A Lightweight Neural Model for Biomedical Entity Linking
Authors: Lihu Chen, Gaël Varoquaux, Fabian M. Suchanek12657-12665
AAAI 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate our model on three datasets (shown in Table 1). The Sh ARe/CLEF corpus (Pradhan et al. 2013) comprises 199 medical reports for training and 99 for testing. As Table 1 shows, 28.2% of the mentions in the training set and 32.7% of the mentions in the test set are unlinkable. The reference knowledge base used here is the SNOMED-CT subset of the UMLS 2012AA (Bodenreider 2004). The NCBI disease corpus (Do gan, Leaman, and Lu 2014) is a collection of 793 Pub Med abstracts partitioned into 693 abstracts for training and development and 100 abstracts for testing. We use the July 6, 2012 version of MEDIC (Davis et al. 2012), which contains 9,664 disease concepts. The TAC 2017 Adverse Reaction Extraction (ADR) dataset consists of a training set of 101 labels and a test set of 99 labels. The mentions have been mapped manually to the Med DRA 18.1 KB, which contains 23,668 unique concepts. Following previous work, we adopt accuracy to compare the performance of different models. |
| Researcher Affiliation | Academia | Lihu Chen,1 Ga el Varoquaux,2 Fabian M. Suchanek,1 1 LTCI & T el ecom Paris & Institut Polytechnique de Paris, France 2 Inria & CEA & Universit e Paris-Saclay, France lihu.chen@telecom-paris.fr, suchanek@telecom-paris.fr, gael.varoquaux@inria.fr |
| Pseudocode | No | The paper includes diagrams of the model architecture (Figure 1) and mathematical formulas but no pseudocode or algorithm blocks. |
| Open Source Code | Yes | All data and code are available at https://github.com/tigerchen52/Biomedical-Entity-Linking. |
| Open Datasets | Yes | We evaluate our model on three datasets (shown in Table 1). The Sh ARe/CLEF corpus (Pradhan et al. 2013) comprises 199 medical reports for training and 99 for testing... The NCBI disease corpus (Do gan, Leaman, and Lu 2014) is a collection of 793 Pub Med abstracts partitioned into 693 abstracts for training and development and 100 abstracts for testing... The TAC 2017 Adverse Reaction Extraction (ADR) dataset consists of a training set of 101 labels and a test set of 99 labels. |
| Dataset Splits | Yes | The NCBI disease corpus (Do gan, Leaman, and Lu 2014) is a collection of 793 Pub Med abstracts partitioned into 693 abstracts for training and development and 100 abstracts for testing. |
| Hardware Specification | Yes | We implemented our model using Keras, and trained our model on a single Intel(R) Xeon(R) Gold 6154 CPU @ 3.00GHz, using less than 10Gb of memory... The CPU is described in the Experimental Settings, and the GPU we used is a single NVIDIA Tesla V100 (32G). |
| Software Dependencies | No | The paper mentions implementing the model using Keras and compares against BERT models (BERT, Tiny BERT, Mobile BERT, ALBERT) but does not provide specific version numbers for these software dependencies (e.g., Keras version, TensorFlow/PyTorch version, specific BERT library versions). |
| Experiment Setup | Yes | Each token is represented by a 200-dimensional word embedding... The dimension of the character LSTM is 64... In the CNN layer, the number of feature maps is 32, and the filter windows are [1,2,3]. The dimension of the context LSTM and entity embedding is set to 32 and 50 respectively. We adopt a grid search on a hold-out set from training samples to select the value τ, and and find an optimal for τ = 0.75. During the training phase, we select at most 20 entity candidates per mention, and the parameter of the triplet rank loss is 0.1. For the optimization, we use Adam with a learning rate of 0.0005 and a batch size of 64. To avoid overfitting, we adopt a dropout strategy with a dropout rate of 0.1. |