exe1023 / dialevalmetrics

A Comprehensive Assessment of Dialog Evaluation Metrics

This repository contains the source code for the following paper:

A Comprehensive Assessment of Dialog Evaluation Metrics

Prerequisties

We use conda to mangage environments for different metrics.

Each directory in conda_envs holds an environment specification. Please install all of them before starting the next step.

Take the installation of conda_envs/eval_base for example, please run

conda env create -f conda_envs/eval_base/environment.yml

Note that there are some packages could not be installed via this method.

If you want find any packages such as bleurt, nlg-eval, and packages downloaded by spaCy are missing, please install it with official instructions.

We apologize for any inconvenience.

Data Preparation

The directory of each qualitiy-annotated data is placed in data, with the data_loader.py for parsing the data.

Please follow the below instructions to downlaod each dataset, place it into corresponding directory, and run the data_loader.py directly to see if you use the correct data.

DSTC6 Data

Download human_rating_scores.txt from https://www.dropbox.com/s/oh1trbos0tjzn7t/dstc6_t2_evaluation.tgz .

DSTC9 Data

Download and Place the data directory https://github.com/ictnlp/DialoFlow/tree/main/FlowScore/data into data/dstc9_data.

Engage Data

Download https://github.com/PlusLabNLP/PredictiveEngagement/blob/master/data/Eng_Scores_queries_gen_gtruth_replies.csv and rename it to engage_all.csv.

Fed Data

Download http://shikib.com/fed_data.json .

Grade Data

Download and place each directory in https://github.com/li3cmz/GRADE/tree/main/evaluation/eval_data as data/grade_data/[convai2|dailydialog|empatheticdialogues].

Also download the human_score.txt in https://github.com/li3cmz/GRADE/tree/main/evaluation/human_score into the corresponding data/grade_data/[convai2|dailydialog|empatheticdialogues].

Holistic Data

Download context_data_release.csv and fluency_data_release.csv from https://github.com/alexzhou907/dialogue_evaluation .

USR Data

Download TopicalChat and PersonaChat data from http://shikib.com/usr

Metric Installation

For baselines, we use the nlg-eval. Please folloow the instruction to install it.

For each dialog metrics, please follow the instructions in README in the corresponding directory.

Running Notes for Specific metrics

bert-as-service

PredictiveEngage, BERT-RUBER and PONE requires the running bert-as-service.

If you want to evaluate them, please install and run bert-as-service following the instrucitons here.

We also provide a script we used to run bert-as-service run_bert_as_service.sh, feel free to use it.

running USR and FED

We used a web server for running USR and FED in our experiments.

Please modify path in usr_fed/usr/usr_server.py and usr_fed/fed/fed_server.py to start the server, and modify the path in usr_fed_metric.py.

How to evaluate

1. After you downlaod all datasets, run gen_data.py to transform all datasets into the input format for all metrics. If you only want to evaluate metric metric and dataset dataset, run with gen_data.py --source_data dataset --target_format metric

2. Modify the path in run_eval.sh as specified in the script, since we need to activate Conda environment when running the script. Run eval_metrics.sh to evaluate all quality-anntoated data.

3. Some metrics generate the output in its special format. Therefore, we should run read_result.py to read the results of those metrics and transform it into outputs. As step 1, you can specify the metric and data by read_result.py --metric metric --eval_data dataset.

4. The outputs/METRIC/DATA/results.json holds the prediction score of each metrics (METRIC) and qualitiy-anntoated data (DATA), while running data_loader.py directly in each data directory also generates the corresponding human scores. You can perform any analysis with the data (The jupyter notebook used in our analysis will be released) .

For example, outputs/grade/dstc9_data/results.json could be

    'GRADE': # the metric name
    [
        0.2568123, # the score of the first sample
        0.1552132, 
        ...
        0.7812346
    ]

Results

All values are statistically significant to p-value < 0.05, unless marked by *.

