From Persona to Person: Enhancing the Naturalness with Multiple Discourse Relations Graph Learning in Personalized Dialogue Generation

Installation and Environment Setup

• OS: Ubuntu 22.04

• GPU

  • NVIDIA RTX 4090 VRAM 24GB
    • For Dialogue Graph Encoder training
  • NVIDIA A100 VRAM 80GB
    • For Generator training

• Python: >=3.9

• Required packages: follow the requirements.txt file.

1. Clone this repository:

    git clone https://github.com/IKMLab/MUDI
    cd MUDI

2. Create a new conda environment and activate it:

   conda create -n mudi python=3.9
   conda activate mudi

3. Install the required packages:

   pip install -r requirements.txt

Dataset Preparation

You could simply download the all dataset from the following link:

• 💾 Google Drive

  # from Google Drive
  # https://drive.google.com/file/d/1nscytAhDEPdDCn5K9nfT7d_RFTXoQZcZ/view?usp=drive_link
  pip install gdown
  gdown --id 1nscytAhDEPdDCn5K9nfT7d_RFTXoQZcZ
  unzip dataset.zip

You can also download the dataset manually and put it in the dataset/RCC/ and dataset/ConvAI2/ directories.

• RCC: We utilize the Reddit Coherence Chain (RCC) dataset, a large-scale chit-chat dialogue dataset designed for dialogue alignment in the pre-training phase of our Dialogue Graph Encoder. The RCC dataset can be accessed here.

  • In our experiments, we use the 5-turn RCC dataset.

• ConvAI2: We utilize the ConvAI2 dataset, a personalized chit-chat dialogue dataset designed for personalized dialogue generation. The ConvAI2 dataset can be accessed here.

  • In our experiments, we use the 'original' version of the ConvAI2 dataset. But you can also use the 'revised' versions of the ConvAI2 dataset.

  • We provide the coherence annotated ConvAI2 dataset in the dataset/ConvAI2/llama3/ and dataset/ConvAI2/mixtral/ directories. (Google Drive)

The dataset folder should be organized as follows:

.
├── dataset
│   ├── ConvAI2
│   │   ├── llama3
│   │   │    ├── train_self_original_coherence.json
│   │   │    └── valid_self_original_coherence.json
│   │   ├── mixtral
│   │   │    ├── train_self_original_coherence.json
│   │   │    └── valid_self_original_coherence.json
│   │   ├── train_self_original.json
│   │   └── valid_self_original.json
├── └── RCC
│       ├──reddit_conversations_v1.0_5turns
│       │   ├── reddit_conversations.5turns.train.txt
│       │   ├── reddit_conversations.5turns.test.txt
│       └── └── reddit_conversations.5turns.dev.txt
├── ...

Data Preprocessing

1. Please convert the raw RCC training set and validation set to the JSON format by running the following code:

   python src/data/parse.py -i dataset/RCC/reddit_conversations_v1.0_5turns/reddit_conversations.5turns.train.txt -o dataset/RCC/reddit_conversations_v1.0_5turns/train.json --dataset rcc

   Argument	Explanation
   -d, --dataset	Dataset name (convai2 or rcc). Choices are 'convai2', 'rcc'.Required.
   -i, --input_file_path	Path to the input file. Only txt file is allowed.Required.
   -o, --output_file_path	Path to save the converted file. Only json file is allowed.Required.

   When the RCC dataset is converted to the JSON format, the data format should be as follows:

   {
       "dialogue": [
           "what is your secret that nobody else knows ?",
           "it 's a secret nobody should know .",
           "go on , you know you want to .",
           "so this is what happened - \" \" i hope you liked it .",
           "you 're the strong silent type are n't you ."
       ]
   }

2. We utilize the Sentence-Transformer as an encoder to extract contextualized global semantics from both utterances and personas, thereby initializing the node features.

   You can obtain the encoding dataset by running the following code:

   For RCC dataset:

   python src/data/preprocess.py -d rcc -i dataset/RCC/reddit_conversations_v1.0_5turns/train.json -o dataset/RCC/reddit_conversations_v1.0_5turns/train.pkl

   For ConvAI2 dataset:

   python src/data/preprocess.py -d convai2 -i dataset/ConvAI2/llama3/valid_self_original_coherence.json -o dataset/ConvAI2/llama3/valid_self_original_coherence.pkl

   Please preprocess the training and validation sets respectively.

