Transliterated Pairs Acquisition in Medical Hebrew - Research on improving medical IE lexicons

Aug 28, 2026

Uploaded by Ivy

Transliterations of English-based names and terms into Hebrew create major challenges for information extraction in the medical domain, including slow integration into technical lexicons and inconsistent phonetic spellings that generate many variants and unknown tokens. The study introduces an automatic semi-automatic method to acquire transliterated word pairs together with their source forms. Using a Hebrew medical lexicon from Infomed and a Hebrew morphological analyzer, experiments on an annotated corpus show markedly improved term detection performance while addressing spelling and segmentation errors.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages3

Reading Time8 min

AccessFree to download

Background and Problem

Impact of transliteration on information extraction

Medical domain and lexicon limitations in Hebrew

Data and Resources

Hebrew medical lexicon from Infomed

Annotated corpus statistics

Baseline detection with exact match

Method and Improvements

Morphological analysis to improve segmentation

Error analysis focused on transliterated terms

Semi-automatic acquisition of transliterated term pairs

Related Work and Prior Approaches

Transliteration pair acquisition tasks

Methods in Japanese, Korean, Arabic, and Hebrew

Transliterated Pairs Acquisition in Medical Hebrew Raphael Cohen, Yoav Goldberg and Michael Elhadad Department of Computer Science, Ben-Gurion University, Beer-Sheva, Israel. In most domain-specific texts encoded in a non-Latin alphabet, many proper names, named entities and open-class lexical items are transliterated from English. We investigate some of the problems caused by the high frequency of transliterations in Hebrew in the medical domain. The phonetic transcription of a word from a source language using a different script is called transliteration. Transliterations affect Information Extraction (IE) in two ways. First, it takes time for a transliterated word to make it into a technical lexicon, making recognition difficult. A second problem is the variability of ways a foreign word can be rendered phonetically, leading in most cases (except for very short words) to many possible spellings of the word and, therefore, making lexicon-based recognition difficult. In this paper, we present a method for automatically acquiring transliterated words and their source word in order to improve a technical lexicon, addressing both problems: spelling variants and unknown tokens. Information Extraction tasks commonly occur in specific domains (Financial, Medical, Technology) . Such domains contain technical words, multi-word expressions and proper names specific to the domain. These words in the specific corpora may have different features than the words in a general corpus. In English, state of the art medical IE methods, based on medical term detection, are lexicon based (Jain et al, 1996, Teufel and Elhadad, 2002) . These methods rely on UMLS (Unified Medical Language System) a well maintained collection of over 30 medical vocabularies. In Hebrew, such a vocabulary is not available. We use a lexicon acquired from the medical terms lists in the popular medical information site [Infomed.co.il. This](Infomed.co.il. This) lexicon includes 6,010 medical terms and 1,981 medication names in Hebrew. To examine the performance of the medical term detection task, we use a small annotated corpus of [Questions/Answers from Infomed.co.il](Questions/Answers from Infomed.co.il), comprising 75 documents, 11,508 words and 1,530 annotated medical terms. Term detection using the Infomed lexicon and exact match yields recall of 16% and precision of 84.7% . Using the Infomed lexicon and a Hebrew Morphological Analyzer (Adler and Elhadad 2006) (not adapted to the medical domain) allows us to search lexicon terms after the morphological analyzer segmented the terms from their prefixes. The improved segmentation yielded higher recall at 33% and higher precision at 89% on instances (31% recall and 88% precision on distinct terms). The recall results remain extremely low compared to a similar simple baseline in English text. Preliminary error analysis suggests that a significant portion of our recall errors fall on transliterated medical terms. This stems from spelling variations and segmentation errors on the unknown words. Loose matching to the lexicon, with small edit distance, to address the spelling variation problem isnot feasible since 928 term pairs in the lexicon already have an edit distance of 1 (24% of the terms) . Therefore, if we were to allow matching between terms and the dictionary even if they have an edit distance of 1 would introduce a large number of false positives or ambiguous matches. To address this problem of low recall, we present a method for semi-automatic acquisition of transliterated term pairs using an English medical lexicon and our corpus. We will show that using this enhanced lexicon improves the performance on the term extraction task significantly. Most previous work concerning transliterations focused on transliteration pair acquisition, i.e., recognizing that two words (source, target) are equivalent, as one is a transliteration of the other. Transliteration pair acquisition includes two sub-task

FAQs

Why do transliterations harm medical information extraction in Hebrew?

They delay recognition until variants enter the technical lexicon and they produce many spelling variants due to phonetic variability, making lexicon-based matching difficult.

How is the Hebrew medical lexicon constructed for the experiments?

The lexicon is built from medical term lists on Infomed.co.il, containing 6,010 medical terms and 1,981 medication names in Hebrew.

What is the key result of using the proposed enhanced lexicon?

Improved term extraction performance is reported, particularly by addressing spelling variants and segmentation errors that disproportionately affect transliterated medical terms.

Aug 28, 2026

Uploaded by Ivy