AI helps to detect politicians misusing public money

Irina Kolesnikova
April 5th, 2022

Hidden political advertising detected by AI

Riigikontroll (The National Audit Office) is an independent institution acting in the interests of the Estonian taxpayers. It investigates and reports how the government spends the taxpayer’s money.

regional fund logo

Among other supervisory activities, its personnel monitors publicly funded channels such as local and state-level sites, publications, newsletters, and contracts, both on- and offline, to find illegally funded promotions of politicians or parties.

Their partnership with MindTitan seeks to improve the processing of all publicly available information channels to find potential red flags of misused public money.

The project is funded by the European Regional Development Fund.

The problem: Too many channels to monitor

Due to cases of illegal expenditure of public money for political promotion, there is a need to supervise the financial processes to detect any instances of splurging.

There are almost 100 types of information sources connected to the spending of public money, e.g., news, contracts, and others. Checking local governments and ministries’ websites, local newspapers, radio and TV programs, and official social media accounts manually requires many work hours.

It occurs not only for flagging the direct appearance of exact words or faces that indicate unwanted promotion at the taxpayers’ expenses to be checked but also for detecting red flags in other areas, such as suspicious contract activities. The volumes of data for supervision grow exponentially during elections.

 

 

In order to improve efficiency, MindTitan is developing an AI to detect illegal usage of public money to promote political individuals or parties. The main part of the solution is natural language processing models that will sift names, promises, and sentiment context, alongside the image analysis model, spotting spikes in publicity appearance frequency.

The expected outcome

The AI algorithms highlight and visualize the potential misuse of public money, to be checked manually by a human agent if needed. The application displays potential candidates selected and classified by AI after sifting and processing texts and pictures from multiple sources.

AI, which has made continuous monitoring possible, not only collects data and aggregates it but also analyzes it and provides a risk assessment if some model is triggered by specific irregularities. These risks are then also aggregated and shown to the user. The major benefit of AI is performing a risk assessment more efficiently with less human effort.

The scraper component

 The scraper component allows the system to fetch data from multiple publicly available sources:

 

The component collects all publicly available textual data and media as well as provides it for analysis to the machine learning components. It is the source of most system inputs (election dates, candidates, and reference photos of politicians are uploaded manually).

The classification models

The natural language processing and image recognizing models of this project work with free texts either scraped from open public digital sources and channels or provided transcriptions of public-funded radio and TV programs prepared beforehand.  

First, the text goes through name entity recognition models. In all the channels provided, the models find any entity (person, agency, party, etc.) mentioned in the text.

The name entity recognition simplified

As many people are mentioned on open public sources, to reduce false positives we filter the NER detections with a manually uploaded list of people of interest, including people currently performing public duties (and hence with access to public money) and candidates (who can be mentioned by government members) in a later step.

The models find the exact names of politicians or parties uploaded to the system manually as references in all the channels provided.

Monitoring those channels regularly, the AI algorithm defines the trend in terms of its’ appearance frequency and, again, if there is a crest in the trend, the system highlights it as suspicious activity.  

The promise detection and sentiment analysis models simplified.

Afterward, the texts go into the promise detection and sentiment analysis models simultaneously.

The promise detection model recognizes both future and past statements of achievements. This is important during election periods, as it is vital to keep those mentions supervised thereby preventing a concealed promotion paid with public money.

The sentiment analysis model allows checking the emotional background and context of politicians’ or parties’ mentions. To keep a clear and honest depiction of the political field for Estonian taxpayers, it is necessary to halt insults or excessive unreasonable praise if there are any.

All the data delivered from AI algorithms go to the database and is kept there until deemed useful by the NAO auditors, although regular erasing cycles are foreseen. An auditor from The National Audit Office can access the available data and trends anytime from a web browser application and supervise all the activity of the political field in publicly-funded channels, such as news and open contracts published on official websites as well.

MindTitan team discussing AI strategy

Conclusion

After the implementation, the scraper component and classification models will automatically sift a lot of sources to find red flags for suspicious financial activity.

The AI-powered system for Riigikontroll is expected to save a significant amount of time for auditors and offer higher supervisory efficiency with less human effort for the sake of the impartial, transparent and trustworthy political field for Estonia.

Go back