Skip to content
Branch: master
Find file History
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Type Name Latest commit message Commit time
..
Failed to load latest commit information.
controllable_seq2seq
mturk
tasks
Analysis_n_Graphs.ipynb
README.md
eval_wordstat.py
get_bucket_lowerbounds.py
inspect_wordstats.ipynb
interactive.py
make_control_dataset.py
train_controllable_seq2seq.py

README.md

Controllable Dialogue

Paper information

Abigail See, Stephen Roller, Douwe Kiela, Jason Weston. What makes a good conversation? How controllable attributes affect human judgments. To appear in NAACL 2019.

Abstract

A good conversation requires balance -- between simplicity and detail; staying on topic and changing it; asking questions and answering them. Although dialogue agents are commonly evaluated via human judgments of overall quality, the relationship between quality and these individual factors is less well-studied. In this work, we examine two controllable neural text generation methods, conditional training and weighted decoding, in order to control four important attributes for chitchat dialogue: repetition, specificity, response-relatedness and question-asking. We conduct a large-scale human evaluation to measure the effect of these control parameters on multi-turn interactive conversations on the PersonaChat task. We provide a detailed analysis of their relationship to high-level aspects of conversation, and show that by controlling combinations of these variables our models obtain clear improvements in human quality judgments.

Citation

If you use the dataset or models in your own work, please cite with the following BibTex entry:

@inproceedings{see2019what,
  author={Abigail See and Stephen Roller and Douwe Kiela and Jason Weston},
  booktitle={North American Chapter of the Association for Computational Linguistics (NAACL)},
  title={What makes a good conversation? How controllable attributes affect human judgments},
  url={https://arxiv.org/abs/1902.08654},
  year={2019},
}

Code Instructions

Once you have installed ParlAI, follow the instructions below.

Download the data

Running the commands to train or chat with the models will automatically download the data for you. Alternatively, you can manually download the data by running python projects/controllable_dialogue/tasks/build.py. This will download the following files to data/controllable_dialogue:

  • train.txt: This is Convai2 training data, with extra annotations for three CT controllable attributes (question, lastuttsim, avg_nidf). It is in parlai format.
  • valid.txt: Similarly to train.txt.
  • arora.pkl: This is a pickle file containing information necessary to compute Arora-style sentence embeddings, needed for the response-relatedness control methods.
  • word2count.pkl: This is a pickle file containing information necessary to compute NIDF measures, needed for the specificity control methods.
  • personas_validation.txt: This file contains all the ConvAI2 validation set personas, provided for convenience (useful for talking to the model interactively).
  • ConvAI2_parlaiformat/:
    • train.txt: This is the ConvAI2 training set (data/ConvAI2/train_self_original_no_cands.txt) converted to parlai format.
    • valid.txt: This is the ConvAI2 validation set (data/ConvAI2/valid_self_original_no_cands.txt) converted to parlai format.
  • wordstat_files/: This directory contains json files with generated output and automatic metrics computed for the various pretrained models.
  • evaluation_logs/: This directory contains logs and evaluations from the human evaluations.

(Alternatively) Making the data yourself

For reproducibility, in this section we provide the commands to create the data yourself.

Note: Due to changes in ParlAI, there might be some small differences between the generated files obtained in this section, and the downloaded files in the previous section.

First, convert the ConvAI2 data to ParlAI format:

mkdir -p data/controllable_dialogue/ConvAI2_parlaiformat

python parlai/scripts/convert_data_to_parlai_format.py \
--task convai2:SelfOriginal:no_cands \
--datatype train:ordered \
--outfile data/controllable_dialogue/ConvAI2_parlaiformat/train.txt

python parlai/scripts/convert_data_to_parlai_format.py \
--task convai2:SelfOriginal:no_cands \
--datatype valid \
--outfile data/controllable_dialogue/ConvAI2_parlaiformat/valid.txt

Next, create word2count.pkl:

python projects/controllable_dialogue/controllable_seq2seq/nidf.py

This will create a file called word2count.pkl in your data/controllable_dialogue directory. It might take a while, especially the part when it goes through the Twitter dataset counting words.

Next, create arora.pkl:

python projects/controllable_dialogue/controllable_seq2seq/arora.py

This will create a file called arora.pkl in your data/controllable_dialogue directory. It might take a while - in particular, if necessary it will download GloVe vectors and store them in ParlAI/data/models/glove_vectors.

