Swedbank’s Analytics & AI @ sTARTUp Day 2020

Lehar Oha
Swedbank AI
Published in
3 min readJan 21, 2020
Source

On 30th of January this year we will be presenting some of our work at the sTARTUp Day, one of the biggest business festivals in the Baltics. The sTARTUp day gathers regional enterprises, entrepreneurs, and academia in Estonia’s renowned university town Tartu.
In our talk we will introduce Analytics & AI at Swedbank, discuss shortly our efforts on anomaly detection, and dive a little into Natural Language Processing (NLP).

As the name suggests, NLP is of great use when working with large amounts of semantic data. In banking, knowledgeable insights from such type of data allows for a better understanding of inefficiencies and thereby enhancing customer satisfaction. Tools provided by NLP enable businesses to discover obstacles, remove flaws, find opportunities as well as possibilities for automation.

Tools within NLP can be used in pattern matching, for example extracting key details from documents, an otherwise cumbersome task. For more complex extraction tasks named entity recognition can be of great help. Such models allow for finding organization identities and events from semantic data. Additionally, NLP allows for general text classification, e.g. classification of language in any written text, helping to automatically route e-mails to domain-specific human experts or doing sentiment analytics on app reviews.

In the unsupervised domain, topic models can help to find common themes in prohibitively large amounts of texts, e.g. whether customer messages to the bank are about credit cards, consumer loans or a combination of several topics. Virtual assistants are grouping of multiple NLP technologies that are designed to help users with frequently occurring and well-defined tasks.

Recently, research within NLP has had significant progress thanks to transfer learning and pre-trained models. Essentially, models in transfer learning are trained on one kind of task and subsequently used on a slightly different but related task. Traditional machine learning is predominantly isolated one-task learning, there is no cumulative effect as opposed to accumulative human knowledge. With the help of transfer learning, learning fresh tasks relies on previously learned experiences.
Transfer leaning has shown a series of state-of-the-art results in many NLP tasks and delivered several breakthroughs. Firstly and most generally, transfer learning tends to be faster when new models are needed to be trained. Secondly, the need for large sums of (annotated) semantic data is somewhat relaxed as some “common knowledge” is already captured and thereby transferred: the core idea of transfer learning. Additionally, transfer learning is of great help in multilingual problems, e.g. when serving clients in six different languages.

To follow up on the last advantage above and showcase a concrete example, we have acquired reviews of financial service apps from Google Play Store in the Baltics and Sweden. We use this data along with an application of open source research tools from Facebook, specifically LASER: Language-Agnostic Sentence Representations.
Expressed briefly, our aim is to enable transferring of reviews into vector representations, i.e. mapping of similar reviews onto close vectors. Such representations are designed to work across languages, a benefit of cross-lingual pre-training. By using this methodology, we can search for reviews with similar meanings. A concrete example is given below where we give a text input in Estonian, select a similarity metric and get similar reviews in other languages we are interested in.

In addition, we added a trained classifier (binary sentiment detection) based on pre-trained LASER embeddings on rich language data (English) and used the model to infer sentiment scores in other languages, for example Estonian. Although human language is a very complex phenomenon and transfer of learning in the multilingual semantic data space is far from perfect, our initial results have indeed been promising.

The above serves to show an example from the rapidly advancing field of NLP and its potential to ease a number of daily tasks. Swedbank’s Analytics & AI continues to monitor similar developments in other fields and where possible, leverages them for the benefit of our customers and colleagues.

Thanks to Mehrdad Mamaghani and Markus Reimegård.

--

--

Lehar Oha
Swedbank AI

Data Scientist at Analytics & AI @ Swedbank Group