The Universal Wealth Management Platform (UWMP) project has the objective of creating a new service model in the financial domain. An integral part of this service model is the creation of a new Virtual Customer Assistant, that is able to assist customers via natural language dialogues. This paper is a report of the activities performed to develop this assistant. It illustrates a general architecture of the system, and describes the most important decisions made for its implementation. It also describes the main financial operations that it is able to assist customers with. Finally, it delineates some avenues for future work.
• Human-centered computing → Natural language interfaces; • Computing methodologies → Information extraction; Discourse, dialogue and pragmatics.
The purpose of Wealth Management firms is to advise clients on investment strategies, execute orders on their behalf and help them custody their financial assets. This type of work has always been closely associated to personal relationships and confidentiality. Today however, we expect that the coming of age of a new, digital-savvy generation of wealthy individuals will change the relations between clients and wealth management firms.
Digital Assistants such as Amazon Alexa and Apple Siri have popularized the notion of software applications helping users with everyday tasks. In the realm of business, such applications are also known as Virtual Customer Assistants and they support text or voice interactions to deliver information or to act on behalf of the customer. The goal of a Virtual Customer Assistant is twofold: first, to take over routine interactions so that human service agents can engage in more value-adding activities; second, to improve customer experience by reducing friction, i.e. eliminating any factor that can deter or slow down interactions between clients and ifrm. There are more than 200 software companies ofering Digital Assistants for customer interaction. In the financial domain some prominent cases are IBM Watson, Microsoft Virtual Assistant, boost.ai, Creative Virtual. Most of these assistants interact via the bank’s mobile app and web site, while a few in the USA use Facebook Messenger.
For example, Royal Bank of Canada has a chatbot called
Arbie that helps users open a new account. Bank of
America’s assistant called Erica [
Objectway is a global Fintech 100 software provider of the wealth, investment & asset management industry. The company launched its Universal Wealth Management Platform (UWMP) initiative to implement a new, fully-digital service model destined to financial institutions. The new digital model complements the existing model based on financial advisors and physical interactions, by ofering clients selfserve functionality on a wide range of investment services.
One of the activities of the UWMP project in support of the new digital service model was to implement a service that enables clients to interact with wealth management firms using Natural Language dialogue. The goal was to empower clients with informational and transactional capabilities that were always-on and accessible from everywhere. The new service would have to satisfy the requirement of confidentiality, so priority was set on text interactions. Work on voice was postponed due to potential issues with revealing sensitive client information. Another requirement was to ensure high quality of interactions, since failures would have an adverse efect on the perception of quality and trust that clients place on wealth management firms. This advised us to focus on the most recurring, time-consuming routine tasks that clients faced when interacting with the firm, such as requesting information on account balance and performance.
To try out more complex dialogues, we also decided to explore tasks of a transactional nature, although we expect the adoption of such features in real life to be slow. Investment transactions increase the dificulty of interaction due to the number and variety of financial products, order types and the associated ambiguity that derives from such complexity. The rest of this paper is organized as follows: the system architecture is describe in Section 2, while Section 3 shows how the system works. Conclusions are outlined in Section 4. 2
Chatbots can be classified as open-domain or goal-oriented.
Open-domain agents handle generic conversations, while
goal-oriented chatbots interact with users via natural
language conversations to assist them in their tasks. Goal-oriented
CAs can then be classified as informational, transactional
or advisory. Advisory chatbots learn from users, and can
recommend products based on the interaction, such as in
[
This architecture is called modular, as it is made up of several components, each with a specific task.
The Speech Recognizer component has the responsibility of transcribing the user’s voice input into a text message. This is an optional component used by conversational agents that support a voice-based interaction.
The Natural Language Understanding (NLU) component takes in input the user’s text message, and outputs a semantic representation of that message. The objective is to understand the meaning of the message itself. This means performing several Natural Language Processing tasks. In the NLU component, Intent Recognition is done to understand the action or the request that the user is making. For example, the message "Pay 200 USD to Alice" means that the user is trying to start a payment. Entity Recognition is also a common NLU task, which is used to recognize mentions to entities such as people, organizations, or numbers. From the previous example message, it should be able to extract "200 USD" as a monetary amount, and "Alice" as a person name. Sentiment Analysis is performed to understand the emotional state expressed through the message. This could be used to make the agent react diferently to diferent emotional states.
For example, a customer support agent can decide to hand of the request to a human operator, in case it detects that the user is frustrated.
The Dialog Manager (DM) component maintains an
estimate of the current state of the conversation. Based on
this state and the current user message, it decides the next
appropriate action to perform in order to complete the user’s
request. For example, the DM might respond by asking a
question (if some information is needed), or by performing a
transaction (if all information has been provided). Typically,
Goal-Oriented agents employ a frame-based Dialog Manager
[
Our agent implements the NLU, DM and NLG, with
Intent and Entity recognition functionalities implemented in
the NLU component. An analysis phase was conducted in
order to determine the most appropriate tools for
developing the aforementioned components. Several options were
vetted, based on their ability to cover the requirements of
the project. In particular, we opted for solutions that
require little training, since traditional supervised methods
require large amount of domain-specific conversational
corpora, such as [
The Entity Recognition (ER) module used in the system is
implemented using Stanford CoreNLP3 and Apache Lucene4.
