1. New NACE classifications
Background: The industry classification of establishments is an important variable for the majority of SSB’s business statistics. A new NACE classification standard is being introduced soon, (potentially from 2025) for industry activity. This introduces several challenges but also provides opportunities to access the way in which we approach classification tasks. There are several areas where ML/LLM may be able contribute to improved quality and provide more efficient processes:
- Automating classification of NACE based on descriptive texts and other sources.
- Using ML models to detect incorrect classifications.
- Aiding and automating the transition to new NACE classifications.
The emergence and popularity of LLMs in the past few years provides opportunities to perform these types of tasks in a novel way. There is potential for finetuning these models in a way to perform predictive classification tasks with much less training data than more traditional supervised learning methods.
Project suggestion: Look at ways to use LLMs for predicting new NACE codes (or using current NACE as a backup depending on timing/access to new data)
2. Innovation prevalence among Norwegian companies
Background: There have been a number of studies in other countries on predicting innovative companies based on data from their website. This has not been tested in Norway to our knowledge. Norway is a highly digital society and the coverage of websites for companies is likely to be high, making Norway a good candidate for this type of data collection. Many have used DNN for prediction after data extraction but newer LLMs may be a good approach for this task.
Source: Maria, L. E. (2019). Case Study of Big Data Methods for Innovation Detection in the Netherlands
Project suggestion: Use data from webscraping of company websites combined with text mining and prediction algorithms to predict innovation prevalence in Norway. Test different models and the use of LLMs in this context.
3. Extracting data from pdf/tiff documents
Background: Manual tasks to find specific data in tiff documents is common in SSB. This is particularly common with economic statistics needing additional information from annual accounts, which have previously only been available in picture tiff format. A survey among the statistical division in April 2024 indicates that around 4.6 man-years is used on manually searching these files for information. SSB has recently gained access to these files in more computer readable formats, however the format is not standardized (many are pdf but also xls, doc, jpeg, png). NLP opens options for automating this process in a way to significantly reduce this time-consuming task.
In one such case study, around 1300 tiff picture files are manually scanned for information on pension costs. NLP could most likely help in this task of extracting the correct information.
Project suggestion: Investigate the use of NLP to extract specific information from pdf/and or picture files (tiff, jpeg, png). A specific case could be to extract pension costs from annual accounts.