IOCC: Aligning Semantic and Cluster Centers for Few-shot Short Text Clustering

Imagine you are trying to sort a huge pile of mixed-up postcards into different boxes. Each postcard has a very short message written on it (like "Buy milk" or "Meeting at 5"). Your goal is to group similar messages together perfectly.

The problem, as the paper explains, is that these short messages are like tiny, blurry snapshots. Because they are so short, it's hard to tell exactly what they mean just by looking at them. Traditional sorting methods often get confused, placing a postcard in the wrong box because they can't find the "true heart" of what that group of messages is really about. They end up creating messy piles where the center of the group doesn't actually represent the main idea.

The Solution: IOCC
The authors propose a new method called IOCC to fix this. Think of IOCC as a smart, two-step sorting machine designed specifically for these tricky, short notes. Its main goal is to make sure the "center" of each pile (the average idea of the group) perfectly matches the "true meaning" of the notes inside it.

Here is how the two main parts of IOCC work, using a simple analogy:

1. The "Smart Matchmaker" (Interaction-enhanced Optimal Transport)

Imagine you have a bunch of students (the text samples) and you need to assign them to study groups.

• The Old Way: You just guess which group they belong to based on a quick glance.
• The IOCC Way (IEOT): This module acts like a super-smart matchmaker. Instead of just looking at one student in isolation, it looks at how the students interact with each other. It asks, "If Student A is talking to Student B, and Student B is talking to Student C, do they all belong in the same circle?"
• By analyzing these connections, it creates a "draft list" (pseudo-labels) of who belongs where. It then picks the most confident students from these drafts to build a "prototype" or a Pseudo-Center for each group. Think of this as finding the "ideal student" that represents the best version of that study group.

2. The "Magnet Trainer" (Center-aware Contrastive Learning)

Now that you have these "ideal prototypes" (the Pseudo-Centers), the second module kicks in.

• Imagine each text note is a small metal ball, and the Pseudo-Center is a powerful magnet.
• CACL acts as a trainer that pulls the metal balls (the text representations) closer to their matching magnets.
• As the training continues, the notes that belong together get pulled tighter into their specific group, while notes from other groups are pushed away.

The Result: A Perfect Dance

The magic happens because these two modules work together like a dance partner.

1. The "Smart Matchmaker" suggests where the groups should be.
2. The "Magnet Trainer" pulls the notes into those spots.
3. As the notes move, the "Matchmaker" gets a better look and refines the group centers again.

This back-and-forth loop gradually shrinks the gap between where the groups are and where they should be (the true semantic meaning). Eventually, the piles become incredibly clear and well-separated.

The Proof

The authors tested this system on eight different sets of data (like sorting medical notes, news headlines, and more). They found that IOCC was much better at sorting these short texts than previous methods.

• On a particularly difficult set of medical notes (the Biomedical dataset), it improved accuracy by 7.34%.
• It was also more stable (less likely to make random mistakes) and efficient.

In short, IOCC solves the problem of "blurry" short texts by constantly aligning the physical grouping of the data with the true meaning of the words, resulting in much cleaner and more accurate clusters.