Skip to content

Latest commit

 

History

History
220 lines (170 loc) · 7.92 KB

README.md

File metadata and controls

220 lines (170 loc) · 7.92 KB
Raw

ACCENT: An Automatic Event Commonsense Evaluation Metric for Open-Domain Dialogue Systems

This repository contains the code, models, and datasets for our paper ACCENT: An Automatic Event Commonsense Evaluation Metric for Open-Domain Dialogue Systems.

Quick Links

Overview

We present ACCENT, an event commonsense evaluation metric empowered by commonsense knowledge base ATOMIC. ACCENT uses event-relation tuples as the symbolic intermediate representation and consists of two components:

  1. Event-relation extraction model which extracts event-relation tuples from the target response and its dialogue history.
  2. Compatibility test method which scores each tuple by checking its compatibility against the knowledge base.

Requirements

You can install the required packages through

pip install -r requirements.txt

Datasets

We construct the first publicly available event commonsense evaluation dataset for open-domain dialogues. The dialogue histories in DECO (Dialogue Event COmmonsense dataset) are from human-human dialogues; the dialogue histories in ConTurE Subset are from human-machine dialogues.

The dataset is arranged in JSON format and each sample follows the following format:

{
  'history': 4 utterances separated with '</UTT>',
  'response': target response,
  'event_cs_raw': [annotation1, annotation2, annotation3],
  'event_cs': final labelfor event commonsense score obtained by average aggregation
}

We further request annotators to thoroughly annotate event-relation tuples for DECO test set, the tuples for each sample is divided by relation:

'tuples':
    {
        rel1: [(head, rel1, tail, 'single' / 'pair'), ... ],
        ...
        rel12: [(head, rel1, tail, 'single' / 'pair'), ... ]
    }

Experiments

Commonsense Evaluation Metric

ACCENT consists of an event-relation extraction model and a compatibility test model . To download the models, run cd models; bash download_model.sh.

To test our metric on DECO test set and ConTurE Subset, run the following commands:

# DECO
python -m metric.pipeline --verbose \
  --data_dir data/deco/deco_test.json \
  --saved_tuple_path deco_extracted_tuples.csv \
  --evt_model_dir EchoShao8899/t5_event_relation_extractor \
  --comet_dir models/comet-atomic_2020_BART \
  --embedder 'sentence_bert' or 'simcse'

# ConTurE Subset
python -m metric.pipeline --verbose \
  --data_dir data/conture/conture_subset.json \
  --saved_tuple_path conture_extracted_tuples.csv \
  --evt_model_dir EchoShao8899/t5_event_relation_extractor \
  --comet_dir models/comet-atomic_2020_BART \
  --embedder 'sentence_bert' or 'simcse'

Use ACCENT to evaluate your own dialogues

We provide the following two ways to use ACCENT on your own dialogues:

  1. Formulate your own dataset into JSON format following DECO and ConTurE Subset. 
    [{
  'history': 4 utteraces separated with '</UTT>',
  'response': target response,
  'event_cs': ground truth score (optional)
}, ...]
    Run the following command: 
    python -m metric.pipleine --verbose \
  --data_dir {your own data dir} \
  --saved_tuple_path {saved results path} \
  --evt_model_dir EchoShao8899/t5_event_relation_extractor \
  --comet_dir models/comet-atomic_2020_BART \
  --embedder 'sentence_bert' or 'simcse'
  2. Check the following code snippet to use ACCENT and interpret its score: 
    from metric.accent import ACCENT

metric = ACCENT(
    comet_dir='models/comet-atomic_2020_BART', 
    evt_model_dir='EchoShao8899/t5_event_relation_extractor',
    use_gpu=True,
    embedder='sentence_bert'
)

history = "{utt1}</UTT>{utt2}</UTT>{utt3}</UTT>{utt4}</UTT>"  # Dialogue history with 4 previous utterances.
target = "target response to evaluate"

score, tuples, cs_tails, tuple_scores = metric.score(history, target, debug_mode=True)

print(f'Score: {score}')

if len(tuples) > 0:
    print('Break-down results:')
    tuples = [f'({t[0]}, {t[1]}, {t[2]})' for t in tuples]
    d = max([len(t) for t in tuples])
    for t, s in zip(tuples, tuple_scores):
        print(t.ljust(d), ':', s)

Train your own event-relation extractor

We train the event-relation extractor using the human-extracted tuples for DECO training set and some manually selected negative samples. You can follow scripts/train_event_relation_extractor.sh to train the model with your own data.

Baselines

In this repository, we also include the implementation of some baseline metrics, including FED, Cross-encoder, and MLP Regressor. You can find them in metric.

Event-Relation Extraction

We conduct additional experiments to thoroughly examine the event-relation extraction part of ACCENT (Table 3 in the paper).

To test the event-relation extraction part of ACCENT, run the following command (See the commands above to find how to get deco_extracted_tuples.csv):

python -m eventualization.evt_evaluate \
    --ground_truth_path data/deco/deco_test.json \
    --saved_tuple_path outputs/deco_extracted_tuples.csv

Baselines

We also provide the implementation of two baseline methods for the event-relation extraction part.

  • Baseline 1: ASER Extractor

    Before you run the extractor, you need to train the relation classifiers. Run the following command:

    python -m eventualization.train_lstm_relation_classifier \
    --data_dir data/deco/deco_train.json \
    --epoch 50 --model_dir models/aser_classifiers

    Then run the following command to extract tuples for DECO test set:

    python -m eventualization.aser_runner \
    --stanfordnlp_dir {Stanford core NLP toolkit dir} \
    --data_dir data/deco/deco_test.json \
    --saved_tuple_path outputs/aser_extraction_result.csv

  • Baseline 2: ATOMIC search

    Run the following command to get the result:

    python -m eventualization.one_hop_search \
    --data_dir data/deco/deco_test.json \
    --saved_tuple_path outputs/one_hop_search_result.csv

CSKB Population

The compatibility test approach in ACCENT also outperforms baseline methods for the CSKB Population task (Table 4 in the paper). We use the CSKB population benchmark provided in the EMNLP'21 paper. The benchmark training set can be found here and the test set can be found here. Please download them to data before you run the code for CSKB Population.

Run the following command to evaluate our method on the test set.

bash train_cskb_population_comet.sh
python population/evaluate_cskb_my_comet.py \
  --comet_dir models/my_cskb_population_comet \
  --evaluation_file_path data/evaluation_set.csv

Baselines

We include the implementation of Bert and KG-Bert as the CSKB Population baselines. Please look at scripts/train_bert.sh and scripts/train_kgbert.sh to evaluate them.

Bugs or Questions

If you have any questions related to the code, the data, or the paper, feel free to email Sarik or Yijia. If you encounter any problem when using the code, you can also open an issue to describe your problem.