USR Data

	USR-TopicalChat				USR-Pearsonachat
	Turn-Level		System-Level		Turn-Level		System-Level
	P	S	P	S	P	S	P	S
BLEU-4	0.216	0.296	0.874*	0.900	0.135	0.090*	0.841*	0.800*
METEOR	0.336	0.391	0.943	0.900	0.253	0.271	0.907*	0.800*
ROUGE-L	0.275	0.287	0.814*	0.900	0.066*	0.038*	0.171*	0.000*
ADEM	-0.060*	-0.061*	0.202*	0.700*	-0.141	-0.085*	0.523*	0.400*
BERTScore	0.298	0.325	0.854*	0.900	0.152	0.122*	0.241*	0.000*
BLEURT	0.216	0.261	0.630*	0.900	0.065*	0.054*	-0.125*	0.000*
QuestEval	0.300	0.338	0.943	1.000	0.176	0.236	0.885*	1.000
RUBER	0.247	0.259	0.876*	1.000	0.131	0.190	0.997	1.000
BERT-RUBER	0.342	0.348	0.992	0.900	0.266	0.248	0.958	0.200*
PONE	0.271	0.274	0.893	0.500*	0.373	0.375	0.979	0.800*
MAUDE	0.044*	0.083*	0.317*	-0.200*	0.345	0.298	0.440*	0.400*
DEB	0.180	0.116	0.818*	0.400*	0.291	0.373	0.989	1.000
GRADE	0.200	0.217	0.553*	0.100*	0.358	0.352	0.811*	1.000
DynaEval	-0.032*	-0.022*	-0.248*	0.100*	0.149	0.171	0.584*	0.800*
USR	0.412	0.423	0.967	0.900	0.440	0.418	0.864*	1.000
USL-H	0.322	0.340	0.966	0.900	0.495	0.523	0.969	0.800*
DialogRPT	0.120	0.105*	0.944	0.600*	-0.064*	-0.083*	0.347*	0.800*
Deep AM-FM	0.285	0.268	0.969	0.700*	0.228	0.219	0.965	1.000
HolisticEval	-0.147	-0.123	-0.919	-0.200*	0.087*	0.113*	0.051*	0.000*
PredictiveEngage	0.222	0.310	0.870*	0.900	-0.003*	0.033*	0.683*	0.200*
FED	-0.124	-0.135	0.730*	0.100*	-0.028*	-0.000*	0.005*	0.400*
FlowScore	0.095*	0.082*	-0.150*	0.400*	0.118*	0.079*	0.678*	0.800*
FBD	-	-	0.916	0.100*	-	-	0.644*	0.800*