It is able to do both Named Entity Recognition and Entity
Linking. The ER function can recognize entities such as
numbers and person names, and also mentions to entities that
are contained in a dictionary (e.g. financial products). It uses
a combination of classifiers in order to recognize entity
mentions in the user message. A regex (regular expression)
classifier is used to recognize entities and keywords using exact
matching. Additionally, it uses a text classifier based on
Conditional Random Fields (CRF) [
The Dialog Manager component was also custom-made for this project. As said before, it uses frames to store the goal-specific information, and slot filling to acquire said information during the dialog. Given a user message, the Dialog Manager can either: perform a transaction, request a slot, or request a disambiguation. Disambiguation is required when multiple entities are found as possible candidate values for a slot. For example, given the user utterance "I want to buy 200 BMW shares", the ER component may recognize two diferent products: "BMW AG" and "BMW ORD". Therefore, the user will be asked to clarify which one should be selected. During the conversation, the Dialog Manager also performs checks on the slot values, and can auto-fill slot values based on some conditions (e.g. when the user makes a payment to a known payee). 3
The main objective of the Virtual Customer Assistant created for the UWMP project is to assist customers of banks or other financial institutions. The bot is reactive, that is, it responds to requests made by users. It can respond both to information requests, and can perform financial transactions. This means that it can be categorized both as a informational and transactional agent.
From the informational side, the agent can be used to obtain general information about the user’s financial situation, such as the balance of the user’s cash account. Figure 2 shows an example of this use case. Given the message "What is my cash account balance?", the NLU component will recognize the account balance intent, and the agent will provide the requested information, responding with: "Your cash account balance is 4890.00 EUR". Users can also keep track of the performance of their investments, such as the value, the performance, and the profit and loss . The agent can also provide a report for a specific time period provided by the user. For example, a user can type "What was my portfolio performance in the last three weeks?", or "What is my portfolio performance from May 1st to August 31st?". This is possible thanks to the time expression recognition component of the Entity Recognizer, described in Section 2.
Users can also perform financial transactions directly from the chatbot, as said earlier. The most notable transactions are payments and investments. Indeed, users can send payments by writing for example "Send a payment of 200 USD to Alice Smith". In this case, the NLU component will recognize the payment intent, and some of the required entities, such as the payee "Alice Smith", and the amount of "200 USD". Alternatively, the user may not provide any information at the start, e.g. by saying "I want to make a payment". The Dialog Manager reacts to this accordingly, by collecting all the mentioned entities, and prompting the user for the missing information via slot filling. This can be seen in the example in Figure 2. In this case, the user provided the payee, but not the amount. Thus, the Dialog Manager responds by asking: "Please tell me the amount to transfer". If the user wants to make a payment to an already known payee, some slots will be automatically filled, making repeated payments easier. If multiple payees are found, the Dialog Manager will prompt the user to specify the correct one. When all the required information is collected, the Dialog Manager will ask for a ifnal confirmation, e.g. by saying "I prepared a payment of 200 USD to Alice Smith, alias Aunt Alice, at Barclays account...". An authorization code is required to complete the transaction, to increase security. When the correct code is entered, the agent completes the transaction, and returns a positive feedback to the user, as seen in Figure 2. The agent also lets the user invest in funds or trade stocks. Investments can be made by specifying a product, the number of stocks (or the amount of money to invest), and the direction (buy or sell). For example, given the message "I want to buy 200 BMW shares, the investment intent will be recognized by the NLU component, as well as the amount "200", and the "buy" keyword, indicating the direction of the trade. The "BMW" entity will be extracted by the Entity Recognizer, and the Entity Linking component will try to match it to two diferent products: "BMW AG" and "BMW ORD". In this case, the Dialog Manager will request the user to specify the correct one, e.g. by saying "I found several matches for this product.". This can be seen in the example in Figure 2. Just like for payments, missing information is acquired via slot filling.
In this paper, we presented a Virtual Customer Assistant developed for the UWMP project. The aim of the project is to assist users in completing financial operations of varying complexity using natural language dialogues. The agent supports a wide range of informational and transactional operations, and several components were developed to support them. As future work, we plan to investigate the introduction of a voice-based interaction model. Another avenue for future work is the introduction of advisory functionalities, such as financial product recommendations, that enable the agent to take a more proactive role in the interaction. Finally, the Virtual Customer Assistant will be subject of in-vitro and in-vivo experiments, that will assess its ability to aid users in their financial tasks.
This work has been funded by the project UNIFIED WEALTHMANAGEMENT PLATFORM - OBJECTWAY SpA - Via Giovanni Da Procida nr. 24, 20149 MILANO - c.f., P. IVA 07114250967.