princeton-nlp / Cognac

Repo for paper: Controllable Text Generation with Language Constraints

Cognac

Repo for paper: Cognac: Controllable Text Generation with Language Constraints

Overview

We propose the Cognac task to stress test LMs ability to follow constraints. Green highlight specifies the topic to be covered. Red highlight specifies the constraint to conform to. GPT-3 generates continuation that mentioned a politician, thus violating the constraint. Our method, CognacGen, generates continuation that satisfies both the topic requirement and the constraint.

Setup

We use Python version 3.8. Install the dependencies with pip:

pip install -r requirements.txt

Download necessary resources:

python -m spacy download en_core_web_lg
python -m nltk.downloader wordnet

Cognac Benchmark

WordNet

Download the WordNet data here. The folder contains files train.jsonl, dev.jsonl, and test.jsonl that include instances of instructions with topics and constraints. The file topic_to_leafs.json contains the WordNet hierarchy (used to verify if the generation is conformant). The data is loaded in the code here.

Wikidata

Coming soon...

CognacGen

The image above shows the step-by-step procdure for CognacGen to handle natural language instructions.

Stage 1: the LM generates a list of guidance examples from the queries that specify the topic and constraint. During self-guidance distillation, the topic and constraint prefixes are tuned using the guidance example as target and the instruction with demonstrations as input.

Stage 2: The guidance model (blue LM & the tuned prefixes) generates guidance examples from the test instance. The guidance examples are used to construct trie trees for both the topic (green) and the constraint (red). The generation (blue) LM’s next token probability is modified by the tries.

Overall Structure

The main run script is main.py. Some important hyperparameters are described below:

• eval_version: determining if control code or natural language instruction is used as context
• guidance: the combination of guidances to use. in means the inclusion of topic is applied. ex means the exclusion of constraint is applied. wd (weighted decoding) is used in CognacGen. Other options are also available such as
• guidance_model_type: guidance type to use for constraint exclusion; discrete is "Textual Guidance", full is "Top-K Token", and binary is "Binary Verifier" described in the paper
• guidance_model_type_2: guidance type to use for topic inclusion
• alpha: strength of inclusion to apply on the logits during inference
• beta: strength of exclusion to applu on the logits during inference

Run Control Code Setting on WordNet

python -m src.main \
    --name "your_run_name" \
    --dataset_split "dev" \
    --dev_path "./data/wordnet/dev.jsonl" \
    --hierarchy_path "./data/wordnet/topic_to_leafs.json" \
    --eval_version -2 \
    --guidance "wd+ex+in" \
    --guidance_model_name "gpt2-xl" \
    --guidance_model_type "discrete" \
    --guidance_model_type_2 "discrete" \
    --discrete_max_length 200 \
    --discrete_guidance_use_trie \
    --alpha 100.0 \
    --beta 5.0 \
    --top_p 0.92 \
    --temperature 0.7

Note that the control code setting applies only stage 2 and does not fine-tune the guidance model.

Run Natural Language Instruction Setting on WordNet

python -m src.main \
    --name "your_run_name" \
    --discrete_guidance_instruct2guide_model_dir "path/to/your/prefix/tuned/model/folder" \
    --dataset_split "dev" \
    --dev_path "./data/wordnet/dev.jsonl" \
    --hierarchy_path "./data/wordnet/topic_to_leafs.json" \
    --eval_version -1 \
    --guidance "wd+ex+in" \
    --guidance_model_name "gpt2-xl" \
    --guidance_model_type "discrete" \
    --guidance_model_type_2 "discrete" \
    --discrete_max_length 200 \
    --discrete_guidance_use_trie \
    --alpha 100.0 \
    --beta 5.0

The prefix-tuned model in stage 1 can be downloaded here.

Run Control Code Setting on Wikidata

Coming soon...

Run Natural Language Instruction Setting on Wikidata

Coming soon...

Prefix-Tuning the Guidance Model

Coming soon...

Questions

Please contact Howard Chen ([e-mail]) if you have any questions.

Citation

@inproceedings{chen2022cognac,
   title={{Cognac}: Controllable Text Generation with Language Constraints},
   author={Chen, Howard and Li, Huihan and Chen, Danqi and Narasimhan, Karthik},
   booktitle={arXiv},
   year={2022}
}