As part of the European project ACCSESS, Fraunhofer has investigated the perception of carbon capture, utilisation and storage (CCUS) technologies in Europe. In addition to a traditional survey, Fraunhofer conducted a social media analysis and a sentiment analysis. This was to ensure that the data was as representative of people’s true attitudes towards CCUS as possible.
Author: Alberto Sánchez (Fraunhofer IAO)
The involvement of the population is crucial to the success of CCUS in Europe
CCUS is a technology for capturing carbon dioxide emissions from industrial processes and either storing them underground to prevent their release into the atmosphere or providing them to industries that need carbon. It plays a crucial role in mitigating climate change by reducing greenhouse gas emissions, particularly in industries where emissions are challenging to get rid of.
An important part of environmental science research projects nowadays is to find out the current state of public perception of CCUS. The actual process that is being carried out in Europe to safely develop carbon capture technologies, and thereby reduce the environmental impact of many industrial activities, necessitates a social science perspective. This includes assessing the depth of general knowledge and identifying the main benefits and concerns of the technology. With this information, it is possible for researchers to develop educational materials specifically targeted at the general public, as the involvement of the population is crucial to the active and successful development of CCUS in Europe.
“Nobody knows what CCUS is”
During the Public Perception and Business Models Joint Event held on 14th November 2023 in Brussels, several European workgroups discussed the current state of public perception of CCUS, as well as the challenges and lessons learned during their processes.
The presentation by Fraunhofer and the European project ACCSESS about their work on public perception of CCUS began with this sentence: “CCUS provokes little discussion among the population. Its benefits and potential for tackling climate change are widely recognized, but its perception in social media has a clear predominant negative tendency”.
A key and recurring comment from many participants in the conference was on the fact that citizen surveys often fail to achieve their aim of getting enough respondents to ensure that the survey is truly representative of the population. To address this issue, the solution proposed and implemented by ACCSESS was to include extra methods of uncovering public perceptions in addition to the traditional survey: namely, a social media analysis and sentiment analysis. Both these analyses were first developed in the fields of computer science and linguistics and have been adopted in recent years by numerous distinct disciplines.
The social media analysis involves examining user-generated content from social media platforms, such as posts, comments and shares, to gain insights into public opinions, trends and behaviours. Social media analysis allows a greater number of opinions from different regions to be obtained in a faster and more effective way. Furthermore, it has been proven that social networks offer anonymity in many cases, allowing users to express their true opinions without fear of being judged by people or entities close to them.
The sentiment analysis focuses on identifying and categorising the sentiment expressed in social media content. Since users who write about CCUS do so without having a pre-defined survey that sets a path to follow when structuring how they express their opinions or the possible responses allowed, this method enables the collection of opinions that may not be covered by the original survey, but that can still be very useful and provide a lot of information. Using natural language processing and machine learning techniques, the sentiment conveyed in a post or comment is determined to be positive, negative or neutral. This enables researchers to gauge public attitudes and perceptions towards specific subjects, such as CCUS, providing valuable insights for decision-making and strategy development.
Avoiding bias: Revolutionising data collection for environmental sciences
At the time of defining the first steps and main components of the social media and sentiment analysis project, the lack or unusualness of applying these methods in the area of environmental sciences was noticed. This reaffirmed the need to introduce these methods to this research field.
Given that ACCSESS’ main survey was distributed to city partners and personal contacts of Fraunhofer, the results showed a high prevalence of respondents from academia and public administration. In addition, the number of responses only amounted to 124. It was considered that there was a high probability that people educated and active in the field of environmental sciences would have a more positive opinion of CCUS than people with less knowledge in this field. Avoiding bias was therefore another task delegated to the social media project.
For this project, the usual steps were followed:
- Development of annotation guidelines,
- Scraping of tweets,
- Manual annotation of tweets,
- Analysis of the results and training of a machine learning model.
First, a Twitter (now known as “X”) scraping program was developed using the official Twitter API and the Python programming language, as well as other necessary packages. Tweets containing information about CCUS were collected using a search query for all languages supported by Twitter, translated into English, and placed in an Excel .CSV-file with a special drop-down list to facilitate the annotation process.
Then, two independent analysts manually annotated the tweets, following some simple and short instructions. After the initial annotation process, the team responsible for this task in the project came together with the annotators and developed an official annotation guide for the whole process. With the guidelines in place, 3,374 tweets were scraped and annotated in two months. The annotations followed the structure of the survey, namely perceived CCUS benefits and concerns, with the overall sentiment of the tweet added at the end.
Once sufficient data had been collected over two months, the results were analysed and translated into a report with graphs and comparisons between the social media and the survey data. Findings from the final analysis include that more respondents from the survey had no opinion on CCUS than had an opinion, and there were significantly more negative opinions in the social media data, particularly viewing CCUS as a:
- scam,
- waste of time,
- green washing.
Process automation and lessons learned: maximising AI integration benefits
The final part of the procedure was to automate the whole process so that any person without much knowledge of the above-mentioned technologies could use it. For this purpose, an attempt was made to train a state-of-the-art BERT language model with the manually labelled tweets, so that it could automatically annotate tweets on the CCUS topic in the future. However, for that task, the tweets contained in the ACCSESS database were insufficient and the model did not perform well.
For research groups considering applying these methods, it should be noted that the training phase of such a model requires a lot of annotated data, which in turn means time, money and effort. It should also be kept in mind that tweets should be retrieved together with their respective tweet-IDs, as due to privacy policies, tweet texts cannot be published, only tweet-IDs.
By avoiding some errors, social media analysis and sentiment analysis can bring great complementary data and information to projects in environmental sciences and other fields. As AI applications are gaining more and more popularity within months, the possibility of implementing them in this process should be explored.
For tasks such as translating tweets, creating an annotation environment or retrieving tweets, traditional programming methods like the use of APIs, exporting data to Excel files or applying Boolean operators are sufficient. However, AI models, e.g. Large Language Models (LLMs) such as Open-AI’s GPT-4 or Meta’s LLaMA 2, can be helpful in processes like analysing language elements, creating search queries for the Twitter API, or annotating large amounts of tweets, which would allow a model to be trained.
Want to hear more from the ACCSESS project? Subscribe to our newsletter and follow us on LinkedIn to stay up to date on the latest news and developments.
