# A New Data Representation Based on Training Data Characteristics to Extract Drug Named-Entity in Medical Text

**Authors:** Sadikin Mujiono, Mohamad Ivan Fanany, Chan Basaruddin  
**arXiv:** [1610.01891](https://arxiv.org/abs/1610.01891)  
**Date:** 2016-10-06  
**Summary page:** https://www.summarizepaper.com/en/paper/1610.01891v1/

## Key points

- Drug name recognition is a crucial task in information extraction from medical texts
- Challenges in medical text include unstructured nature, rapid introduction of new terms, and wide variations in drug names
- Lack of labeled datasets, external knowledge sources, and multiple token representations for drug names exacerbate challenges
- Many existing approaches struggle to achieve satisfactory F-score performance levels (below 0.75)
- A new study introduces innovative data representation techniques to overcome these challenges
- Three distinct techniques proposed based on word distribution characteristics and word similarities derived from word embedding training:
- Evaluation with Multi-Layer Perceptrons (MLP)
- Utilization of Deep Belief Networks (DBN) and Stacked Denoising Encoders (SAE)
- Representation of sentences as sequences using Long Short Term Memory (LSTM) recurrent neural network model
- The third technique utilizing LSTM achieves the best F-score performance with an average of 0.8645
- This research provides valuable insights into enhancing drug name recognition in medical text mining applications and sets a new benchmark for performance in this domain

## Summary

In the field of information extraction from medical texts, one crucial task is drug name recognition. Medical text presents unique challenges compared to other domains due to its unstructured nature and rapid introduction of new terms. Additionally, a wide range of variations in drug names further complicates this task. These challenges are exacerbated by the lack of labeled datasets and external knowledge sources, as well as the presence of multiple token representations for a single drug name. While various approaches have been proposed to address these obstacles, many still struggle to achieve satisfactory F-score performance levels (below 0.75). To tackle this issue, a new study introduces innovative data representation techniques aimed at overcoming these challenges. Three distinct techniques are proposed based on word distribution characteristics and word similarities derived from word embedding training. The first technique involves evaluating data representation with a standard neural network model known as Multi-Layer Perceptrons (MLP). The second technique utilizes two deep network classifiers - Deep Belief Networks (DBN) and Stacked Denoising Encoders (SAE). Lastly, the third technique represents sentences as sequences and is evaluated using a recurrent neural network model called Long Short Term Memory (LSTM). Among these techniques, the third approach stands out by achieving the best F-score performance compared to existing methods. With an average F-score of 0.8645, this technique demonstrates significant improvement in extracting drug name entities from medical texts. This research contributes valuable insights into enhancing drug name recognition in medical text mining applications and sets a new benchmark for performance in this challenging domain.