   Argument	Explanation
   -d, --dataset	Dataset name. Choices are 'convai2', 'rcc', 'daily_dialog'.Required.
   -i, --input_file_path	Path to the input file.Required.
   -o, --output_file_path	Path to save the dataset after preprocessing. Only pickle files are allowed.Required.
   --augment	Augment the dataset. Only for ConvAI2 dataset. Default is True.

Dialogue Graph Encoder Training

Pre-training

In order to pre-train the Dialogue Graph Encoder, you can run the following code:

sh scripts/pretrain_gnn.sh

Fine-tuning

Please indicate the path to the pre-trained Dialogue Graph Encoder in the --pretrained_model_path argument. You can run the following code to fine-tune the Dialogue Graph Encoder:

sh scripts/train_gnn.sh

Argument	Explanation
--data_dir	Path to the dataset directory. Default is 'dataset/ConvAI2/'.
--processed_train_data_dir	Path to the processed training data directory.
--processed_valid_data_dir	Path to the processed validation data directory.
--train_data_name	Training data name under the data_dir. Default is 'train_self_original_coherence.pkl'.
--valid_data_name	Validation data name under the data_dir. Default is 'valid_self_original_coherence.pkl'.
--processed_train_data_name	Processed training data name under the processed_train_data_dir.
--processed_valid_data_name	Processed validation data name under the processed_valid_data_dir.
--ckpt_dir	Checkpoint directory. Default is 'checkpoints/gnn'.
--num_workers	Number of workers. Default is 4.
--seed	Random seed. Default is 42.
--wandb	Use wandb or not. Default is False.
--wandb_entity	Wandb entity.
--wandb_project	Wandb project.
-a, --wandb_run_name	Wandb run name.
--process_mode	Label (coherence relations) preprocess mode. Default is 'single_filter'.
--k_hop	Keep how many k-hop neighbors. Default is 3.
--reverse_edge	Reverse edge or not. Default is False.
--directed	Directed graph or not. Default is True.
--train_mode	Model training mode. Choices are 'finetuning', 'pretraining'. Default is 'finetuning'.
--do_inference	Do inference only. Default is False.
--cpu	Use CPU or not, if True, then use CPU, else use GPU (cuda). Default is False.
--distributed	Use distributed training or not. Default is False.
--pretrained_model_path	Path to the pretrained model.
--pretrained_utterance_encoder	Pretrained model for utterance/persona encoder. Choices are 'none', 'bert', 'roberta'. Default is 'none'.
--layer_type	Type of GNN encoder layers. Choices are 'GAT', 'GATv2', 'DialogueGAT'. Default is 'DialogueGAT'.
--num_layers	Number of GNN encoder layers. Default is 2.
--num_heads	Number of attention heads if using GAT like layers. Default is 4.
--embedding_dim	Embedding dimension of GNN layers. Default is 512.
--batch_size	Batch size. Default is 512.
--epochs	Number of epochs. Default is 100.
--lr	Learning rate. Default is 0.0001.
--weight_decay	Weight decay. Default is 0.01.
--optimizer	Optimizer. Choices are 'adam', 'adamw', 'sgd', 'adagrad', 'rmsprop', 'sparse_adam'. Default is 'adamw'.
--coh_rel_cls_weight	Loss weight for coherence relations classification. Default is 1.0.
--link_prediction_weight	Loss weight for link prediction. Default is 1.0.
--next_resp_type_direct_weight	Loss weight for next response type prediction. Default is 1.0.
--next_resp_type_seq_weight	Loss weight for next response type prediction. Default is 1.0.
--endure_times	The maximum endure epochs of loss increasing on validation. Default is 10.

Note: If you encounter the error: FileNotFoundError: [Errno 2] No such file or directory: 'dataset/RCC/reddit_conversations_v1.0_5turns//processed_train/processed_train.pt', It is because the processed data is not found. Please delete the dataset/RCC/reddit_conversations_v1.0_5turns/processed_train/ directory and rerun the code. Other similar errors can be solved in the same way.

