feel about their wealth management team by reading customer feedback, the NLP algorithms might need to be tweaked to “understand” certain financial terms and common trading jargon.
This might well be a much harder task than it appears, especially because customers might ask for the same request in several different ways. NLP algorithms can be trained to identify and categorize these customer inquiries automatically. If a significant number of the messages had sentences that started with variations of the phrase “my spouse,” such as “my partner” or “my husband/wife,” NLP algorithms can be trained to label such inquiries and send an alert to the team handling joint accounts.
It is also important to note that the accuracy of most AI software capabilities are measured up to a reasonable “level of confidence;” the software might not be capable of performing a task accurately a 100% of the time.
For an NLP software being applied to categorize a large volume of documents, the algorithm might have to be adjusted to account for edge-cases where the software might be unable to make a decision.
For example, If a large financial firm uses NLP software to automatically label and extract the top customer complaints from open-ended forms on their web portal, the software might be trained by using datasets of labeled sentences.
The software might identify that terms similar to the word “rude” were commonly associated with the complaints for a customer service team in a particular branch, thus allowing the financial firm to take action and introduce better training programs for that branch.
Another option might be to use human experts to train the software to identify more complex associations faster. The software can then start automatically labeling new customer messages by predicting a probability score for each label by analyzing relevant parts of each customer message.
Considering the case where a customer complaint message said the online trading platform did not have a good interface, but the customer service reps were very helpful in resolving the issue, the software might identify positive and negative sentiments in the sentence, but human subject-matter experts might be better at identifying which internal department label needs to be added onto parts of each sentence.
There will also exist cases where the software has a low probability of predicting the correct ideas for each message. Having human financial subject-matter experts categorize these “low confidence cases” might allow the software to “learn” to identify and categorize more accurately.
NLP for Analyzing Trader-Client Interactions
NLP might essentially allow large financial firms to automatically read and categorize documents which contain free-form text. Hoopes mentions the example of a large finance firm which needed to identify customer complaints that might involve cases of regulatory non-compliance.
Financial advisory services are highly regulated and recent changes seem to be leaning in the direction of having the onus on financial firms to monitor the performance of their advisors.
If a large financial advisory received a customer complaint, such as charging a fee that was not clearly stated initially, the firm could be in violation of certain regulations, attracting a fine or a bad regulatory rating from supervisory agencies.
NLP software can run through several millions of incoming customer service requests to identify and categorize ones that need immediate action to avoid regulatory issues. The software can be trained to “read” a message and automatically populate a table or schema from the text to list out customer requests that need action from the compliance team.
Hoopes gives the example of a financial services firm trying to understand the interactions between traders and clients. These companies need to identify broadly what their traders are saying to their clients and what advice is being given to each client.
There has been a recent regulatory push from supervisory bodies towards ensuring that traders aren’t giving bad advice to clients. NLP can help understand these customer interactions at a massive scale and gain insights such as the ability to gauge the performance of financial advisors.
NLP and machine learning could be used to read through transcripts of trader-client interactions and identify which parts relate to any financial advice being given. For example, if a trader says “I would advise you to place a buy order for these trades,” the software could be trained to label and tag parts of this sentence in the context of financial advice.
The NLP algorithm would most likely identify the words “buy” and “trade” in the context of a sentence beginning with the phrase “would advise you to.” It would then likely categorize it as financial advice.
Similarly, the software might also help identify and extract the parts of a conversation where the trader is suggesting a trading action to the client. The firm could read through these advisory messages from their traders to detect any instances where the wrong advice might have been given, possibly indicating the need to improve trader training processes.
Additionally, Hoopes explains that capturing the trader-client interaction from call or chat transcripts might help with identifying fraud. For example, NLP-based software might help financial services firms identify that a particular series of sentences always show up in conversation records for cases of fraud.
Using historical evidence and public datasets, finance firms can generate a list of common words, phrases, or topics that are associated with fraud.The NLP algorithms crawl through the messages to identify sentences that contain any of these words of phrases.
The software might also automatically cross-reference historical fraud cases and find previously unidentified patterns in conversation that lead to cases of fraud.
A few vendors also offer software that allows financial institutions to dig deeper into their data and identify undiscovered fraud patterns through diagnostic tools. These might be in the form of a dashboard that allows the non-technical employees and financial subject matter experts to edit the way the algorithm labels.
Employees can also see ranked lists for certain phrases or topics that mean the same thing and make additions or changes. The algorithm learns to label sentences better with more such inputs from subject matter experts.
Business leaders in finance might also need to be aware of the range of capabilities of both NLP and what a particular vendor can offer. The most common approach to NLP software offered by vendors catering to the finance sector seem to grant financial firms the ability to summarize text by extracting parts of a document that the software deems as useful.
For instance, JPMorgan announced they internally-developed a contract abstraction software tool called COiN, which uses NLP to automatically extract the most useful portions of contracts.
Most such software also allow for a level of automatic categorization of documents. The software “learns” to categorize the document by reading through the document and learning from examples categorized by humans.
Hoopes noted that even among the NLP vendors there might be levels to what capabilities are offered for financial firms. Some vendors offer software that can perform the extraction and categorization tasks mentioned above. But the software might require the expertise of data scientists whenever the algorithms might need to be tweaked to accommodate new data.
The “Black Box” of Machine Learning in Finance
Some vendors offer NLP software that are already tailored to a particular industry, but the inner working of the algorithms might be somewhat unclear. Financial firms that need a quick integration might choose such a vendor, although it might come at the price of not being able to fully understand how the software is coming to the conclusions it is.
More often than not, AI vendors today offer NLP software that is a “black box.” The software might take in data as input and the algorithms might be tweaked to calculate a desired output, but it is very challenging to understand each step in the decision-making process of the algorithm. Additionally, modifying the software to account for a new data category would require data science expertise, time and resources each time.
Some vendors, such as Gamalon, offer NLP and machine learning software they claim could offer an alternative to the black box. Gamalon’s approach involves allowing subject-matter experts to tweak the algorithm through their UI without requiring coding experience.
Concluding Thoughts on AI Search Applications in Finance
NLP might help financial firms search large volumes of structured, semi-structured, and unstructured documents and extract data from them.
A lot of the business in the finance industry still gets done on paper. Advanced optical character recognition and computer vision software can now help financial firms digitize these document, allowing NLP software to search them.
This could open up a new avenue for larger financial firms to gather insights from sources that might have been previously untapped due to the inability to gather data recorded on paper.
In some cases even with digital data capturing systems in place, the large organizations cannot manually handle the sheer volume of incoming customer service tickets coming in from a variety of channels.
NLP-based search software could be a key to allowing banks, insurance firms, and other financial institutions access to their volumes of digital and physical records, allowing them to search customer support tickets and trader-client interactions at scale.
This article was sponsored by Gamalon, and was written, edited and published in alignment with our transparent Emerj sponsored content guidelines. Learn more about reaching our AI-focused executive audience on our Emerj advertising page.
Header Image Credit: ESMT Berlin