Injecting domain knowledge in language models for task-oriented dialogue systems

This project contains the code to reproduce the results in:

[EMNLP 2022] Injecting domain knowledge in language models for task-oriented dialogue systems

By Denis Emelin, Daniele Bonadiman, Sawsan Alqahtani, Yi Zhang, Saab Mansour

@Inproceedings{Emelin2022,
 author = {Denis Emelin and Daniele Bonadiman and Sawsan Alqahtani and Yi Zhang and Saab Mansour},
 title = {Injecting domain knowledge in language models for task-oriented dialogue systems},
 year = {2022},
 url = {https://www.amazon.science/publications/injecting-domain-knowledge-in-language-models-for-task-oriented-dialogue-systems},
 booktitle = {EMNLP 2022},
}

Abstract

Pre-trained language models (PLM) have advanced the state-of-the-art across NLP applications, but lack domain-specific knowledge that does not naturally occur in pre-training data. Previous studies augmented PLMs with symbolic knowledge for different downstream NLP tasks. However, knowledge bases (KBs) utilized in these studies are usually large-scale and static, in contrast to small, domain-specific, and modifiable knowledge bases that are prominent in real-world task-oriented dialogue (TOD) systems. In this paper, we showcase the advantages of injecting domain-specific knowledge prior to fine-tuning on TOD tasks. To this end, we utilize light-weight adapters that can be easily integrated with PLMs and serve as a repository for facts learned from different KBs. To measure the efficacy of proposed knowledge injection methods, we introduce Knowledge Probing using Response Selection (KPRS) – a probe designed specifically for TOD models. Experiments1 on KPRS and the response generation task show improvements of knowledge injection with adapters over strong baselines.

---

Required libraries

• nltk
• numpy
• sacrebleu
• tensorboardX
• torch
• transformers
• tqdm
• word2number

Codebase

• create_kprs_benchmark contains python scripts used to construct the KPRS benchmark files, as well as their perturbed variant for the planned knowledge-update experiments
• data_handling contains python scripts used to create and modify data used across all performed experiments (excluding the KPRS benchmark)
• experiments contains python scripts relevant to the experiments performed as part of the project
• training_scripts_ec contains bash scripts used to run experiments on the EC2 instance
• memory_adapter contains code relevant to the memory-network adapter variant that we decided not to pursue further

create_kprs_benchmark

• combine_training_dialogues helper script used to aggregate all MultiWoZ 2.2 training files into a single file for convenience
• collect_contexts collects dialogue contexts and system responses from the MultiWoZ 2.2 train / test / dev data used to construct the KPRS samples
• create_samples generates KPRS samples based on the collected dialogue contexts and system responses, by identifying sets of high-likelihood distractor items in the MultiWoZ 2.2 databases
• filter_train filters the KPRS training data by removing samples that contain entities mentioned in dev / test samples
• perturb_databases creates perturbed variants of MultiWoZ 2.2 databases by reassigning entity names across all database entries (used in the knowledge-update experiments)
• create_perturbed_samples creates perturbed KPRS samples, where positive responses are consistent with the perturbed databases, while the negative responses are consistent with the original databases (used in the knowledge-update experiments)
• sample_for_manual_eval helper script used to sample KPRS samples for manual quality control
• util: a collection of utility functions, primarily for managing MultiWoZ databases

data_handling

• prepare_db_facts derives atomic and composite facts for the adapter training / knowledge injection step from the MultiWoZ 2.2 databases
• create_dialogue_state_tracking_data derives dialogue state tracking samples from the MultiWoZ 2.2 train / dev / test data
• create_response_generation_data derives response generation samples from the MultiWoZ 2.2 train / dev / test data
• annotate_response_generation_samples annotates response generation targets with sets of entities that are supported by the databases and appropriate given the dialogue context
• merge_samples combines single-domain and multi-domain samples into a single file (used in the multi-domain experiments)
• util a collection of utility functions, primarily for managing MultiWoZ databases

experiments

• adapter_model implements the adapter-enhanced language model as well as the various methods for combining adapter and language model representations
• adapter_generation_utils a modified variant of the HuggingFace Transformers generation adjusted to support the adapter model
• pretrain_adapter trains adapters on facts derived from database contents; also used to train the sequentially fine-tuned baselines
• finetune_on_downstream_task fine-tunes adapter models and baselines on down-stream tasks
• evaluate_kprs defines the evaluation methods for the KPRS task
• evaluate_dialogue_state_tracking defines the evaluation methods for dialogue state tracking
• evaluate_response_generation defines the evaluation methods for response generation
• util a collection of utility functions used primarily for data preparation and serving

training_scripts

Script names are meant to be self-explanatory. X_multidomain scripts are used to run multi-domain experiments.

---

Running experiments

To run experiments, execute the corresponding bash script in the training_scripts_ec directory

• Specify the target domain in the --active_domains argument
  • Supported domains are restaurant, hotel, attraction, train for single-domain experiments and mixed for multi-domain experiments
• For pretraining adapters, specify the fact format in the --fact_format argument
  • Supported formats are atomic, composite, and atomic_and_composite
• For fine-tuning, add the --plm_only argument to fine-tune the LM without adapters
• To specifiy the combination function, use the --adapter_combo_method argument
• Supported methods are mean, gate, gate_hidden, concatenate, expert, attention, and gru for use with single adapters, and mean_multi, gate_multi and gate_hidden_multi for use with multiple active adapters
• To load-in pre-trained adapter models, specify the relevant checkpoint in the --adapter_model_checkpoint argument
  • When using multiple adapters, provide paths to all pre-trained models (only the adapter parameters will be loaded in)

Security

See CONTRIBUTING for more information.

License

This project is licensed under the Apache-2.0 License.