How to answer regulators questions while accurately gauging the impact of climate change on corporate sustainability? That is a key question faced by Sequantis, a specialist provider of monitoring tools and services for investors. In enabling analysts to directly extract useful data from activity reports with a minimum of fuss, reciTAL’s platform can radically reduce analysis times and enhance the accuracy of the information collected. Discussing this productive partnership are Sequantis CEO Nicolas Fournier and his reciTAL opposite number Gilles Moyse.
What triggered your interest in the analysis of the environmental impact of companies?
NICOLAS FOURNIER: The climate question and its impact on economic and financial activities is one that’s been asked for some time now. When the Bank of England asked British insurers to undergo a stress test in 2019, forcing them to assess their ability to withstand climate change, they had to get themselves in working order for it. A few different scenarios were dreamed up, with some manufacturing companies having to break down their activity to identify the details of their economic operations and their initiatives for fighting climate change. At Sequantis we started by assessing everything we had available to us in meeting this requirement. This is an area in which companies publish official documents such as activity reports, which are packed full of data.
GILLES MOYSE: When Nicolas spoke to me about this project in late 2019, reciTAL had just embarked on a use case with its Extract solution for the extraction of LIBOR clauses in structured loan agreements. If Extract could work for loan agreements, it could potentially extract data from corporate reports.
How did you find the relevant information in these reports?
NF: The platform extracts fragments from documents in response to a request on a given subject. The analyst asks questions and the platform automatically brings up answers for them, highlighted in yellow: insets, graphics, tables, and parts of paragraphs. The solution can restrict the field of possible answers to a set of data that the analyst can then focus on. The idea is to ensure that a 500-page document can be read in 30 minutes instead of half a day, which is what it would normally take.
GM: The only reason this has never been done before is not that people couldn’t be bothered to do it but because it was technically impossible. Corporate documents are so bulky that even with the best will in the world you can’t read everything in sufficient detail. What we offer is a bespoke summary of a document in response to specific questions. It saves an awful lot of time.
It’s a simple process but I guess there must be a few pitfalls to be avoid along the way.
NF: The activity report is an amazing example of an unstructured document, with all these different formats, text and sentences written in totally different ways and tables presented in different ways, depending to the stakeholder. If one of them talks about renewable energy, another talks about solar and wind, and another about photovoltaics. What we’ve done is provide a structure for unstructured data.
How do you use this data once you’ve got it?
NF: What we’re looking to express are the negative effects that a company might have on the environment and also the way in which climate change can impact on its activity. If sea levels rise by 150cm, then there’s every chance that is going to impact on a factory located on the coast. That’s the kind of information we’re looking for in trying to deduce the risks to companies. The list of everything we can potentially identify is just huge. We can find as many as 200 pieces of data for each company. Given that we monitor 35,000 companies, that’s 7,000,000 pieces of data to extract. We’ve got a lot of work ahead of us!
GM: What’s so exciting about this project is that it gives us the ability to structure data extracted from documents written by people. Entire audit businesses are founded on that very principle. Annual reports are mines of information that can feed into so many different lines of analysis. The idea is to put questions to documents as if they were oracles. Ask them something and they give you an answer. That’s the goal.
What role does AI play in your solution?
GM: It operates on three levels: understanding documents, semantic indexing, and answering the questions asked.
The first building block is breaking down the document. The human eye can do it without any problem, but for a machine the task of reading a table or identifying a graphic involves a whole learning process. It’s a cross between computer vision and automatic language processing, where we use deep learning, an area in which we’re at the cutting edge.
The second building block is founded on modes of semantic indexing. Learning through deep neural networks allows the machine to identify the semantic proximity between two words, such as “car” and “automobile”. The neural network learns language by itself, saving the user from having to enter long lists of words, synonyms, acronyms, typing errors, etc.
The third building block involves finding data through a question asked in natural language and not simply by matching key words. The platform can now propose a combination of these AI building blocks.
What role do humans have in this machine learning model?
NF: There are two levels to the human aspect. Firstly, there’s pre-process because you have to know what you’re looking for and what the right question is. Secondly, you have to know how to interpret results and then broaden your research. Let’s be clear here – we’re not looking to remove people from the process but to make things achievable. It increases our ability to say that a company is environmentally ethical because we will truly have looked at everything it’s done, without missing a little line with a potentially big impact.GM: These machines are light years behind [RJ(1] of humans when it comes to comprehending things. They answer questions without having a goal or aim, whereas the analyst does have an objective. Our platform is a data mining solution. We ask a question and it looks for the answer, making the expert’s job that much easier.
Paris, November 27th 2020.
Want to know more about our Extract solution ? Request a demo.