This repository hosts the implementation of the paper "Evaluating Coherence in Dialogue Systems using Entailment", published in NAACL'19.
DialogEntailment is a microframework to automatically evaluate coherence in dialogue systems. Our implementation includes the following metrics:
- Semantic Similarity, derived from [Dziri et al., 2018], estimates the correspondence between the utterances in the conversation history and the generated response. The metric is acquired by computing the cosine distance between the embedding vectors of the test utterances in the dialogue history and the generated response.
- Word-level metrics, introduced in [Liu et al., 2016], incorporates word embeddings to measure three metrics: A (average), G (greedy), and E (extrema) (will be added later to the repo)
- Consistency by textual entailment: we cast a generated response as the hypothesis and the conversation history as the premise, projecting thus the automatic evaluation into an natural language inference (NLI) task.
Note that in the paper, we reported distance for the semantic similarity, but in the code, we named the metric SemanticDistance (i.e., the lower the better). We also provided SemanticSimilarity that actually computes the similarity.
DialogEntailment is shipped as a Python package and can be installed using pip:
git clone git@github.com:nouhadziri/DialogEntailment.git
pip install -e .
python -m spacy link en_core_web_lg en
- Python >= 3.6
- SpaCy >= 2.1.0
- allennlp >= 0.8.3
- pytorch-pretrained-bert
- scikit-learn
- tqdm
- smart_open
- pandas
- seaborn
We build a syntenthized entailment corpus, namely InferConvAI, from the ConvAI dialogue data [Zhang et al., 2018], described in details in the paper. The dataset is formatted in both tsv (similar to MultiNLI) and jsonl (following SNLI). To download InferConvAI, please use the following links:
- InferConvAI_v1.3_tsv.tar.gz (84MB download / 236MB uncompressed)
- InferConvAI_v1.3_jsonl.tar.gz (74MB download / 274MB uncompressed)
Check out convai_to_nli.py to see how the synthesized inferenece data is generated from the utterances.
We adopt two prominent models that have shown promising results in commonsense reasoning:
- The Enhanced Sequential Inference Model (ESIM) [Chen et al., 2016] entangled with ELMO [Peters et al., 2018] contextualized word embedding. The implementation is obtained from the AllenNLP library. You can run the following command to train the ESIM model with this configuration:
training/allennlp.sh -s <MODEL_DIR> [--overwrite] [--config <CONFIG_FILE>]- BERT [Devlin et al., 2018]: We fine-tuned a pre-trained BERT model using 🤗 Transformers (when it was called,
pytorch-pretrained-BERT). We modified run_classifier.py to support the entailment task. Here is how to train the model, followed by other arguments that can be passed to the program:
python -m dialogentail.huggingface --do_eval --do_train --output_dir <MODEL_DIR> --train_dataset default: InferConvAI train data
--eval_dataset default: InferConvAI validation data
--model bert-base-uncased, bert-large-uncased (default: bert-base-uncased)
--train_batch_size default: 32
--eval_batch_size default: 8
--num_train_epochs default: 3
--max_seq_length default: 128
You may run the dialogentail module to replicate the plots provided in the paper:
python -m dialogentail --bert_dir <BERT_DIR> --esim_model <ESIM_MODEL> [--plots_dir <DIR>]
For the ESIM model, you need to input model.tar.gz which is generated by allennlp in the model directory once the training is finished.
Note that loading the BERT model and the ESIM model in the same process requires massive amount of memory, so we recommend to run the above command for each model separately.
The default test data is 150 dialogues drawn from Reddit (used in THRED for human evaluation). We also provided a 150-dialogue test data from OpenSubtitles. You can change the test data by the --response_file argument. To use our OpenSubtitles data, simply pass --response_file opensubtitles.
For your own test data, the file format should be the following for each test sample (see our Reddit data for more information):
Line N: TAB-separated utterances in the conversation history Line N+1: the ground-truth response Line N+2: Response generated by Method_1 Line N+3: Response generated by Method_2 ... Line N+m+1: Response generated by Method_m
Run the program with the following arguments:
--response_file Path to your test file
--generator_types The names of 'm' generative models
By default, the program evaluates the following m=4 models:
- Seq2Seq [Vinyals & Le, 2015],
- HRED [Serban et al., 2016],
- TA-Seq2Seq [Xing et al., 2017],
- THRED [Dziri et al., 2018].
To measure the correlation with human judgment, you need to provide a pickle file
containing the mean evaluation ratings of your human judges. More precisely, the pickle file consists of a python list
containing triples ('Method_i', sample_index, mean_rate).
If you have m generative models and N test samples, the size of the list would be N * m:
[('Method_1', 1, 2.1), ('Method_2', 1, 3.4), ..., ('Method_m', 1, 2.6), ('Method_1', 2, 0.2), ...]
To pass your own human judgment file, use --human_judgment <PATH_TO_PICKLE_FILE>. For the OpenSubtitles test data, you may simply set the argument to opensubtitles to use the provided human judgment.
Please cite the following paper if you used our work in your research:
@inproceedings{dziri-etal-2019-evaluating,
title = "Evaluating Coherence in Dialogue Systems using Entailment",
author = "Dziri, Nouha and
Kamalloo, Ehsan and
Mathewson, Kory and
Zaiane, Osmar",
booktitle = "Proceedings of the 2019 Conference of the North {A}merican Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers)",
month = jun,
year = "2019",
address = "Minneapolis, Minnesota",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/N19-1381",
doi = "10.18653/v1/N19-1381",
pages = "3806--3812",
}