The increasing availability of parallel bilingual corpora and of automatic methods and tools for their processing makes it possible to build linguistic and terminological resources for low-resourced languages.
with French and English terms.
The acquisition of terminology has gone through a
very active period and provides nowadays several
automatic tools and methods
Our main objective is to create terminological resources for Ukrainian, for which very little digitized or electronic resources are available. Yet, the terminology extraction tools usually require the morpho-syntactic tagging of texts, which can be problematic if the corresponding automatic tools are not available for a given language. For in
UMR8163 STL
France does not perform the syntactic and morphological disambiguation of the tags. Hence, it becomes impossible to use it for the pre-processing of corpora before the traditional terminology acquisition process.
In this situation, we propose first to compile terminological resources for Ukrainian in order to build the basis for the observation of the specificities of terminological units in this language. Such observations will allow to develop and parameter the terminology extraction tool for Ukrainian.
The motivation of our work is double. We want
2. design specific methods for the acquisition of such terminological resources.
The work is carried out with medical data, and in three languages (Ukrainian, French, and English).
The work we present starts from the exploitation of two kinds of corpora (Section 2.1): Wikipedia in Ukrainian which provides several useful kinds of information (such as term labels and their codes) with a high level of quality, and the parallel corpus MedlinePlus. The term detection and extraction can be either manual or automatic. Since, there is no appropriate POS-tagging and term extraction tools for Ukrainian, we propose to use such tools in French and English, and to take advantage of these to transfer English and French extracted terms on the Ukrainian corpus.
Indeed, the transfer methodology can be
considstance, the UGtag Part-of-Speech (POS) tagger
ered as suitable for such objectives. Suppose we
have parallel and aligned corpora with two
languages L1 and L2, and we have several types of
syntactic or semantic annotations and information
associated to L1. The transfer approach permits
to transpose these annotations or information from
L1 to L2, and to obtain in this way the
corresponding annotations and information in the L2
text. From this point of view, L1 is considered as
the source language while L2 is considered as the
target language. This kind of approach is
particularly interesting when working with low-resourced
languages for which less tools and semantic
resources are available. An increasing
availability of parallel bilingual corpora, and of automatic
methods and tools for their processing makes it
possible to build linguistic and terminological
resources using the transfer methodology
We propose to apply this method for the acquisition of bilingual or trilingual terminologies involving Ukrainian. In our work, each corpus is exploited through dedicated methods. The MedlinePlus corpus provides the basis for the building of the terminology, while the Wikipedia corpus permits to enrich this information and helps the wordlevel alignment of the MedlinePlus corpus.
Terminology-related research on Ukrainian is
an active area, although the main
terminological work shows mainly theoretical and
linguistic orientation
In the following of this paper, we first present the material used for the acquisition of bilingual terminology (section 2), and the methods designed for achieving this objective (section 3). We then discuss the results we obtain (section 4), and conclude with directions for the future work (section 5). 2 2.1
• MedlinePlus: parallel medical corpus from MedlinePlus. These data are built by MedlinePlus from the National Library of Medicine1. They contain patient-oriented 1www.nlm.nih.gov/medlineplus/healthtopics.html Corpus Wikipedia/UKmed MedlinePlus/UK MedlinePlus/FR MedlinePlus/EN
Size (occ of words) 246,368,411 43,184 53,067 46,544 brochures on several medical topics (body systems, disorders and conditions, diagnosis and therapy, demographic groups, health and wellness). These brochures have been created in English and then translated in several other languages, among which French and Ukrainian; • Wikipedia: medicine-related articles from Wikipedia. This corpus is extracted from the Ukrainian part of the Wikipedia using medicine-related categories, such as Медицина (medicine) or Захворювання (disorders). The corpus potentially covers a wide range of medical notions. In Figure 1, we indicate an example of the source pages which propose the navigation frame on the left, the text with explanations and the infobox with illustration and coding on the right.
In Table 1, we indicate the size of the corpora. Not surprisingly, the Wikipedia corpus is much larger although only part of its information is exploited, as we will see in the next section. 2.2
The UMLS (Unified Medical Language System)
The methods we propose for the extraction of bilingual terminology are adapted to each kind of corpora and of data they contain: the MedlinePlus corpus (section 3.1) and the Wikipedia corpus (section 3.2). We then present their crossfertilization (Section 3.3), and the evaluation of the results (Section 3.4). 3.1 Prior to the exploitation of the MedlinePlus data, the documents are first transformed in a suitable format: • the source PDF documents are converted in the text format; • in each language, the documents are segmented in paragraphs; • the alignments French/Ukrainian and English/Ukrainian are generated, in which nth paragraph from one language is associated with the nth paragraph from the other language; • the alignment between the two pairs of languages is then verified manually.
In Figure 2, we present an excerpt from the English/Ukrainian aligned corpus.