GRADE Data

	GRADE-ConvAI2				GRADE-DailyDialog				GRADE-EmpatheticDialogue
	Turn-Level		System-Level		Turn-Level		System-Level		Turn-Level		System-Level
	P	S	P	S	P	S	P	S	P	S	P	S
BLEU-4	0.003*	0.128	0.034*	0.000*	0.075*	0.184	1.000*	1.000	-0.051*	0.002*	1.000*	1.000
METEOR	0.145	0.181	0.781*	0.600*	0.096*	0.010*	-1.000*	-1.000	0.118	0.055*	1.000*	1.000
ROUGE-L	0.136	0.140	0.209*	0.000*	0.154	0.147	1.000*	1.000	0.029*	-0.013*	1.000*	1.000
ADEM	-0.060*	-0.057*	-0.368*	-0.200*	0.064*	0.071*	1.000*	1.000	-0.036*	-0.028*	1.000*	1.000
BERTScore	0.225	0.224	0.918*	0.800*	0.129	0.100*	-1.000*	-1.000	0.046*	0.033*	1.000*	1.000
BLEURT	0.125	0.120	-0.777*	-0.400*	0.176	0.133	1.000*	1.000	0.087*	0.051*	1.000*	1.000
QuestEval	0.279	0.319	0.283*	0.400*	0.020*	0.006*	-1.000*	-1.000	0.201	0.272	1.000*	1.000
RUBER	-0.027*	-0.042*	-0.458*	-0.400*	-0.084*	-0.094*	-1.000*	-1.000	-0.078*	-0.039*	1.000*	1.000
BERT-RUBER	0.309	0.314	0.885*	1.000	0.134	0.128	-1.000*	-1.000	0.163	0.148	1.000*	1.000
PONE	0.362	0.373	0.816*	0.800*	0.163	0.163	-1.000*	-1.000	0.177	0.161	1.000*	1.000
MAUDE	0.351	0.304	0.748*	0.800*	-0.036*	-0.073*	1.000*	1.000	0.007*	-0.057*	1.000*	1.000
DEB	0.426	0.504	0.995	1.000	0.337	0.363	1.000*	1.000	0.356	0.395	1.000*	1.000
GRADE	0.566	0.571	0.883*	0.800*	0.278	0.253	-1.000*	-1.000	0.330	0.297	1.000*	1.000
DynaEval	0.138	0.131	-0.996	-1.000	0.108*	0.120	-1.000*	-1.000	0.146	0.141	-1.000*	-1.000
USR	0.501	0.500	0.995	1.000	0.057*	0.057*	-1.000*	-1.000	0.264	0.255	1.000*	1.000
USL-H	0.443	0.457	0.971	1.000	0.108*	0.093*	-1.000*	-1.000	0.293	0.235	1.000*	1.000
DialogRPT	0.137	0.158	-0.311*	-0.600*	-0.000*	0.037*	-1.000*	-1.000	0.211	0.203	1.000*	1.000
Deep AM-FM	0.117	0.130	0.774*	0.400*	0.026*	0.022*	1.000*	1.000	0.083*	0.058*	1.000*	1.000
HolisticEval	-0.030*	-0.010*	-0.297*	-0.400*	0.025*	0.020*	1.000*	1.000	0.199	0.204	-1.000*	-1.000
PredictiveEngage	0.154	0.164	0.601*	0.600*	-0.133	-0.135	-1.000*	-1.000	-0.032*	-0.078*	1.000*	1.000
FED	-0.090	-0.072*	-0.254*	0.000*	0.080*	0.064*	1.000*	1.000	-0.014*	-0.044*	1.000*	1.000
FlowScore	-	-	-	-	-	-	-	-	-	-	-	-
FBD	-	-	-0.235*	-0.400*	-	-	-1.000*	-1.000	-	-	-1.000*	-1.000

DSTC6 Data

	DSTC6
	Turn-Level		System-Level
	P	S	P	S
BLEU-4	0.131	0.298	-0.064*	0.050*
METEOR	0.307	0.323	0.633	0.084*
ROUGE-L	0.332	0.326	0.487	0.215*
ADEM	0.151	0.118	0.042*	0.347*
BERTScore	0.369	0.337	0.671	0.265*
BLEURT	0.326	0.294	0.213*	0.426*
QuestEval	0.188	0.242	-0.215*	0.206*
RUBER	0.114	0.092	-0.074*	0.104*
BERT-RUBER	0.204	0.217	0.825	0.093*
PONE	0.208	0.200	0.608	0.235*
MAUDE	0.195	0.128	0.739	0.217*
DEB	0.211	0.214	-0.261*	0.492
GRADE	0.119	0.122	0.784	0.611
DynaEval	0.286	0.246	0.342*	-0.050*
USR	0.184	0.166	0.432*	0.147*
USL-H	0.217	0.179	0.811	0.298*
DialogRPT	0.170	0.155	0.567	0.334*
Deep AM-FM	0.326	0.295	0.817	0.674
HolisticEval	0.001*	-0.004*	0.010	-0.002
PredictiveEngage	0.043	0.004*	-0.094*	-0.409*
FED	-0.106	-0.083	0.221*	0.322*
FlowScore	0.064	0.095	0.352*	0.362*
FBD	-	-	-0.481	-0.234*

