public_meetings: a meeting corpus of aligned pairs of transcriptions and reports.

• Paper: Align then Summarize: Automatic Alignment Methods for Summarization Corpus Creation.
• Code: https://github.com/pltrdy/autoalign
• Cite the paper:

@inproceedings{Tardy2020,
author = {Tardy, Paul and Janiszek, David and Est{\`{e}}ve, Yannick and Nguyen, Vincent},
booktitle = {Proceedings of the 12th Conference on Language Resources and Evaluation (LREC 2020)}, 
keywords = {Alignment,Corpus Annotation,Summarization},
pages = {6718--6724},
title = {{Align then Summarize: Automatic Alignment Methods for Summarization Corpus Creation}},
year = {2020}
}

Getting started

• from pip:

pip install public_meetings

• from sources:

git clone https://github.com/pltrdy/public_meetings.git
cd public_meetings
pip install .

About

This corpus contains meetings, made of pairs of (a) automatic transcriptions from audio recordings, (b) meeting reports written by a professional.
Both texts are way too long to be reasonibly processed (e.g. by neural models) so we worked on the automatic segmentation and alignment to get suitable pairs for meeting summarization training/evaluation.

We present a public extract of our data in this repository. The segmentation/alignment can be found at https://github.com/pltrdy/autoalign.

Reading the data

We provide 22 aligned meetings that can be loaded easily:

import public_meetings

meetings = public_meetings.load_meetings()

Meetings are identified by a hash e.g.:

meeting = meetings['81540075987931464031780e046c0d8f']

Each meetings has been automatically aligned first meeting['initial'] then post-edited by a human annotator, meeting['final']. Each alignment has a transcription (aka. ctm) and a report side (aka. doc) that contains segments (usually several sentences).

meeting['final']['doc'][i]['text']      # text of the i-th document segment
meeting['final']['doc'][i]['id']        # id of the i-th document segment

meeting['final']['ctm'][j]['text']      # text of the j-th transcription segment
meeting['final']['ctm'][j]['id']        # id of the j-th transcription segment
meeting['final']['ctm'][j]['aligned']   # doc segment id corresponding to the j-th transcription segment

Commands

public_meetings provides utility functions from a single entry point:

public_meetings [command]

Commands are listed in the following sections.

prepare: process all meetings to src/tgt files.

The prepare command is meant to prepare meetings for summarization models (either for training or inference).
It basically load every meetings and write the transcription side in a [prefix].src.txt file and the report side in a [prefix].tgt.txt. Many parameters can be set to filter segments, on their number of words/sentences (both min and max values).

Example from the paper:

./prepare.py \
    -mw 10 -Mw 1000 \
    -ms 3 -Ms 50 \
    -overlap_prct 0 -n_draw 0 \
    -remove_unk \
    -sentence_tag \
    -remove_names \
    -remove_headers \
    -remove_p

full usage:

public_meetings prepare -h
usage: prepare [-h] [-dir DIR] [-mw MW] [-Mw MW] [-ms MS] [-Ms MS]
               [-remove_tags] [-remove_unks] [-remove_names] [-remove_headers]
               [-remove_p] [-sentence_tags] [-overlap_prct OVERLAP_PRCT]
               [-n_draw N_DRAW] [-output OUTPUT] [-verbose]

optional arguments:
  -h, --help            show this help message and exit
  -dir DIR, -d DIR      Aligned meeting root
  -mw MW                Min #words
  -Mw MW                Max #words
  -ms MS                Min #sentences
  -Ms MS                Max #sentences
  -remove_tags          Remove every tags i.e. <*>
  -remove_unks          Remove unknown tags i.e. <unk>
  -remove_names         Remove names i.e. <nom>*</nom>
  -remove_headers       Remove headers i.e. <h>*</h>
  -remove_p             Remove paragraph tags i.e. <p> and </p>
  -sentence_tags        And sentence separators <t> and </t>
  -overlap_prct OVERLAP_PRCT, -oprct OVERLAP_PRCT
  -n_draw N_DRAW
  -output OUTPUT        Output path prefix
  -verbose, -v

segmentation: process the transcription side in a linear segmentation fashion.

We use this before running linear segmentation experiments. It only considers transcription side of meetings, and write it to source (one segment per line) and reference (one segment per line + segmentation separator ==========).

You just have to set an output_root directory to recieve the text files, and, optionnally a different meeting_root

Example:

public_meetings segmentation -o ./public_meetings_txt

Full usage:

public_meetings segmentation -h
usage: segmentation [-h] [-meeting_root MEETING_ROOT] -output_root OUTPUT_ROOT

optional arguments:
  -h, --help            show this help message and exit
  -meeting_root MEETING_ROOT, -m MEETING_ROOT
                        Meeting root directory
  -output_root OUTPUT_ROOT, -o OUTPUT_ROOT
                        Output root directory