Then in French and English, we can use the
existing terminology extraction tools which results
bootstrap the acquisition of bilingual terminology.
Hence, we use the YATEAterm extractor
The exploitation of the MedlinePlus parallel and aligned corpus is performed in several ways (Figure 3).
Transfer 1 First, the simplest situation is when the two aligned lines contain term candidates in either language: these terms are recorded as candidates for the alignments. For instance, in Figure 2, the pairs {Tiredness, Втома} and {Pain, Біль} are issued from this kind of alignment.
Transfer 2 Secondly, when the paragraphs
contain complex expressions or sentences, the
processing is done as follows (Figure 4):
1. the paragraph-aligned corpora are aligned at
the word level using GIZA++
For instance, in Figure 2, the term Cancer cells is automatically extracted from the English corpus. GIZA++ proposes that Cancer cells is aligned with Ракові клітини. Thus, through the wordaligned text, we can propose that Cancer cells is the translation of Ракові клітини. This processing is performed on the two pairs of languages (French/Ukrainian and English/Ukrainian).
As indicated in Table 1, the size of our corpora is rather small for the statistical alignment performed by GIZA++. For this reason, we provide GIZA++ with a bilingual dictionary in order to help the alignment at the word level (see Section 3.3). Besides, in preliminary experiments, we also observe that word level alignment errors lead to the extraction of Ukrainian stopwords as term candidates (на (on), або (or), etc.). To remove such obvious errors, we filter out such candidates if they occur in a list of 385 stop-word forms issued from an existing resource dedicated to the localization of graphical interfaces2. 3.2
The Wikipedia corpus is used to complete and to
help the method applied to the MedlinePlus
corpus. The content we propose to exploit is included
in infoboxes (on the right in Figure 1) and is
reachable through the MediaWiki source code of the
Wikipedia. This provides the label of the
medical terms in Ukrainian and their MeSH codes. The
process is the following (Figure 4):
1. the infobox content is extracted and parsed3
in order to obtain the term label and its MeSH
code,
glish terms,
ogy.
2. the MeSH code is used to query the UMLS,
and to get the corresponding French and
En3. the term pairs French/Ukrainian and
English/Ukrainian are then built and provide
good candidates for the bilingual
terminolThis part of the method exploits specific and
intentionally created content for a given medical
notion in Ukrainian: term for a given medical
notion and its MeSH code. This information is
reliable. For instance, in Figure 1, the term нанізм
is extracted, as well as its MeSH code D004392.
Through the UMLS, the corresponding English
terms are dwarfism and nanism, while the
corresponding French term is nanisme. Notice that
similar method has been used for the building of
medical terminology in the Arabic language
)
• the single-word terms extracted by other approaches can be provided to GIZA++, as an additional bilingual dictionary, in order to help the alignment of MedlinePlus at the word level.
During preliminary experiments, we test several combinations of parameters for the pre-processing and the alignments.
French corpus, the Part-of-Speech is performed by
TreeTagger
Thus, based on the results of the preliminary experiments, we choose to pre-process the English corpus with TreeTagger and the French corpus with TreeTagger and Flemm.
the Giza++ word level alignment. We only present the results obtained with this configuration in the following. 3.4
The evaluation is performed manually in order to check whether the candidates extracted for building the bilingual terminologies are correct. It has been performed by an Ukrainian native speaker having knowledge in medical informatics. Terms are validated independently in each language, but we also evaluation the bilingual and trilingual relations between the Ukrainian, English and French terms. With this kind of evaluation, precision of the results can be computed, i.e. the ratio between the correct answers and all the answers. 4
Table 2 presents the results and the precision for the extracted terms by the three methods. Table 3 presents the results and the precision concerning the pairs and triples of terms. 4.1
The exploitation of the Wikipedia infobox allow to collect 357 Ukrainian medical terms among which 177 are single-word terms. By querying UMLS with the MeSH codes, those terms are associated with 1428 French terms (among them, 339 single-word terms) and 3625 English terms (among them, 448 single-word terms). The number of French and English terms compared to the number of Ukrainian terms are due to the synonyms proposed by MeSH. As for the bilingual pairs of terms, we obtain 1,515 Ukrainian/French term pairs and 3,789 Ukrainian/English term pairs, including, respectively, 270 and 405 pairs between single-word terms. Since each Ukrainian term is associated with at least one French and English terms, this allows to build 28,840 triples. We consider that the precision of this terminology is 1 because the collecting manner. 4.2
The use of the first method of transfer (Transfer 1) allows to extract 436 Ukrainian terms with a high precision unsurprisingly (0.966). These terms are associated with 316 French terms and 354 English terms in 282 triples between Ukrainian, French and English terms, 63 pairs only between Ukrainian and French terms and 115 pairs only between Ukrainian and English terms, with 0.954, 0.937 and 0.965 precision, respectively. Thus, the Transfer 1 method allows to collect 334 Ukrainian/French term pairs (among them 108 pairs between single-word terms) and 380 Ukrainian/English term pairs (among them 135 pairs between single-word terms). We observe that these relations can involve synonyms in either language: {фаллопієва труба, trompes de fallope/trompe utérine} (fallopian tube), {втрата слуху/втрачається слух, hearing loss}, {втома, fatigue/tiredness}. Besides, in Ukrainian, several case forms can be associated to a same English of French form: {вагітність, pregnancy} and {вагітності, pregnancy}.
