TreeSeg: Hierarchical Topic Segmentation of Large Transcripts

• Official repo for "TreeSeg: Hierarchical Topic Segmentation of Large Transcripts"
• Paper: https://arxiv.org/abs/2407.12028

TreeSeg is an algorithm for segmenting large meetings transcripts in a hierarchical manner using embeddings and divisive clustering. It produces a binary tree of segments by recursively splitting segments into sub-segments. The approach is completely unsupervised. While this implementation is using OpenAI's embeddings model, any embeddings model can be used as an alternative.

This repo makes multiple contributions:

1. Code for extracting utterances and ground truth segmentation from ICSI and AMI.
2. TinyRec, a small scale dataset of self-recorded session complete with topic annotations at two levels.
3. Versions of two existing unsupervised meetings segmentation approaches (BertSeg and HyperSeg) adapted to the hierarchical segmentation setting.
4. The original, efficient implementation of TreeSeg, a novel unsupervised, intrinsically hierarchical topic segmentation approach for large meeting transcripts.
5. Code for reproducing the experiments in the TreeSeg paper.

Note: You will need an OpenAI API key to run some of these experiments, since TreeSeg and BertSeg require access to an embeddings model. Take a look at the "running" instructions below.

ICSI and AMI datasets:

AMI:

• Download the AMI corpus from here: https://groups.inf.ed.ac.uk/ami/AMICorpusAnnotations/ami_public_manual_1.6.2.zip
• Create a folder named AMI in the datasets dir (datasets/AMI/) and extract it there.

ICSI:

• Download the ICSI corpus from here: https://groups.inf.ed.ac.uk/ami/ICSICorpusAnnotations/ICSI_plus_NXT.zip
• Create a folder named ICSI in the datasets dir (datasets/ICSI/) and extract it there.

The two datasets are available via the classes ICSIDataset and AMIDataset

Parsing these datasets is challenging. We did our best to extract utterances faithfully.

Baselines

Both baselines are adapted to receive a desired number of segments K. The top K-1 splitting points are used to segment the meeting into K segments.

Unsupervised Topic Segmentation of Meetings with BERT Embeddings

• Paper: https://arxiv.org/abs/2106.12978
• Original repo: https://github.com/gdamaskinos/unsupervised_topic_segmentation

The code is adapted from the original repo and differs in two major ways:

• It uses OpenAI instead of Roberta embeddings.
• In the original repo the ICSI/AMI datasets were assumed to reside in some database and were thus unavailable. We have adapted the code to use our version of ICSI and AMI as described above.

Unsupervised Dialogue Topic Segmentation in Hyperdimensional Space (Interspeech 2023)

• Paper: https://arxiv.org/abs/2308.10464
• Original repo: https://github.com/seongminp/hyperseg

The original repo did not contain extracted data for ICSI/AMI. We have adapted the code to our version of ICSI and AMI as described above.

Random

A naive stochastic baseline choosing transition point at random.

Equidistant

A naive determinstic baseline choosing equidistant transition points.

Running

• The entry point is main.py

• You can invoke it as follows python3 main.py --model MODEL_NAME --dataset DATASET_NAME --mid MEETING_ID --fold {dev,test}

• Alternatively instead of providing a meeting id yo ucan set the --eval flag which will process all meetings in the fold.

• model: treeseg TreeSeg, bertseg (BERT embeddings), hyperseg (hyperdimensional vectors), random (random segmentation), equi (equidistant segments), view (will print meeting in-order, segment by segment)

• dataset: icsi or ami

• mid: 0-indexed meeting id as an integer (e.g. 1 will evaluate on the second meeting in the corpus)

• For bertseg you will need to set up an environment variable named OPENAI_API_KEY with your OpenAI key, so that the embeddings can be extracted. Calls to the embeddings endpoint are inexpensive. We are using OpenAI's v2 (ADA) model. For more information on pricing, take a look here: https://openai.com/api/pricing/

• The original segment transitions are pruned so that segments have a minimum size of 10 utterances. In the end a plot is displayed comparing the original, pruned and inferred segment transitions points.