Next, create data/controllable_dialogue/train.txt and valid.txt:

python projects/controllable_dialogue/make_control_dataset.py \
--fromfile_datapath data/controllable_dialogue/ConvAI2_parlaiformat/train.txt \
--outfile data/controllable_dialogue/train.txt \
--controls question,lastuttsim,avg_nidf

python projects/controllable_dialogue/make_control_dataset.py \
--fromfile_datapath data/controllable_dialogue/ConvAI2_parlaiformat/valid.txt \
--outfile data/controllable_dialogue/valid.txt \
--controls question,lastuttsim,avg_nidf

This will create files called train.txt and valid.txt in your data/controllable_dialogue directory.

The pretrained models

Running the commands in the next section to chat with the pretrained models will automatically download them for you. In data/models/controllable_dialogue you will find the following models, along with their .opt files:

  • twitter_pretrained_baseline: A seq2seq model trained on the Twitter dataset.
  • convai2_finetuned_baseline: The twitter_pretrained_baseline model, after fine-tuning on the ConvAI2 dataset.
  • control_avgnidf10b10e: The convai2_finetuned_baseline model, after adding parameters for CT specificity control (10 buckets, embedding size 10), and fine-tuned on the ConvAI2 dataset with loss_CT as described in Section 5.1 of the paper.
  • control_questionb11e10: Similarly to control_avgnidf10b10e, except this is CT question-asking control (11 buckets, embedding size 10).

The directory also contains a dictionary file:

  • dict_twit30k_train_split: This is the dictionary used for all models.

Chat with the pretrained models

This section provides the commands to talk to the model configurations described in the paper. You can refer to Table 5 in the paper to see how these commands correspond to the configurations described there.

Running any of these commands will also download the pretrained models, if necessary.

Talk to the greedy search baseline model:

python projects/controllable_dialogue/interactive.py \
-mf models:controllable_dialogue/convai2_finetuned_baseline \
--beam-size 1

Talk to the beam search baseline model:

python projects/controllable_dialogue/interactive.py \
-mf models:controllable_dialogue/convai2_finetuned_baseline

This setting uses beam size 20 by default.

Talk to the repetition-controlled (WD) baseline:

python projects/controllable_dialogue/interactive.py \
-mf models:controllable_dialogue/convai2_finetuned_baseline \
-wd extrep_2gram:-3.5,extrep_nonstopword:-1e20,intrep_nonstopword:-1e20

You can change the weights for these three WD repetition features to be any real number (positive or negative). Here -1e20 represents -infinity. In addition, there are other repetition WD features you can use if you wish: see the keys of WDFEATURE2UPDATEFN in controllable_seq2seq/controls.py.

Talk to the question-controlled CT model (with WD repetition control):

python projects/controllable_dialogue/interactive.py \
-mf models:controllable_dialogue/control_questionb11e10 \
-wd extrep_2gram:-3.5,extrep_nonstopword:-1e20,intrep_nonstopword:-1e20 \
--set-controls question:7

Here question:7 means the '70% questions' bucket. You can set this anywhere between 0 and 10.

To talk to the "z=10 (boost)" version mentioned in the paper:

python projects/controllable_dialogue/interactive.py \
-mf models:controllable_dialogue/control_questionb11e10 \
-wd extrep_nonstopword:-1e20,intrep_nonstopword:-1e20 \
--set-controls question:10 --beam-reorder best_extrep2gram_qn

Talk to the specificity-controlled CT model (with WD repetition control):

python projects/controllable_dialogue/interactive.py \
-mf models:controllable_dialogue/control_avgnidf10b10e \
-wd extrep_2gram:-3.5,extrep_nonstopword:-1e20,intrep_nonstopword:-1e20 \
--set-controls avg_nidf:7

Here avg_nidf:7 means the 7th specificity bucket (where higher is more specific). You can set this anywhere between 0 and 9.

Talk to the specificity-controlled WD model (with WD repetition control):

python projects/controllable_dialogue/interactive.py \
-mf models:controllable_dialogue/convai2_finetuned_baseline \
-wd extrep_2gram:-3.5,extrep_nonstopword:-1e20,intrep_nonstopword:-1e20,nidf:4

Here nidf:4 means using the NIDF WD feature with weight 4. You can use any real number as a weight (positive or negative).

Talk to the response-relatedness WD model (with WD repetition control):

python projects/controllable_dialogue/interactive.py \
-mf models:controllable_dialogue/convai2_finetuned_baseline \
-wd extrep_2gram:-3.5,extrep_nonstopword:-1e20,intrep_2gram:-1e20,intrep_nonstopword:-1e20,partnerrep_2gram:-1e20,lastuttsim:5

Here lastuttsim:5 means using the response-relatedness WD feature with weight 5. You can use any real number as a weight (positive or negative).

