Attention Metric Approaches

Here's how our attention scores can accurately capture the recognition a topic receives in the media.

2 Attention Metric

2.1 Definition

For each processed news story, our topic model generates a vector of 200 probabilities. These probabilities represent the proportion of text associated with each topic that the model detects. Based on these values, we calculate an attention metric to measure the popularity of a topic at a specific granularity level and/or geographical zone. Currently, we provide attention values daily and monthly for the US, with the possibility of extending this coverage to other time zones through a prior agreement.

To minimise errors during the inference process, we consider only news stories that contain at least 100 words and have topics that capture at least 10% of the content. This helps ensure the accuracy of the scores.

The daily attention score is calculated by summing the probabilities associated with a specific topic throughout the day. The aggregated value is normalized by considering the count of all the processed news stories on the same day. The attention is a value ranging in the interval [0, 1], which facilitate the interpretation of the metric.

2.2 Attention metric and historical events

Our objective is to validate whether the attention score assigned to each topic accurately captures the recognition that a specific topic receives in the media.

In other words, we aim to visualise whether there is a potential change in the trend of attention over time when significant events related to a given topic occur.

In both charts, Figures 3 and 4 (next page), the displayed series represent the centroids of the cluster (the mean attention of all topics included in the cluster), using a monthly granularity.

To provide additional clarity in the analysis, we have included periods of the US elections pre-campaigns since the attention towards certain topics may be influenced by the interest they generate during these campaigns. >

Figure 3 illustrates the attention received by Cluster 42 from 2015 to 2023. This cluster is related to Crude Oil and groups topics such as Crude Oil, Oil & Gas, Mining, Iron Ore Mining, and Oil & Gas exploration. Bear in mind that the clusters were named manually, using word clouds displaying the most relevant terms for each topic. It appears that the attention towards the Crude Oil cluster is affected by the occurrence of wars, natural disasters, or events that directly impact crude oil prices.

Figure 4 represents the same information but focuses on Cluster 30 related to US Defense. This cluster encompasses topics such as Syria, Asia Politics, Terrorism, NATO, Defense, Russia, the Middle East, Iranian Sanctions, North Korea, and Afghanistan and Iraq wars.