Session Outline

Our guards write reports on everything that happens — from criminal events like burglary with damages or threats to people to warm freezers, water leaks and issuing parking tickets — generating tens of millions of reports globally every year. The guards categorize the event using preset categories and also write a free text comment to provide detail of the event in their local language. Freetext is information dense but challenging to automate information extracting from, especially on multilingual data. To enrich our guard reports to maximize client value, we use a Transformer model (XLM-R), which has been pre-trained on 100+ languages, and fine tune it to better understand guard terminology and expression. By adding another neural network, we can train the model to classify the reports for a number of interesting categories. Despite only fine-tuning training the model in a few languages, the model can classify well in many other languages, meaning that the pre-training multilingual properties carries over to our real-world task. To maximize the model impact with minimized effort, we developed a methodology of creating labels in a collaborative fashion and training the model which we will unpack. We will also share our learnings which we will use in other projects utilizing pre-trained language models. We have productionized these models in multiple countries, performing several tasks

Key Takeaways

  •  We employ a pre-trained Transformer language model to solve a multitude of real problems greatly outperforming classical methodologies in both performance and value scalability and greatly outperforms LLMs in speed and cost
  • Transformer language model can be fine tuned to understand domain specific data and their multilingual capabilities from pre-training carries through to the task despite single language fine-tuning
  • We need a relatively small amount of labels to reach a high classification performance by developing an efficient way of generating labels in a collaborative manner
  • Iterating over label generation and model analysis will ensure the most time efficient model development, with a minimum amount of labels and maximum generalization performance


Speaker Bio

Jonas Alström Mortin-Expert Data Scientist | Securitas Digital

Jonas works as an Expert Data Scientist in NLP at Securitas. He has almost 15 years of experience in data science and business intelligence, across academic research (physics, atmospheric sciences), consulting, and product companies in the security industry. Having previously lead/contributed as a data scientist in projects like “Crime and Risk prediction” and “Real alarm classification”, Jonas has in the last couple of years specialized in working with text data using methods that span very simple to latest gen LLMs. Jonas holds a PhD in Atmospheric Sciences.

October 25 @ 17:10
17:10 — 17:40 (30′)

Day 1 | 25 Oct 2023 | MACHINE LEARNING + MLOPS

Jonas Alström Mortin-Expert Data Scientist | Securitas Digital