While artificial intelligence is steadily automating industries, many financial transactions within the global capital markets are still conducted the old-fashioned way: in conversations between people. Every day, tens of millions of these conversations are carried out via chat messages exchanged by financial professionals using Instant Bloomberg (IB).
Researchers in Bloomberg’s AI Group used the publicly available STAC (Strategic Conversation) dataset to study the dynamics of interactive conversations related to negotiated financial instrument transactions– also known as Dialogue Acts – in the online version of the game The Settlers of Catan, where trade negotiations are conducted in chats. Senior Research Scientist Ozan İrsoy is the lead author of the paper highlighting its findings titled “Dialogue Act Classification in Group Chats with DAG-LSTMs.” It was presented by Senior Research Scientist Mu-Hsin Wei, one of the paper’s co-authors, during the 1st Workshop on Conversational Interaction Systems (WCIS 2019) at the 42nd International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR 2019) in Paris.
Additional co-authors from the company’s AI Group include Rakesh Gosangi, Haimin Zhang, Peter Lund, Duccio Pappadopulo, Brendan Fahy, Neophytos Nephytou, and Camilo Ortiz. The group hopes to extrapolate from their findings to develop a method for identifying behaviors that can help automate downstream processes for traders and financial institutions.
During the workshop, Camilo Ortiz, who leads all of the Communications-related efforts in Bloomberg’s AI Group, also delivered an Invited Talk titled “Dialogue Understanding and Finance,” in which he introduced some key concepts explaining how the global capital markets function and the role of social interactions in trading, as well as use cases for dialogue understanding in the finance industry.
Understanding a Conversation
“When you have a conversation – or an IB group chat, in particular – in which people are negotiating, there’s a state of that conversation that you want to follow,” explained Ortiz. “In simple terms, conversations start with an inquiry that gets acknowledged and responded to. Sometimes, an order is created that needs to be executed.”
Bloomberg is moving the state-of-the-art by developing methods in the field of dialogue understanding and adapting them for the content of IB chats. Upon client request to implement the add-on features that use these machine learning models, clients can utilize workflows in which Bloomberg pulls information pertinent to a financial transaction from an IB chat.
“We are always exploring new technologies that will improve the efficiency and effectiveness of our communications channels,” said Heidi Johnson, Global Head of Community, Collaboration and Compliance Products who supported the development of the white paper. “IB is an important part of our clients’ workflows. By studying and understanding conversation patterns, we can further enhance this offering to meet the needs of our clients.”
“[If people are] talking about a particular financial instrument [on IB], we want to surface the Bloomberg function so the trader can execute the trade – but you need to ignore the small talk amongst chat participants,” said Ortiz.
Each post is an act defining a purpose. Detecting those chat posts that mention financially-related content, such as inquiries, as opposed to requests for random, non-financial information, help identify the speaker’s intent.
Bloomberg’s natural language processing models also help provide context within a chat by analyzing the words surrounding a product mention. A statement like “I’m trading McDonald’s at this price” demonstrates intent, but the word “Apple,” for example, could reference a fruit or a company.
Modeling the Data
All negotiations have similarities. However, they differ in intent, the language that may be used, and the entities bought and sold, and the distribution of the instruments. For example, in the game The Settlers of Catan, players barter the natural resources they hold (sheep, wood, iron ore, etc.) similar to how financial traders discuss buying and selling securities, thereby making the conversations between players a good dataset for this analysis.
“But that’s where the similarities end,” expressed İrsoy.
Dissecting chats among multiple people has its challenges. Messages can be taken out of context. Responses aren’t always to the prior post in the chat, but may sometimes refer to a message a few posts back. One person could also be responding to multiple people, creating interwoven threads. When someone confirms or refuses an order, they may do that in several different posts.
Entire conversations are treated as an ordered sequence of posts, and each post is an ordered sequence of words or tokens, said İrsoy. “Words are represented by word vectors, and we use state-of-the-art neural architectures to turn sequences of vectors into other vectors. Given the sequence of vectors, you get one post vector, and given the sequence of post vectors, you can get a conversation vector, and then a classification label. This can give you a probability for one of the possible labels, such as ‘inquiry.’”
Posts from the same users are linked and every post is tagged by the models. Prior posts provide context as to whether someone started a new thread or responded to a previous one. The goal is for the models to hopefully deduce whether the speaker repeated dialogue or started a new conversation.
To improve the model’s context representation, skip connections identify someone’s history in an entire conversation and provide a graph structure so that posts are not bunched together in sequential order.
“The baseline approach allows you to represent context by treating posts in a sequence – a chain of posts,” noted İrsoy. “In our novel approach which uses additional skip connections, it is no longer a chain, but a graph, and the historical representation of each user provides a richer context allowing the models to read the conversations as humans would.”
Automating the Workflow
There are many applications for Dialogue Act classification, including conversation summarizing, question answering, and workflow automation. Posts are enriched using any available metadata.
For example, if someone in an IB group chat is discussing commodity futures, the order form in a Bloomberg Terminal function can be populated with that information, eliminating the need for the user to type the information in another screen on the Terminal.