Interactive Spoken Content Retrieval

---

Getting Started

Run baseline result on dnn recognition results with jieba segmentation

Example data

• ./data/PTV.query.txt (query file, an unsegmented query on each line)
• ./data/PTV_transcription_charSeg (directory of dnn recognition transcripts)

1. Segment query and transcripts with jieba

# Segment query, retain line for each query
$python ./data/segment_words.py -i ./data/PTV.query.txt -o ./data/PTV.query.jieba.txt -r
# Segment transcripts, separate each word by line
$python ./data/segment_words.py -i ./data/PTV_transcription_charSeg -o ./transcripts/PTV.dnn.onebest.jieba

Now we have our jieba segmented query at './data/PTV.query.jieba.txt' and transcripts at './transcripts/PTV.dnn.onebest.jieba'

2. Run indexing with our indexer

The indexing script is located at './scripts/run_indexing_for_segmented_transcripts.py'

# Create vocabulary with query & transcript words,
# run indexing(background, doclengs, inverted_index),
# then store it to the collection directory
$python ./scripts/run_indexing_for_segmented_transcripts.py -q ./data/PTV.query.jieba.txt -t ./transcripts/PTV.dnn.onebest.jieba -o ./collections/PTV.dnn.onebest.jieba/

Indexing will create 3 pickle files in the collection directory

• lex.pickle
• document.pickle
• indices.pickle

3. Create query & answer pickle file for ranking evaluation

The scripts is located at './data/build_queries.py'

# Loads lex dict from collection directory, then process query text & group them by answer, save to query pickle
$python ./data/build_queries.py -l ./collections/PTV.dnn.onebest.jieba/lex.pickle -q ./data/PTV.query.jieba.txt -a ./data/PTV.ans -o ./queries/PTV.dnn.onebest.jieba.query.pickle

Now we have our query_and_answer pickle at ./queries/PTV.dnn.onebest.jieba.query.pickle

4. Run ranking script

Ranking script is located at './scripts/run_klranker_baseline.py'

# Loads query_answer pickle & transcript indices and outputs Mean Average Precision for the queries
$python scripts/run_klranker_baseline.py -q ./queries/PTV.dnn.onebest.jieba.query.pickle -d ./collections/PTV.dnn.onebest.jieba/

5. Psuedo relevance feedback

Script is located at './scripts.run_pseudo_relevant_feedback.py'. This script assumes top 10 ranked documents are relevant and performs query regularized mixture expansion for second pass ranking. Please see './iscr/dialoguemanager.py' for details.

# Loads the same query & transcript pickle from 4
$python scripts/run_pseudo_relevant_feedback.py -q ./queries/PTV.dnn.onebest.jieba.query.pickle -d ./collections/PTV.dnn.onebest.jieba

Options

• Read old format: parse old files to current format
  • See './data/read_old_format.py' for details
• Word segmentation
  • Jieba
  • CKIP(中研院斷詞系統):
    • you will need to register an account, see './data/ckipauth.json.backup' and ./data/ckipclient.py'
• Indexing
• Ranker

Installation

Code is written in python35

pip install -r requirements.txt

To use it in another project, first execute

python setup.py install

then import

from iscr.ranker import KLRanker

Development

If there's anything you want to add, fork and submit a pull request!

Testing

pytest is used for testing

python -m pytest

To understand more, please read the pytest documentation

License

This project is licensed under the MIT License - see the LICENSE.md file for details