Generator Training

Please indicate the path to the Dialogue Graph Encoder in the --pretrained_dialogue_encoder_weights_path argument. You can run the following code to train the generator:

sh scripts/train_generator.sh

Argument	Explanation
--data_dir	Path to the dataset directory. Default is 'dataset/ConvAI2/'.
--processed_train_data_dir	Path to the processed training data directory.
--processed_valid_data_dir	Path to the processed validation data directory.
--train_data_name	Training data name under the data_dir. Default is 'train_self_original_coherence.pkl'.
--valid_data_name	Validation data name under the data_dir. Default is 'valid_self_original_coherence.pkl'.
--processed_train_data_name	Processed training data name under the processed_train_data_dir.
--processed_valid_data_name	Processed validation data name under the processed_valid_data_dir.
--ckpt_dir	Checkpoint directory. Default is 'ckpts/generator/ConvAI2/'.
--num_workers	Number of workers. Default is 4.
--seed	Random seed. Default is 42.
--wandb	Use WandB or not. Default is False.
--wandb_entity	WandB entity.
--wandb_project	WandB project.
--wandb_run_name, -a	WandB run name.
--process_mode	Label (coherence relations) preprocess mode. Default is 'single_filter'.
--k_hop	Keep how many k-hop neighbors. Default is 3.
--reverse_edge	Reverse edge or not. Default is False.
--directed	Directed graph or not. Default is True.
--train_mode	Model training mode. Choices are 'finetuning', 'pretraining'. Default is 'finetuning'.
--do_inference	Do inference only. Default is False.
--cpu	Use CPU or not. If True, then use CPU, else use GPU (cuda). Default is False.
--distributed	Use distributed training or not. Default is False.
--pretrained_dialogue_encoder_weights_path	Path to the pretrained dialogue encoder weights.
--pretrained_dialogue_encoder_encoder_weights_path	Path to the pretrained dialogue encoder encoder weights.
--pretrained_utterance_encoder	Pretrained model for utterance/persona encoder. Choices are 'none', 'bert', 'roberta'. Default is 'none'.
--layer_type	Type of GNN encoder layers. Choices are 'GAT', 'GATv2', 'RGAT', 'DialogueGAT'. Default is 'DialogueGAT'.
--num_layers	Number of GNN encoder layers. Default is 2.
--num_heads	Number of attention heads if using GAT like layers. Default is 4.
--embedding_dim	Embedding dimension of GNN layers. Default is 512.
--batch_size	Batch size. Default is 512.
--epochs	Number of epochs. Default is 100.
--lr	Learning rate. Default is 0.0001.
--weight_decay	Weight decay. Default is 0.01.
--optimizer	Optimizer. Choices are 'adam', 'adamw', 'sgd', 'adagrad', 'rmsprop', 'sparse_adam'. Default is 'adamw'.
--coh_rel_cls_weight	Loss weight for coherence relations classification. Default is 1.0.
--link_prediction_weight	Loss weight for link prediction. Default is 1.0.
--next_resp_type_direct_weight	Loss weight for next response type prediction. Default is 1.0.
--next_resp_type_seq_weight	Loss weight for next response type prediction. Default is 1.0.
--endure_times	The maximum endure epochs of loss increasing on validation. Default is 10.
--coherence_attn_strategy	Coherence attention strategy. Choices are 'SP', 'Emb', 'SP+Emb'. Default is 'SP+Emb'.
--graph_encoder_strategy	Graph encoder strategy. Choices are 'Attn', 'Add', 'C', 'P', 'Random', 'None'. Default is 'Attn'.

Inference (Personzied Response Generation)

sh scripts/generate.sh -m <model_name_or_path>

Argument	Explanation
-m, --model_name_or_path	The path or model name.Required.
-t, --tokenizer_name_or_path	The path or tokenizer name.Required.
-o, --output_dir	The output directory where results are saved.Required.
--data_dir	Path to the dataset directory. Default is 'dataset/ConvAI2/'.
--processed_data_dir	Path to the processed data directory.
--data_name	Data file name. Default is 'valid_self_original_coherence.pkl'.
--processed_data_name	Processed data file name.
--num_workers	Number of workers. Default is 4.
--process_mode	Label (coherence relations) preprocess mode. Default is 'single_filter'.
--k_hop	Keep how many k-hop neighbors. Default is 3.
--reverse_edge	Reverse edge or not. Default is False.
--directed	Directed graph or not. Default is True.
--batch_size	Batch size. Default is 4.
--pretrained_utterance_encoder	Pretrained model for utterance/persona encoder. Choices are 'none', 'bert', 'roberta'. Default is 'none'.
--tau	Temperature parameter for sampling in dialogue generation. Default is 0.2.
--top_k_relations	Number of top relations to consider for generating responses. Default is 3.
--coherence_attn_strategy	Coherence attention strategy. Choices are 'SP', 'Emb', 'SP+Emb'. Default is 'SP+Emb'.
--graph_encoder_strategy	Graph encoder strategy. Choices are 'Attn', 'Add', 'C', 'P', 'Random', 'None'. Default is 'Attn'.

