IMD business school for management and leadership courses
A.I. Text Analysis
How does artificial intelligence help us understand companies better than ever before?
Our NLP-based analysis investigates a company’s outlook to determine its readiness for the future.
We have created an algorithm that analyzes more than 70 public news sources from 2010 onward. These include BBC, CNN, New York Times, corporate press releases and various trade magazines. The workflow we created specifically for this purpose fetches media sources through an API (Application Programming Interface). We brought in hundreds of thousands of articles.
Our goal of this is to quantify analyst description of companies’ strategic aspect as well as quantify companies’ self-description of their strategic aspect. This approach is akin to sentiment analysis, but instead of determining discrete outputs (positive, neutral or negative sentiment) we derive continuous measures of strategic thematics, commonly referred to as constructs in the academic literature.
In order to attain high accuracy in our findings, we are leveraging the state-of-the-art transformer architecture that Vaswani et al. first released in 2017. Models derived from this architecture can be used for various natural language processing tasks.
Here is the entire process: To start, we pre-process news articles by feeding them into a pretrained transformer model. With this model, we can convert anaphoric expressions into company names. In simple terms, we replace any ambiguous pronouns in articles with the appropriate company names. This helps us to focus only on the text that is relevant to the companies in question.
Second, we utilize the transformer model to identify negation in the sentence. This will ultimately prevent us from over-inflating our results. As an example, “company A is not moving fast enough.” The negation, or rather, the word “not”, needs to be classified as a negative against our construct “fast-moving”.
To extract quantifiable insights on strategic topics such as digital orientation, decision speed, etc., we derive continuous measures based on relative counts of the words defining them. The composition of these subjects is established using academic peer-reviewed definitions. Our approach follows a long history of research from the academic management literature. In our earlier example, the construct was “fast-moving”.
What’s new with our A.I. approach is that we can unpack and analyze more data than ever before. We compare the scores of different companies in each calendar quarter to ensure a fair baseline, as all companies face similar circumstances during any given point in time.
Questions we asked in our research include:
- Which companies learn more quickly and efficiently?
- Which companies focus on opportunities with immediate benefits rather than developing new capabilities that may not have an immediate payoff?
- Which companies are more committed to adopting digital trends?
Looking at the data, we can see that there are companies who consistently perform better than their peers over long periods of time. We also see quarters where some companies move up in the rankings while others move down. Such observations help us understand why and how some companies are better prepared for the future than others.