Topic Change Detection

[vdpappu]

Let's discuss the potential next steps for improving topic change detection and update the activities here.

[reaganrewop]

The ideal goal for the topic change detection is to, slice the meeting into multiple partition where each partition carries enough information to redeem itself as a discussion.

The following needs to be addressed to achieve this:

• cosine similarity as solo edge weights.
• A mixture of topics in a single segment (not an ideal case.)
• what is the factors for grouping of segments (currently it's the order of the segments by which they were spoken at)
• pruning of the edges. How do we prune the respective irrelevant edge?
• filler sentences in the segment causing overlapping groups.

going with our current implementation, I made few extra implementations to try to fix the last issue.

• handling spillover sentences was rather important because it caused many overlapping groups. I am currently handling this by checking for duplicate segments across the communities and if found, I remove them if majority of the sentences from that segment are placed in a different community.

doing this increased the accuracy of the communities by a large amount and no overlapping groups would be formed.

[reaganrewop]

To improve the current communities approach (the one on staging) or to be precise, to understand what is best for communities, I went through few papers and methods to understand how effective it can be. Based on that I made few changes to the algorithm.

• Instead of fully connected network, we connect two sentences only if they are either from same segment or from the next. This helps to reduce cosine similarity noise.

• Normalizing the graph is now a bit different. we compute local normalization score for each node and then for the overlapping edge values, we average the score.

• community approach relies on self-loops, so that is also added.

• Based on this paper https://arxiv.org/pdf/0812.1770.pdf , we add another resolution parameter t, which helps to control the stability of the network.

Based on the validation set, the accuracy increased form 47 percent to 79 percent.