Abstract

Challenges of Natural Language Processing techniques in official statistics

Session organisers: Michael Reusens (michael.reusens@vlaanderen.be); Marc Callens (marc.callens@vlaanderen.be)

This session invites practitioners and researchers to present research on the challenges related to the innovative application of natural language processing (NLP) techniques to produce official statistics.
Official statistics are traditionally produced using structured data, often produced by conducting surveys. In the past decades, these have been complemented with register and administrative data sources. More recently by applying innovative data science techniques (e.g., NLP) on unstructured data (e.g., text) statistical offices are creating new, mostly experimental, statistics. There is more data available, and previously hard-to-analyze unstructured texts (unstructured survey results, web pages, social media, …) now become usable with the use of NLP (amongst other techniques) in a timely and frequent way
Furthermore, the field of NLP is rapidly evolving resulting in ever-increasing opportunities to extract information from unstructured texts.
However, given the novelty of many NLP methods and underlying data sources, statistical institutes must be careful to ensure the quality of resulting statistics and hence need to invest in more systematic knowledge about their quality.

Topics of interest in this session are, but are not limited to:

Quality frameworks and quality metrics for NLP applications in official statistics
Applications of NLP on new data sources: social media data, web scraped data, ...
Applications of Supervised classification of texts: sentiment analysis, automatic categorization of companies, product classification, ...
Applications of Unsupervised knowledge extraction from texts
Visualization of textual data

The following speakers will shed light on these topics by presenting their work on applying NLP in official statistics:

Statistics Canada (Roshanafshar Shirin), on “Classifying Respondent Comments from Canadian Census of Population”
Central Bureau of Statistics, Netherlands (Piet Daas), on “Categorizing company websites”
INEGI, Mexico (Jael Perez), on "Methodological proposal to codify records of occupation and economic activity of the National Survey of Household Income and Expenses (ENIGH), using Deep Learning".
Statistics Flanders, Belgium (Michael Reusens, Cedric De Boom, Marc Callens), on “Challenges on using Twitter sentiment in official statistics”
Statistics Poland (Dominik Dabrowski), on “Extracting meaningful information from web data on real estate – challenges and experiences from the Web Intelligence Network”