As the precision values suggest, this first transfer method leads to few errors. Their analysis shows that they mainly concern partial match between one language and another involved by the translation: {появу виразок у роті, mouth sores} -- lit. (appearance of) mouth sores, {ви можете спати, dormir/sleep} -- lit. you can sleep. The silence of the method can be explained by two reasons. First, again the variation due to the translation prevents the transfer 1 method to extract term in French or English. For instance, since the title Soins in the French corpus is the English translation of Your care, the French term matches with the line, contrary to the English term. The Transfer 2 method will solve this problem. However, the main reason of the silence is the incapacity of the term extractor to identify French or English terms because its extraction strategy or errors in the POS tagging.
As for the second transfer method (transfer 2), we present the results obtained when the pairs of single-words terms issued from the MedlinePlus corpus and from Wikipedia are used to help the GIZA++ alignment. In that context, the transfer 2 method allows to extract 9,040 Ukrainian terms with 0.454 precision (exact match). Precision of the French and English terms is higher: 0.674 and 0.761 respectively (exact match). Moreover, the number of French and English terms is dramatically lower (about -45% and -40%) than in Ukrainian: the rich morphology of the Ukrainian language provides several inflected Prec. 1 0.966 0.998 0.454 0.84 0.481 Source Wikipedia MedlinePlusT ransfer1
inexact match MedlinePlusT ransfer2
inexact match
forms for a given term ({напад, нападу} -- attack, {припадків, припадки} -- seizure, {костей, кістки} -- bones). Besides, the method allows also to extract synonymous terms ({приступам, припадків} -- attacks/seizures, {биття, удару} -- beats). The precision values with the inexact match (the correct term is included or includes the term candidates) are much higher and gain 0.40 points for the Ukrainian terms and 0.05 for the French and English terms. We assume this difference on Ukrainian candidate terms is mainly due to the alignment quality. As for the interlingual relations, the Transfer 2 method collects 3,724 pairs of Ukrainian/French terms with 0.309 precision, 4,745 pairs of Ukrainian/English terms with 0.401 precision and 4,724 triples with 0.419 precision.
An analysis of the results shows that most of the errors are due to the alignment problems. Indeed, we observe that when the alignment is correct, the Ukrainian terms are correctly extracted by the transfer. Otherwise, the errors occur.
Moreover, even if the documents (patientoriented brochures) are not highly specialized, most of the extracted terms are specific to the medical domain ({трахеотомією, tracheostomy}), {фактори ризику, risk factors}, {шприца, syringe}, {холестерину, cholesterol}). Other terms also refer to close and approximating notions which reflects this type of documents: {діти, children}, {здорову їжу, healthy diet}, {серцевий напад, heart attack}, {склянок рідини, glasses of liquid}. An interesting observation is that some French and English terms correspond to propositions in Ukrainian: {не до кінця приготовлену їжу, undercooked foods} (lit. food which is not fully cooked), {При цьому обстеженні Ви не відчуєте жодного болю, indolore (painless)} (lit. With this exam you will feel no pain).
Finally, all the methods combined allow to build a terminological resource containing 4,588 Ukrainian medical terms and their 34,267 relations with French and English terms. 5
In this work, we propose to exploit two kinds of freely available multilingual corpora in French, English and Ukrainian. Each corpus is exploited with appropriate methods which allows to extract the term candidates and to create term pairs Ukrainian/French and Ukrainian/English. In particularly, French and English corpora are processed with NLP and term extraction tools. Then, thanks to the transfer methods these terms are transposed on the Ukrainian language. We also propose to use existing terminologies and to exploit simple terms for improving the alignment performed at the word level with GIZA++.
Our future work will address the enrichment of
the created resource with terms from other
corpora. Besides, in the Wikipedia corpus, we can use
other codes, such as those from МКХ-10 (ICD10)
or MedlinePlus. This will also augment the
coverage of the term pairs extracted in the current work.
Another perspective of this work is the
improvement of the bilingual alignment of documents at
the word level. In that respect, we plan to
investigate the use of other alignment algorithms, such as
Fast-Align
This work is funded by the LIMSI-CNRS AI project Outiller l'Ukranien. We are thankful to the reviewers for their useful comments which permitted to improve the quality of the paper.