PredictiveEngage-DailyDialog

	PredictiveEngage-DailyDialog
	Turn-Level
	P	S
QuestEval	0.296	0.341
MAUDE	0.104	0.060*
DEB	0.516	0.580
GRADE	0.600	0.622
DynaEval	0.167	0.160
USR	0.582	0.640
USL-H	0.688	0.699
DialogRPT	0.489	0.533
HolisticEval	0.368	0.365
PredictiveEngage	0.429	0.414
FED	0.164	0.159
FlowScore	-	-
FBD	-	-

HolisticEval-DailyDialog

	HolisticEval-DailyDialog
	Turn-Level
	P	S
QuestEval	0.285	0.260
MAUDE	0.275	0.364
DEB	0.584	0.663
GRADE	0.678	0.697
DynaEval	-0.023*	-0.009*
USR	0.589	0.645
USL-H	0.486	0.537
DialogRPT	0.283	0.332
HolisticEval	0.670	0.764
PredictiveEngage	-0.033*	0.060*
FED	0.485	0.507
FlowScore	-	-
FBD	-	-

FED Data

	FED
	Turn-Level		Dialog-Level
	P	S	P	S
QuestEval	0.037*	0.093*	-0.032*	0.080*
MAUDE	0.018*	-0.094*	-0.047*	-0.280
DEB	0.230	0.187	-0.130*	0.006*
GRADE	0.134	0.118	-0.034*	-0.065*
DynaEval	0.319	0.323	0.503	0.547
USR	0.114	0.117	0.093*	0.062*
USL-H	0.201	0.189	0.073*	0.152*
DialogRPT	-0.118	-0.086*	-0.221	-0.214
HolisticEval	0.122	0.125	-0.276	-0.304
PredictiveEngage	0.024*	0.094*	0.026*	0.155*
FED	0.120	0.095	0.222	0.320
FlowScore	-0.065*	-0.055*	-0.073*	-0.003*
FBD	-	-	-	-

DSTC9 Data

	DSTC9
	Dialog-Level		System-Level
	P	S	P	S
QuestEval	0.026*	0.043	0.604	0.527*
MAUDE	0.059	0.042*	0.224*	0.045*
DEB	0.085	0.131	0.683	0.473*
GRADE	-0.078	-0.070	-0.674	-0.482*
DynaEval	0.093	0.101	0.652	0.727
USR	0.019*	0.020*	0.149*	0.127*
USL-H	0.105	0.105	0.566*	0.755
DialogRPT	0.076	0.069	0.685	0.555*
HolisticEval	0.015*	0.002*	-0.019*	-0.100*
PredictiveEngage	0.114	0.115	0.809	0.664
FED	0.128	0.120	0.559*	0.391*
FlowScore	0.147	0.140	0.907	0.900
FBD	-	-	-0.669	-0.627

How to Add New Dataset

Let the name of the new dataset be sample

Create a directory data/sample_data, write a function load_sample_data as follow:

def load_sample_data(base_dir: str):
    '''
    Args: 
        base_dir: the absolute path to data/sample_data
    Return:
        Dict:
        {
            # the required items
            'contexts' : List[List[str]], # dialog context. We split each dialog context by turns. Therefore one dialog context is in type List[str].
            'responses': List[str], # dialog response.
            'references': List[str], # dialog references. If no reference in the data, please still give a dummy reference like "NO REF".
            "scores": List[float] # human scores.
            # add any customized items
            "Customized Item": List[str] # any additional info in the data. 
        }
    '''

Import the function in gen_data.py, and run with python gen_data.py --source_data sample

How to Add New Metrics

Let the name of the new metric be metric

Write a function gen_metric_data to transform and generate the data into the metric directory:

# input format 1
def gen_metric_data(data: Dict, output_path: str):
    '''
    Args:
        data: the return value of load_data functions e.g. {'contexts': ...}
        output_path: path to the output file
    '''

# input format 2
def gen_metric_data(data: Dict, base_dir: str, dataset: str):
    '''
    Args:
        data: the return value of load_data functions e.g. {'contexts': ...}
        base_dir: path to the output directory
        dataset: name of the dataset
    '''

We have two input formats. Just follow the one which is easier for you.

