evamaxfield / cue-queue Goto Github PK

Ideas:

• NPR Media Dialog Transcripts
• TedTalks dataset

I can't find any instructions on how to use this. From looking at the docs, I assume it does not output segments of a text. Does it?

Instead of choosing N most similar to average cue sentence, try choosing the local maximas with scipy.signal.find_peaks

Describe the Bug

$ pip install git+https://github.com/JacksonMaxfield/cue-queue.git
Collecting git+https://github.com/JacksonMaxfield/cue-queue.git
  Cloning https://github.com/JacksonMaxfield/cue-queue.git to /tmp/pip-req-build-zzf9yt5l
  Running command git clone -q https://github.com/JacksonMaxfield/cue-queue.git /tmp/pip-req-build-zzf9yt5l
  Resolved https://github.com/JacksonMaxfield/cue-queue.git to commit b10d94759c9a8dd6cafd714297f66cc52a4cd571
ERROR: Could not find a version that satisfies the requirement cdp-backend>=3.0.0.dev18 (from cue-queue) (from versions: none)
ERROR: No matching distribution found for cdp-backend>=3.0.0.dev18

Expected Behavior

Install without error

Reproduction

Steps to reproduce the behavior and/or a minimal example that exhibits the behavior.

Environment

Any additional information about your environment.

• OS Version: Ubuntu 20.04

SliceCast seems to be exactly what we are looking for, or at least a starting point to generalize further / stricter engineering of the codebase: https://github.com/bmmidei/SliceCast

https://arxiv.org/abs/2106.12978

https://dl.acm.org/doi/abs/10.1145/3331184.3331208?casa_token=gmvvT7Xf8NgAAAAA:tzJ353rIAbtuSrWlItRMw2vrJX3d6tfCB8bn0q7M_2ayI__xLijQEMeRkFodOjltIEPF5UoZq4H6

Discourse Segmentation of Multi-Party Conversation

https://aclanthology.org/P03-1071.pdf

Cite: https://doi.org/10.1162/tacl_a_00261

https://asistdl.onlinelibrary.wiley.com/doi/10.1002/asi.24496

https://terpconnect.umd.edu/~oard/pdf/acl20-barrow.pdf

Building off of: https://aclanthology.org/2020.lrec-1.829.pdf

Use semantic similarity between each minutes item to a sentence in transcript, then using the trained model for segmentation classification, classify sentences that are breaking. Find the nearest segmentation markings to the most similar breaking sentences.

i.e.

part 1:

s1, s2, s3, sn...
     ^ most similar to t1

part 2:

s1, s2, s3, sn...
     ^ most similar to t1
^a       ^b

where a and b are delimiter sentences (topic breaks)

https://arxiv.org/abs/1803.09337

Also attempt to train and use the model for segmentation

https://segeval.readthedocs.io/en/latest/api/#segeval.boundary_similarity

input is [len_sec_0, len_sec_1, len_sec_n, ...] i.e. [10, 52, 131, 281, 5, 12]

citation: https://aclanthology.org/P13-1167.pdf

Build a classifier for if something is a cue block or if something is a discussion block.

Can test different size block sizes (moving window) from 1, 2, 3, 4, 5, 10, sentences etc.
After finding the block size classifier that performs best, use the trained classifier to generate a signal that is 0 for discussion blocks and 1 for cue blocks.

So a transcript's generated sequence from the classifier may look something like the bottom sequence in the image. (1, 0, 0, 0, 1, 1, 0, 0 ,0, 0, 1, 1, 0, 1)

The top sequence is created by assuming that there will always be an "intro cue", some discussion, and then an "outro cue". So generate this sequence as (1, 0, 1) * M where M is the number of minutes items. I.e. for three minutes items the generated sequence is (1, 0, 1, 1, 0, 1, 1, 0, 1).

Finally perform dynamic time warping / sequence alignment on these two sequences to find best path.

Eval overal performance with PK / WindowDiff.