This code can be used to preprocess AMI dataset for meeting summarization from scratch.

• Download files from ACL18-code.
• Copy files dascim-acl2018_abssumm-3ee827b34472/data/meeting/ami/* to ./data/acl18-ami/*.
• Copy files dascim-acl2018_abssumm-3ee827b34472/data/meeting/lists/list.ami, list.ami.train, list.ami.eval, list.ami.test to **./data/**.

Now, you will have data/acl18-ami/*, data/list.ami, data/list.ami.train, data/list.ami.eval and data/list.ami.test

• Download the AMI dataset from AMI manual annotations v1.6.2 and unzip it to ./data/. Finally you can get ./data/ami_public_manual_1.6.2/*, which contains original AMI files.
• Run python delete_useless_files.py

python delete_useless_files.py: Following previous setting, we only use part of the meeting files, so we first filter out useless files. This script is used to extract words, topics and summaries according to the useful file list ./data/list.ami. Output files are in data/ami/words, data/ami/summary, data/ami/topics

• Run python ami_preprocess.py, results are shown in data/ami/train, data/ami/valid, data/ami/test

python ami_preprocess.py will first create dirs train,valid,test under data/ami/ and then extract speaker, time, utterance, topic from data/ami/words and data/ami/topics, in each line, the data format is speaker \t time \t utterance \t topic.

• First copy stopwords from dascim-acl2018_abssumm-3ee827b34472/resources/stopwords/meeting/filler_words.en.txt to data/ami/stop_words.txt. Note that you should rename the file.
• Run python create_clean_data.py.

create_clean_data.py will first creat output dirs cleaned, cleaned/train, cleaned/valid, cleaned/test at data/ami/, then clean utterances.

• Run python get_summary.py.

python get_summary.py will create a output dir data/ami/final_summary and extract the summary from xml files for each meeting.

Now, you have meetings in data/ami/train, data/ami/valid, data/ami/test and also summaries in data/ami/final_summary.