Evaluation

We have used the exisiting evaluation metrics to evaluate the generated responses. Please download the models and put them in the root directory.

• NLI: consistent_model/
• BARTScore: bart_score.pth

The structure of the root directory should be as follows:

.
├── assets
├── dataset
├── scripts
├── src
├── consistent_model
├── bart_score.pth
├── ...

We also using the QuantiDCE and DEAM to evaluate the coherence of the generated responses. You can download the QuantiDCE and DEAM models from the following links.

• QuantiDCE
• DEAM

Since the official repositories do not provide convenient inference scripts, we have organized the inference results in the coherence_evaluation folder. To proceed with the inference, please place the contents of the corresponding metric's folder into the root directory of the official method's repository and follow the steps outlined below:

• QuantiDCE:

  1. Replace the util/opt.py with the provided opt.py file.

  2. Put the quantidce_inference.py file in the QuantiDCE root directory.

  3. Put the compute_coherence.sh file in the QuantiDCE script directory.

  4. Replace the INPUT_FILE_PATH and OUTPUT_FILE_PATH in the compute_coherence.sh file with the corresponding paths.

  5. Run the compute_coherence.sh file to evaluate the coherence of the generated responses.

     cd script
     sh compute_coherence.sh

• DEAM:

  1. Place the contents of the coherence_evaluation/DEAM/ folder into the root directory of the DEAM root directory.

  2. Run the convert_format.py to convert the data format to the DEAM format.

     python inference/convert_format.py --model_type mudi -i <input_file_path>

     This will generate files in the same directory as input_file_path with the same name, appended with -single_turn.txt and -multi_turn.txt.

  3. Run the deam_inference.py to evaluate the coherence of the generated responses.

     py inference/deam_inference.py --model_path coh_models --mode predict --eval_data_path <input_file_path>

     The output will be default saved in the coh_models/xxx.txt file.

  4. Run the compute_average_score.py to compute the average score of the coherence evaluation scores.

     python inference/compute_average_score.py -i coh_models/xxx.txt

Furthermore, we provide the evaluation script to evaluate the generated responses with the other metrics. You can run the following code to get the main results (coherence evaluation mentioned above is not included in the main results):

sh scripts/evaluate.sh  -i <generated_responses_file_path> -o <output_csv_file_path>

• All the evaluation metrics are saved in the csv file.

Argument	Explanation
-i, --input_file_path	Path to the data for evaluation.Required.
-o, --output_file_path	Path where evaluation outputs are saved.Required.

Analysis

The Effect of proposed DialogueGAT

To validate the performance of the proposed DialogueGAT compared with existing GNN methods in Dialogue Graph modeling, we analyze results in the Next Response Type Prediction (NRTP) and Coherence Relations Classification (RC) tasks on the validation set.

1. You can run the train_gnn.sh script and change the GNN_LAYER_TYPE variable to GATv2 to train the GATv2 model, or DialogueGAT to train the DialogueGAT model:

   sh scripts/train_gnn.sh

   • We report the best results on validation set of using the GATv2 and DialogueGAT models.

The Effect of τ Values in Dynamic Weighted Aggregation

1. You can run the generate.sh and change TAU variable to evaluate the effect of τ values in dynamic weighted aggregation:

   sh scripts/generate.sh -m <model_name_or_path>

2. After generating the responses, you can run the evaluation script evaluate.sh to evaluate the generated responses.

The Effect of Dialogue Graph Encoder

When training the Generator, we will use the Dialogue Graph Encoder to encode the dialogue graph, and get the coherence-aware dialogue embeddings. Thus, we would like to analyze the effectiveness of the Dialogue Graph Encoder and Attention-based Feature-fusion Strategy.

You can run the train_generator.sh script and change the GRAPH_ENCODER_STRATEGY variable to Attn, Add, C, P, Random, or None to train the generator with different strategies of the Dialogue Graph Encoder output:

sh scripts/train_generator.sh

The Difference Strategy of Coherence-aware Attention

In our model, we propose two strategies for coherence-aware attention: SP and Emb. We analyze the performance of these two strategies in each evaluation metric (report in the main results).

Training the Generator

You can run the train_generator.sh script and change the COHERENCE_ATTN_STRATEGY variable to SP, Emb, or SP+Emb to train the generator with different strategies:

sh scripts/train_generator.sh

Inference (Personalized Response Generation)

When the inference stage, you also can change the COHERENCE_ATTN_STRATEGY variable to SP, Emb, or SP+Emb to generate responses with different attention strategies:

sh scripts/generate.sh -m <model_name_or_path>