Aspect Based Sentiment Analysis
A look at the murky waters of sentiment analysis
This is where ABSA (Aspect-Based Sentiment Analysis) comes into play!
Aspect-Based Sentiment Analysis (ABSA) is a facet of sentiment analysis that dissects textual data into different aspects, or components, and determines the sentiment associated with each aspect individually. Rather than categorizing sentiment on a document-level, it does so based on these determined aspects.
ABSA offers a much deeper and more granular understanding of sentiment by breaking it down into specific aspects, such as features, products, and brands, rather than treating the entire text as a whole.
ABSA does not and should not replace traditional sentiment analysis; rather, it complements it by addressing specific business questions that require a more granular perspective on sentiment in online conversations.
How does it work?
Let’s assume we have 100 social media posts, each one labelled as either Positive, Negative, or Neutral
We could visualise these posts in a waffle chart, where each square represents a post and the colour represents document level sentiment
Or…
We could use ABSA to look at the specific sentiment associated to entities (i.e. brands or products) of interest.
To do this, let’s change the colour for entity sentiment, so it is clear when we are looking at document-level sentiment vs entity-level sentiment.
Now we have Positive, Negative, or Neutral
By looking at this more nuanced sentiment, we can see that actually each of these posts could possibly contain different sentiment for different entities
What does this mean?
Let’s look at our document-level posts again
And more specifically this example post
Where the traditional sentiment
matches the ABSA sentiment (remember, in this document we have document-level positive and entity-level positive
meaning that we would have inferred the same result if we did document-level or entity level sentiment analysis
However, there are other cases where we would get different results based on whether we look at document-level or entity level sentiment.
For example, these two posts are negative at a document-level
but the specific entities are positive (i.e., “I’ve had it with technology today. My computer crashed again, but credit to Microsoft’s customer service - they were helpful!”).
The true power of ABSA lies in its ability to pinpoint specific entities relevant to the research project at hand
For instance, by tallying the frequency of each entity mentioned in both positive and negative contexts, we can discern how these are referenced at a much finer level of granularity.
In this sample of 100 posts, Microsoft is mentioned positively 34 times…
… and negatively 8 times
While the higher frequency of Microsoft mentions is expected due to the nature of the data (the original data is from a very MSFT specific project), what provides value is the realisation that Microsoft is mentioned in a positive sense over x4 more frequently than a negative one, in contrast to brands like OpenAI which are mentioned positively only x2 as often as negatively.
Finally we can compare this result to what we would have assumed if we had just looked at document level sentiment for mentions of Microsoft
So whereas at a document-level we might assume Microsoft is mentioned ~2x as much in a positive light than a negative light, in reality when Microsoft is mentioned, 71% of the time it is mentioned positively
Finally, we can expand this to other entities too.
For example, this shows us the number of positive and negative mentions at a document-level for some potentially important entities.
We could infer from this (albeit with minimal data), that there are more positive mentions of ChatGPT than negative
Whereas when we look at this from the view of specific entity aspects, we can see that in reality the specific mentions of ChatGPT skew negative
| Brand or Product | # of Positive Mentions |
|---|---|
| Microsoft | 34 |
| OpenAI | 8 |
| ChatGPT | 6 |
| Bing | 5 |
| Azure | 3 |
| Brand or Product | # of Positive Negative |
|---|---|
| Microsoft | 8 |
| 4 | |
| OpenAI | 4 |
| GPT | 3 |
| ChatGPT | 2 |
| Brand or Product | # of Positive Mentions |
|---|---|
| Microsoft | 34 |
| OpenAI | 8 |
| ChatGPT | 6 |
| Bing | 5 |
| Azure | 3 |
| Brand or Product | # of Positive Negative |
|---|---|
| Microsoft | 8 |
| 4 | |
| OpenAI | 4 |
| GPT | 3 |
| ChatGPT | 2 |
