ISPY Installation

This is a installation guideline for "ISPY: Automated Issue Detection and Solution Extraction".

Dialogue Disentanglement

Before runnning the ISPY prediction, the plain text need to be disentangled into dialogues.

The code of dialogue disentanglement is available in disentanglement. We choose the SOTA model irc-disentanglement to seperate the raw dataset.

The steps of running dialogue disentanglement is as follows:

Step 1: Preprocessing.

Generate the .ascii.txt file (Attention: this function may not modify the name of .txt.):

> python3 make-txt.py <filedir>

Replace .txt into .ascii.txt:

> rename txt ascii.txt <filedir>

Generate .tok.txt:

>  python dstc8-tokenise.py --vocab <vocabdir> --output-suffix .tok <filedir> <filedir>

Step 2: Disentanglement. Predict the link of utterances.

> python3 disentangle.py \                          
  <filename>.1\
  --model example-train.dy.model \
  --test <tokdir> \
  --test-start 0 \
  --test-end 5000 \
  --hidden 512 \
  --layers 2 \
  --nonlin softsign \
  --word-vectors <vecdir> \
  > <filename>.out 2><filename>.1.err

Step 3: Extract separate dialogue messages via link graph.

> python3 graph-to-messages.py <filedir> <filedir>

Step 4: Transfer dialogue messages into separate dialogues.

If you need to check the disentanglement result, use this command:

> python3 merge_file.py <filedir>

Otherwise, you can skip this Step 4 and start to predict the ISPY models.

ISPY Model

The ISPY includes two basic prediction models: issue_classification.py and solution_extraction.py.

Both models are available at models, and our SOTA models are reserved in sota_model. We strongly recommend users of ISPY to retrain these two models on new dataset and submit performance issues to us.

Step 1: Build up ISPY-oriented dataset.

The preprocessing source code is available in predicted_is_pairs. use this command to construct the dataset:

> python3 reformat_dialogs.py <filename>.out

The dataset will be constructed as .tsv files in data directory.

Step 2: Issue-Solution prediction.

Enter models directory. Execute both issue and solution models in sequence.

> python3 issue_classification.py
> python3 solution_extraction.py

The extracted issue-solution pairs will saved in data/result_data