Note that this this feature can take a while to load, especially the first time you run it. This is because we have to load the GloVe vectors from file.

Giving the bot a persona: If you want the bot to have a persona when you talk to it, select one of the lines in data/controllable_dialogue/personas_validation.txt and prepend it to your first utterance. Alternatively you can write the persona yourself - but make sure to use the same format.

Viewing top 10 beam search candidates: If you want to see the top 10 candidates produced by beam search (rather than just the top 1), add the flag --verbose True.

Train a CT model

To train a CT model from scratch:

python projects/controllable_dialogue/train_controllable_seq2seq.py \
-mf /path/to/your/modelfile \
--control-vars avg_nidf

Here we are training a specificity-controlled CT model.

To change control embedding size: The CT control embedding size will default to 10, but you could include e.g. --control-embeddingsize 15 if you wanted to change it.

To change number of buckets: For avg_nidf, the number of buckets will default to 10. If you want to use a different number of buckets, first you need to figure out what the NIDF lower bound should be for each bucket. Suppose you want 8 buckets. First run:

python projects/controllable_dialogue/get_bucket_lowerbounds.py \
--num_buckets 8 \
--control-vars avg_nidf

and then copy and paste the provided lower bounds into projects/controllable_dialogue/controllable_seq2seq/controls.py, similarly to the existing AVG_NIDF_10BUCKET_LBS. Then you can train a model with --control-num-buckets 8.

To train a CT model on multiple controls:

python projects/controllable_dialogue/train_controllable_seq2seq.py \
-mf /path/to/your/modelfile \
--control-vars avg_nidf,question

Here we are training a model conditioned on specificity and question-asking.

To take an existing non-CT model and finetune it as a CT model: First, run this command (in this example, taking the ConvAI2-finetuned baseline and adding specificity control):

python projects/controllable_dialogue/train_controllable_seq2seq.py \
-mf /path/to/your/modelfile \
--init-model models:controllable_dialogue/convai2_finetuned_baseline \
--add-control True \
--control-vars avg_nidf

This command will take the parameters saved in --init-model, load them in the new model (which has randomly initialized weights for the new CT parameters), and then save that model to the given modelfile (-mf). It should be quick. Once that's done, run this command:

python projects/controllable_dialogue/train_controllable_seq2seq.py \
-mf /path/to/your/modelfile \
--add-control False \
--control-vars avg_nidf

You should see your new CT model training. Note: this is how the models in the paper were trained.

Look at generated output and automatic metrics

Once you have downloaded the data, you will find a directory wordstat_files in data/controllable_dialogue. The json files in this directory contain the generated output computed on the ConvAI2 validation set, plus the corresponding automatic metrics. Each json file corresponds to a different model configuration.

Run the following:

cd projects/controllable_dialogue
jupyter notebook

and then open up inspect_wordstats.ipynb. Where it says models_dir, enter the path to your wordstat_files directory. You will be able to recreate the table of automatic metrics from the paper (Table 6), and explore the models' generated output.

Save generated output and automatic metrics to file

If you want to generate json files like those in the previous section, run a command like this:

python projects/controllable_dialogue/eval_wordstat.py \
-mf models:controllable_dialogue/control_questionb11e10 \
-wd extrep_2gram:-3.5,extrep_nonstopword:-1e20,intrep_nonstopword:-1e20 \
--set-controls question:7

This will create a json file containing the output and automatic metrics for the provided model configuration (here, question-controlled CT model with z=7 and WD repetition control). The script eval_wordstat.py always places the json file in the same place as the model file. The script can take a while to complete - so you can set e.g. --num-examples 512 to generate output on a smaller number of examples.

Note: Due to changes in ParlAI, there might be some small differences between the json file created via this method, and the json files downloadable in the previous section.

Human Evaluation code, logs, and analysis

Human evaluation logs should be downloaded automatically after following the download instructions above. You'll find them in the evaluation_logs/ folder.

A Jupyter notebook which generates the graphs and tables for the human experiments is available in the project folder. The notebook should be launched from the ParlAI root directory.

The code for running your own mechanical turk evaluations is also available in the corresponding mturk folder. You will probably want to make changes to the model_config.py and run.py to change which models are being evaluated, and then you can launch the experiment with:

python parlai/mturk/tasks/controllable_dialogue/run.py -r 0.9 --count-complete --hobby --max-resp-time 1200 --max-connections 20 -nc 1200 --sandbox

Change it to --live if you're prepared to spend actual currency. The output must be lightly postprocessed to use it with the analysis tools released. If you intend to do this, please file an issue on the ParlAI GitHub.

You can’t perform that action at this time.