Import the function in gen_data.py and follow comments in the code to add the metric.

Then write a function read_metric_result to read the prediction of the metric:

def read_metric_data(data_path: str):
    '''
    Args:
        data_path: path to the prediction file
    
    Return:
        # You can choose to return list or dict
        List: metric scores e.g. [0.2, 0.3, 0.4, ...]
        or 
        Dict: {'metric': List # metric scores}
    '''

Import the function in read_result.py and follow comments in the code to add the metric.

Then just follow the previous evaluation instructions to evaluate the metric.

dialevalmetrics's Issues

Unable to find "run_eval.sh"

Modify the path in run_eval.sh as specified in the script, since we need to activate Conda environment when running the script. Run eval_metrics.sh to evaluate all quality-anntoated data.

---

I can't find "run_eval.sh". Who knows where it is？

conda install:Found conflicts! Looking for incompatible packages.

when i run "conda env update --name eval_base --file conda_envs/eval_base/environment.yml" on mac,
error reports :
Collecting package metadata (repodata.json): done
Solving environment:
Found conflicts! Looking for incompatible packages.
This can take several minutes. Press CTRL-C to abort.
failed
Solving environment:
Found conflicts! Looking for incompatible packages.
This can take several minutes. Press CTRL-C to abort.
failed

UnsatisfiableError: The following specifications were found to be incompatible with a past
explicit spec that is not an explicit spec in this operation (libgcc-ng):

能不能单独使用其中一个评价指标呀？

比如我想单独直接使用ADEM，运行ADEM里面的eval.py

Unable to reexperiment with GRADE

So thanks for your existing work !

I have get the result of grade_convai and grade_dailydialog, for both pearson and spearman scores are the same with yours. While for grade_empatheticdialogues, I get the different results under the same computation and evaluate methods.

1. First I use nlg_eval to compute baseline metrics for both three datasets
2. I modify the human_correlation.py script in data/grade_data, to compute for all the scores in both three datasets, and as described in above, two of them get the same but last one is different.

Waiting for your suggestions, thanks ！

如何获取bert-ruber分数和PONE分数

你好，请问bert-ruber和PONE有什么区别呀
我发现使用gen_data.py创建bert-ruber的数据时，产生的数据是放在了pone文件夹中，运行相应run之后，会产生三个文件：refer_score_dailydialog_sampled.json、ruber_score_dailydialog_sampled.json、unrefer_score_dailydialog_sampled.json，我想问一下这三个文件分别代表什么？
哪一个是bert-ruber的分数，哪个是pone的分数呀？

Unable to reproduce results

I'm unable to get the desired results by following the steps as given in the README due to multiple issues such as:-

1. Environment creation failures for some of the them
2. gen_data.py is giving errors for DSTC6 and Engage
3. Transformers version issues for FBD, Maude (Eval_base environment) and TF version issues for run_bert_as_service.sh
   and some others as well.
4. Various files such as restore/ensemble.yml, ./data/DailyDialog/keyword.vocab etc. not being present
   So I request to update the repository with the latest codes and files if possible to help get the results as highlighted in your paper.

regr.pkl file not available

when i run "python usr_fed/usr/usr_server.py", error arrears:
Traceback (most recent call last):
File "usr_server.py", line 1, in
import usr
File "/home/zper/DialEvalMetrics/usr_fed/usr/usr.py", line 6, in
import regression
File "/home/zper/DialEvalMetrics/usr_fed/usr/regression.py", line 6, in
with open('regr.pkl','rb')as f:
FileNotFoundError: [Errno 2] No such file or directory: 'regr.pkl'

after i wget https://github.com/Shikib/usr/blob/master/examples/regr.pkl from original usr model github,
error also appears:
pickle invalid load key, '\x0a'.
i wonder if you can provide your pkl file, thanks.

Turn-level DSTC9 annotation data

I noticed there is just dialog-level DSTC9 data used in your work. May I ask if turn-level DSTC9 annotation data is made public? How can I get this data? Thank you very much!