Skip to content

bionlproc/text-summ-for-doc-retrieval

 
 

Repository files navigation

Literature Retrieval for Precision Medicine with Neural Matching and Faceted Summarization (2020 EMNLP Findings, long-paper)

Neural document matching model and neural text summarization model for scoring documents in reranking purpose; REL is a document-relevance classification model. EXT is an extractive summarization model which identifies relevant words to the trained theme from the source document. ABS is an abstractive model which generates query-like sentences by given topic signal.

If you have any question about this code or any topic relate to this paper, please don't hesitate to send me (jiho dot noh at uky dot edu) an email.

Requirements

  • Stanford CoreNLP library
  • Python3.7+
  • PyTorch (v. 1.4.0)
  • Huggingface Transformers (v. 3.0.2)
  • Solr

Data Preparation

We use the PubMed abstracts and the TREC-Precision Medicine document retrieval reference datasets. Followings are the links to the resources.

Download data files

PubMed

Run this script to download the compressed PubMed files.

e.g., Following will download PubMed files numbered with 0001--1015.

mkdir -p data/pubmed/
cd data/pubmed/
python download.py 1015

TREC relevance judgment files

TREC relevance judgment files are in [data/trec_ref/](data/trec_ref].

BMET embeddings

Download from this repo

We have trained special word embeddings for this project, which we dubbed the BMET embeddings. We trained this model with the biomedical literature along with the MeSH (Medical Subject Headings) entity codes used for biomedical concept annotations. The MeSH codes are encoded in a special way, e.g., 𝛆mesh_D00001 for D00001, and the vectors for entity codes share the same vector space with the regular words. You can download pretrained BMET embeddings from the previous link.

Install Stanford CoreNLP

We use the Stanford CoreNLP to pre-tokenize documents. You can download it from here or use a programming language-specific library.

After client installation, let the shell environment know where the library is located:

export CORENLP_HOME="[path_to]/stanford-corenlp-full-2018-10-05"
export CLASSPATH="$CORENLP_HOME/stanford-corenlp-3.9.2.jar:$CLASSPATH"

Replace [path_to] with the path you saved the Stanford CoreNLP. CORENLP_HOME is needed for the Python stanfordnlp package which runs the installed program as a server.

Run preprocessing.py

In order to run proprocessing script, PubMed corpus needs to be imported into a Solr database. Instruction to install and import PubMed corpus into a Solr core is available in this documentation. You can also refer to the script we used for indexing PubMed documents in the Solr system.

preprocessing.py will generate training datasets split into train/valid sets. Preprocessed datsets are also available for downloading [here]

python preprocessing.py --dir_out [PATH_OUT] --dir_pubmed [PATH_PUBMED] --dir_trec [PATH_TREC] --file_emb [PATH_EMB]

e.g.

python preprocessing.py --dir_out data/tasumm --dir_pubmed data/pubmed --dir_trec data/trec_ref --file_emb data/wbmet-1211.vec

Specify the paths to TREC reference files (PATH_TREC), PubMed documents (PATH_PUBMED), and where to store processed data files (PATH_OUT).

Train

We use a pretrained BERT under the hood, and it has some issues of running in CPU mode with the pretrained parameters. So, we assume that you train this model in GPU mode.

REL

REl model is trained for document-query matching which predicts how probable the given document is relevant to a query (patient case). Following command starts training:

% python train.py --model_type rel 
Jun01 11:42 __main__ INFO: [ === Experiment exp06011142 =============================================================== ]
Jun01 11:42 __main__ INFO: [ Start training rel model  ]
Jun01 11:42 __main__ INFO: [ *** Epoch 1 *** ]
Jun01 11:43 utils INFO: [ steps: 100 loss: 0.4340 recall: 0.6704 prec.: 0.4450 lr 0.000010 time: 59.22s ]
Jun01 11:44 utils INFO: [ steps: 200 loss: 0.3455 recall: 0.2976 prec.: 0.6117 lr 0.000010 time: 118.37s ]
...

EXT

EXT model is trained for keyword extraction from the source document. Following command starts training:

% python train.py --model_type ext
Jun01 11:55 __main__ INFO: [ === Experiment exp06011155 =============================================================== ]
Jun01 11:55 __main__ INFO: [ Start training ext model  ]
Jun01 11:55 __main__ INFO: [ *** Epoch 1 *** ]
Jun01 11:56 utils INFO: [ steps: 100 loss: 0.4564 recall: 0.0109 prec.: 0.8920 lr 0.000010 time: 54.62s ]
Jun01 11:57 utils INFO: [ steps: 200 loss: 0.3747 recall: 0.1938 prec.: 0.8701 lr 0.000010 time: 108.54s ]
...

ABS

ABS model is trained for text summarization for pseudo-query generation. It requires a pretrained EXT model and custom embeddings. Running command should be like:

% python train.py --model_type abs --file_trained_ext [path to pretrained ext model] --file_dec_emb [path to wbmet embeddings] --batch_size 6
Jun01 12:07 __main__ INFO: [ === Experiment exp06011207 =============================================================== ]
Jun01 12:08 model INFO: [ Loading a pre-trained extractive model from data/models/ext_26000000_exp05251635.pt... ]
Jun01 12:08 __main__ INFO: [ Start training abs model  ]
Jun01 12:08 __main__ INFO: [ *** Epoch 1 *** ]
Jun01 12:09 utils INFO: [ steps: 100 loss: 16.8745 lr p0/0.000010, p1/0.001000 time: 34.38s ]
Jun01 12:09 utils INFO: [ steps: 200 loss: 6.5567 lr p0/0.000010, p1/0.001000 time: 68.20s ]
...

Analysis

Heatmap of classification scores by EXT

Attention heatmap produced by [unused0] signal (topic of disease)

Attention heatmap produced by [unused1] signal (topic of generic variants and gene regulations)

Evaluation

doc_scorer runs with an optimized model on the TREC evaluation sets. It save the outputs of ranked documents that can be used for evaluating the document re-ranking performance using the standard TREC evaluation methods.

Tools commonly used by the TREC community for evaluating an ad hoc retrieval runs can be obtained from here.

Acknowledgements

If you use this code or find this effort useful or relevant to your research, please cite the following paper:

Jiho Noh and Ramakanth Kavuluru, "Literature Retrieval for Precision Medicine with Neural Matching and Faceted Summarization"

@inproceedings{noh-2020-literature,
    title = "Literature Retrieval for Precision Medicine with Neural Matching and Faceted Summarization"
    author = "Jiho Noh and Ramakanth Kavuluru"
    booktitle = "Findings of the 2020 Conference on Empirical Methods in Natural Language Processing",
    month = oct,
    year = "2020",
    address = "Online",
    publisher = "Association for Computational Linguistics",
}

About

Neural text summarization for document retrieval

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Python 87.1%
  • Perl 12.9%