Imagine that if you are an advertiser or a content creator.

You want to know what the popular stuff people are talking about and people's attitudes towards them so that you can design your product, ads or content accordingly.

What will you do?

Our models can help!

In our project, we use Reddit Singapore as the platform to test our solutions.

Here we have our problem statement.

We build models to detect hot trends on Reddit Singapore and understand people's attitudes towards the trending topics by doing sentiment analysis on the comments.

The flow on the left is the overall structure of our solutions.

First, we input the posts from Reddit into the topic modeling model. The output is the hot topics detected from the posts together with the respective topic descriptions. 

Then the comments are grouped by the topic they belong to and taken as the input of the comment analysis model. The model extracts multiple objects being discussed in the comments and conducts sentiment analysis on them. 

Detect hot topics on Reddit Singapore. Take the posts as input and the model groups posts with similar content into clusters. Then we obtain the trending topics with keywords and popularity. 

The table on the left is a brief summary of the trending topics.

The "Topic" column is the index of each hot topic. Topic "-1" means the cluster of outliers. Topics are sorted by the number of posts on each topic. Then the "Name" column contains the top four keywords representing each topic.

The picture on the right illustrates one example of the hot topics with the representing keywords together with their word scores. The higher the word score is, the more likely the keyword can represent the topic.

There are three key elements.

Topic: The number of topics generated is automatically decided by the algorithm which assigns numbers to the topics starting from 0.  Here is one example.

Keywords: Words with the highest frequencies appeared in this topic, some topics can be directly inferred by them.

Word Score: Word frequency in the clustered posts.

The image on the left illustrates both the number of comments and the number of posts on each topic.

The popularity of each topic can be found by the total number of comments under the posts on this topic. 

Analyze people's attitudes towards the objects being discussed under each topic, these objects are called aspects. Take the comments as input and output the aspects with sentiments. Use visualization to explain people's opinions based on the comments.

After detecting hot topics, we analyzed the sentiments of the comments under topics to get to know about people's attitudes. We used a span-based aspect sentiment triplet extraction (Span-ASTE) model to identify the object being discussed (aspect), its description and the sentiment of the description in a comment sentence.

On the right hand is an example of ASTE.

What's unique?

span-level information:

         Extract a phrase instead of a single word.

         i.e., not enjoy rather than not & enjoy.

aspect-description match:

         Identify each aspect and its corresponding description.

Colour: represents sentiment towards the aspect. The warm colour indicates positive sentiment. The cold colour indicates negative sentiment.

Size: represents word frequency, which is the popularity of the aspect.

X-axis: represents the time when the aspect started to appear in the comments. 

Y-axis: represents the probability of how likely an aspect being a ‘topic’. The aspects with higher scores are more likely to be topics, and aspects with lower scores are less important.

The final delivery of this project mainly consisted of the above two parts: topic modelling and comment analysis. We detect hot trends on Reddit Singapore and find the sentiments in the comments.

Our models can contribute to extracting information from social media platforms with little cost.

1. facilitate precise advertising by identifying target customer groups

2. Adjust inventory for retail stores

3. Facilitate product design by trending topics detection

4. Provide inspiration for content creators

5. Collect feedback from customers to improve the service level