ReviseSum

This is the README for the Findings of EMNLP 2022 paper "Learning to Revise References for Faithful Summarization" by Griffin Adams, Han-Chin Shing, Qing Sun, Christopher Winestock, Kathleen McKeown, and Noémie Elhadad. This research was done while Griffin (PhD student at Columbia University) was an intern with the Amazon Comprehend Medical team.

We will update this README soon with instructions on how to access the datasets and models from the paper.

In the meanintime, if you have any questions about the paper or issues running the code, please don't hesitate to contact griffin.adams@columbia.edu or raise an issue on GitHub! We will try to respond as promptly as possible.

High-Level Approach

We propose a new approach to improve reference quality which involves revising--not remove--unsupported reference content. Without ground-truth supervision, we construct synthetic unsupported alternatives (/perturber) to supported sentences and use contrastive learning to discourage/encourage (un)faithful revisions (/ref_reviser). At inference, we vary style codes to over-generate revisions of unsupported reference sentences and select a final revision which balances faithfulness and abstraction. We extract a small corpus from a noisy source (/preprocess)--the MIMIC-III clinical notes from the EHR--for the task of summarizing a hospital admission from multiple notes. We fine-tune BART and Longformer models on original, filtered, and revised data (/gen_transformers), and find that training on revised data is the most effective data-centric intervention for reducing hallucinations.

A high-level diagram of the revision training strategy is shown below.

Code Setup

pip install -e .

To be able to run models and sync results, you will need to create an account on Weights & Biases. Otherwise, please run all models with -offline flag.

pip install wandb wandb login

Feel free to change the Weights & Biases endpoints in ref_reviser/main.py and gen_transformers/main.py settings to point to your own personal project and entity. Otherwise, they will be available at Our Public Reference Revision Project and Our Public Summarization Project, respectively.

angular2html logger = pl_loggers.WandbLogger( name=args.experiment, save_dir=experiment_dir, offline=args.debug or args.offline, project='mimic-sum', entity='griffinadams', )

Entity Extraction and Linking

For entity extraction and linking, we use the publicly available API provided by Amazon Comprehend Medical. Here are some instructions on how to get started (only Step 1 is necessary and Step 4 is helpful to show the output format). If you would like to change the AWS region, please go to preprocess/entity/extract_ents.py and modify the region_name field:

angular2html def get_prod_hera_client(): client = boto3.client(service_name='comprehendmedical', region_name='us-west-2') return client

New users receive a free tier of 8.5 million characters, which should be enough for evaluating summaries, but may not cover perturber pre-training and inference. Please visit pricing page for more information. You can also replace the function #get_ents in preprocess/entity/extract_ents.py with custom entity extraction code, i.e., SciSpacy, MedCAT, https://ctakes.apache.org/, or CLAMP. If you override #get_ents, it must return an object with the following keys: icd, rx, and ent. icd entities are diagnoses, rx stands for RxNorm which provides normalized names for clinical drugs, and ent returns the output from ACM's DetectEntitiesV2, which inspects clinical text for a variety of medical entities and returns specific information about them such as entity category, location, and confidence score on that information. These can produce overlapping spans and the code prioritizes ICD and RXNorm entities over the ent category. Given that this division is unique to ACM, we recommend including diagnoses in icd, all medications in rx, and all other entity categories in ent. The values of icd, rx, ent are each lists, where a single item represent a single entity span and must include the following keys:

Type: semantic type of entity BeginOffset: the string index of the beginning of the entity span EndOffset: the string index of the end of the entity span Text: raw mention span of the entity, i.e, 'Heparin'. Category: High-level category to decide which semantic groups get filtered out. We only filter out an entity if its category is in {'ANATOMY', 'TIME_EXPRESSION', 'PROTECTED_HEALTH_INFORMATION'}.

Type must be one of the following keys in ENT_TYPE_MAP and will be mapped to one of 5 categories:

angular2html ENT_TYPE_MAP = { 'DX_NAME': 'dx', 'PROCEDURE_NAME': 'procedure', 'TREATMENT_NAME': 'treatment', 'TEST_NAME': 'test', 'BRAND_NAME': 'med', 'GENERIC_NAME': 'med' }

The Type information is only needed for perturber pre-training so if you are just using entities for summary or revision analysis, feel free to put a dummy category for each Type, i.e., DX_NAME.

Overview

Please see separate READMEs for

1. Preprocessing MIMIC-III data into a hospital-course summarization corpus (/preprocess). NB: You will need to request access from PhysioNet to download the raw data using this link.

2. Learning to generate synthetic hallucinations with BART (/perturber).

3. Learning to revise unsupported reference sentences with contrastive learning (/ref_reviser).

4. Running summarization models (BART and Longformer) on MIMIC-III revised, original, and filtered references (/gen_transformers). Also provides flags to train by controlling hallucinations (Fillippova, 2020) and with Loss Truncation (Kang and Hashimoto, 2020).