This section will introduce some of the key aspects of how Re:infer works end to end: what data it can process, how does it process your data, the characteristics of models built in Re:infer, and the types of solutions it can integrate with downstream.
What's covered in this article:
What kind of data works with Re:infer?
Re:infer is a natural language processing (NLP) tool optimised for understanding unstructured textual communications data. As such, it can learn to understand any typical freeform written (or transcribed) communications between teams, businesses and customers.
Re:infer is not designed to handle data specifically from any industry. At Re:infer, we're industry agnostic and have clients across a broad range of industries, including investment and retail banking, insurance, e-commerce and telecoms. The nature of Re:infer's bespoke models mean that users, who are experts in their own domain, can impart their specific industry knowledge into the models they create in our platform.
Most common data types that Re:infer can interpret and add structure to:
- Emails (most typically shared inboxes)
- Tickets (e.g. customer service tickets)
- Chat messages (e.g. customer-agent conversations)
- Customer feedback (e.g. survey responses, complaints, suggestions)
- Call transcripts
Data that Re:infer does not interpret:
- Email attachments (e.g. PDFs)
- Financial data in Excel
- Website / Product click data
Why does Re:infer not interpret email attachments directly?
Email attachments can be incredibly varied, ranging from PDFs, to Excel Spreadsheets, to screenshot images, and introduce a huge amount of additional complexity to understanding the nature and intent of the attachment. Typically understanding the attachment is not an NLP-related problem.
Even turning standardised forms in PDF format into digital text that could be input into a model is a complex problem that specialist OCR (Optical Character Recognition) platforms devote all of their efforts towards.
When emails with attachments are uploaded to Re:infer, our platform filters out and ignores the attached files, relying on understanding the written content of the communication itself.
When our clients have use cases that rely on understanding the communications themselves, as well as the attachments, we work with them to devise integrated solutions that partner with other technologies that can process the attachments. If you're interested in finding out about the possibility of running a similar use case, please reach out to our team.
How does Re:infer work?
The diagram below shows an overview of the journey that your unstructured communications data goes on when it's uploaded to Re:infer:
We start by connecting (typically via live integration) your various sources of unstructured communications data to Re:infer to ingest and store
Deep learning NLP technology then parses and understands the communications
Using unsupervised learning, the platform finds frequently expressed requests and intents
You define and train your own taxonomy of labels (and entities) to capture what’s in your data in a supervised learning stage
Re:infer provides analytics on your data that help you discover insights and opportunities
In-platform validation functionality lets you monitor and improve model performance if needed, at which point the platform retrains and re-infers the predictions across your data.
You can then deploy models in production use cases. Re:infer provides label and entity predictions in a structured format via its API to facilitate automations via downstream systems or RPA, or to feed data into databases or specialist data visualisation tools.
Characteristics of Re:infer models:
- Every time a user labels data in Re:infer, it retrains and re-predicts across the entire dataset that was updated
- We've optimised this training process so that every model updates rapidly while users are training, ensuring they're provided with the very latest predictions and insights as quickly as possible
Easy to train
- Compared to alternative NLP solutions or custom built models developed programmatically, Re:infer models are incredibly easy to train
- Not only does Re:infer have a variety of training modes in our zero-code user interface that help users create the best possible set of training data, Re:infer actually guides users on any corrective actions required to help improve their model's performance, using next best action recommendations
- Our models are extremely flexible, allowing users to quickly update their taxonomies by adding, merging or removing concepts with a few clicks
- Users can also copy existing models within their instance of Re:infer and then tweak them for their own objectives
Require minimal training data
- We've spent a significant amount of effort on optimising our models to be able to predict concepts with very training examples, making the training process quicker and easier
- Compared to open-source solutions, Re:infer's models require a significantly lower proportion of training data per concept to perform well
- Models in Re:infer are highly scalable - very large datasets in Re:infer typically don't require much more training data than smaller datasets, and our infrastructure is designed to scale as your data requirements do
- Once you've well trained your concepts in your model, it can work on an infinite amount of this same kind of data, meaning that you benefit more from Re:infer the more data gets added
Multiple predictions per communication
- Re:infer is not a single classification tool, it understands labels independently from each other and for each communication it can predict as many labels as are relevant
- This helps businesses obtain both broader and more specific insights from their communications data, and allows them to analyse instance where labels co-occur (or don't) with each other
Integrating with Re:infer downstream via our API
Re:infer provides label and entity predictions in a structured format via its API to facilitate automations via downstream systems or RPA, or to feed data into databases or specialist data visualisation tools.
Here are some of the tools that are typically integrated downstream with Re:infer via our API:
- RPA software
- Workflow tools
- Data analytics and visualisation tools
- Data warehouses / Large databases
- And any other tool that wants to extract data from Re:infer's API (with the right permissions)!
Get started with Re:infer
Does your company already use Re:infer?
Read Getting a Re:infer account to learn how to sign up and get started using the platform
Ready to keep learning?