The Comprehension of Medical Words

Cross-lingual Experiments in French and Xhosa

Natalia Grabar

, Izak van Zyl

, Retha de la Harpe

and Thierry Hamon

3,4

CNRS UMR 8163 STL, Universit

e Lille 1&3, 59653 Villeneuve d’Ascq, France

Cape Peninsula University of Technology, Cape Town, South Africa

LIMSI-CNRS, BP133, Orsay, France

Universit

e Paris 13, Sorbonne Paris Cit

e, France

Keywords:

Health Literacy, Readability, Consumer Health Informatics, Natural Language Processing, Xhosa, French.

Abstract:

This paper presents cross-lingual experiments in automatic detection of medical words that may be difﬁcult to

understand by patients. The study relies on Natural Language Processing (NLP) methods, conducted in three

steps, across two languages, French and Xhosa: (1) the French data are processed by NLP methods and tools

to reproduce the manual categorization of words as understandable or not; (2) the Xhosa data are clustered

with a non-supervised algorithm; (3) an analysis of the Xhosa results and their comparison with the results

observed on the French data is performed. Some similarities between the two languages are observed.

1 INTRODUCTION

Health information is integral to individual lifeways

in light of pressing health concerns and the pursuit of

maintaining healthy lifestyles. Moreover, health in-

formation is already widespread in society and is dis-

seminated through a variety of media, scientiﬁc and

medical research, news, the internet, radio and TV

programs. However, the availability of this informa-

tion does not guarantee its correct understanding and

use. Standard medical and health language indeed

conveys very specialized and technical notions elabo-

rated and usually disseminated by healthcare profes-

sionals. These notions often remain opaque and non

understandable for non-expert users, and especially

for patients (Berland et al., 2001). This is despite the

fact that they may have an important inﬂuence on the

success of patients’ medical care and their quality of

life. For such reasons, the difﬁculty related to the ac-

cess and use of health information must be addressed

and surmounted. This extends to a process of en-

hanced communication between medical profession-

als and patients; a better understanding of the medi-

cal care delivered to patients; better management of

chronic diseases (AMA, 1999).

Another aspect is related to multilingual and mul-

ticultural contexts in societies and in hospitals. Typi-

cally, health care providers operate within distinct lan-

guage and cultural modalities, which may often di-

verge from the patient context. This disparity usually

impedes communication leading to patient dissatis-

faction and potential misdiagnosis and inappropriate

medication prescriptions. Existing studies have ad-

dressed, inter alia, Spanish-speaking communities in

North America (Woloshin et al., 1995; Flores et al.,

1998), and Xhosa-speaking communities in South

Africa (Levin, 2006b; Levin, 2006a; Schlemmer and

Mash, 2006). Miscommunication and language dis-

parities reveal additional complexities as they are em-

bedded within cultural, lingual, and specialized and

technical aspects of communication.

The possibility of simpliﬁed health information

use has been poorly researched up to now in the Com-

puter Sciences ﬁeld. The studies of the kind are done

in connection within the Consumer Health Informat-

ics. They join at least two aspects: (1) Semantic

Interoperability (through study and creation of rela-

tions between expert and non-expert languages in or-

der to improve the communication), and (2) Com-

puter Sciences and Disability, because Computer Sci-

ences provide methods for helping patients with un-

derstanding the health information (Zeng and Par-

manto, 2003). Other studies address the automatic

distinction between specialized and non-specialized

documents and words (Miller et al., 2007; Franc¸ois

and Fairon, 2013), and focus on the building of expert

and non-expert aligned medical vocabularies (Zeng

334

Grabar N., van Zyl I., de la Harpe R. and Hamon T..

The Comprehension of Medical Words - Cross-lingual Experiments in French and Xhosa.

DOI: 10.5220/0004803803340342

In Proceedings of the International Conference on Health Informatics (HEALTHINF-2014), pages 334-342

ISBN: 978-989-758-010-9

 2014 SCITEPRESS (Science and Technology Publications, Lda.)

et al., 2006; Elhadad and Sutaria, 2007; Del

eger and

Zweigenbaum, 2008) in one given language (i.e., En-

glish and French in the cited studies), or on the build-

ing of bilingual patient-oriented French-English ter-

minology (Messai et al., 2006). Also, a speciﬁc

task has been proposed withinin the SemEval chal-

lenge aiming the substitution of words by their sim-

pler equivalents in general language texts in English

(Specia et al., 2012). In the present study, we pro-

pose to focus on the distinction between comprehen-

sible and non-comprehensible words from the medi-

cal ﬁeld in two languages from two distant families,

Xhosa and French. We rely on the Computer Sciences

and on Natural Language Processing (NLP) methods

and tools. The main hypothesis is that, when work-

ing with different languages and applying different

methods and resources, it remains possible to observe

some similarities between the languages.

In the following of this study, we ﬁrst present the

background and objectives of our study (section 2),

we then describe the material used (section 3), and the

methods we propose to achieve the objectives (section

4). We then discuss the obtained results (section 5),

and present conclusions (section 6).

2 BACKGROUND

2.1 Speciﬁcities of the Two studied

Languages

We indicate here some speciﬁcities and differences

between Xhosa and French. Xhosa is a Bantu lan-

guage, spoken in South Africa. It is a tonal language,

which means that the same sequence of consonants

and vowels can have different meanings when said

with a rising or falling or high or low intonation.

Tone is contrastive in Xhosa, with several examples

of words that differ only in tone (Van der Stouwe,

2009): goba can mean be happy or put lotion on,

and imithi can mean its pregnant or trees according

to tones. Furthermore, Xhosa is an agglutinative lan-

guage, which means that preﬁxes and sufﬁxes are at-

tached directly to root words. For instance, in abant-

wana bayadlala, the noun ntwana (child) is modiﬁed

by the plural afﬁx aba-, while the verb dlala (play)

is modiﬁed by the afﬁx baya- meaning third plural

person in present tense. The Xhosa words belong

mainly to two syntactic classes (nouns and verbs),

while the afﬁxes (preﬁxes and sufﬁxes) convey addi-

tional meanings, like in the presented example. From

the point of view of NLP, the Xhosa language has

recently attracted attention of researchers (Allwood

et al., 2003; Roux et al., 2004; Moropa, 2007; Bosch

et al., 2008; Pretorius and Bosch, 2009), but up to now

there is a paucity of freely available resources and

tools. French is a Latin language, spoken in Europe

and in countries across the world. French grammar

organizes words into several syntactic classes (nouns,

verbs, adjectives, adverbs, prepositions...). Its words

can be modiﬁed with inﬂectional ({abdomen, ab-

domens}), derivational ({abdomen, abdominal}) and

compositional ({abdomen, abdominoplastie}) pro-

cesses. French is not an agglutinative language:

words and syntactic classes remain free in the sen-

tences, although they have deﬁned places there. The

semantic ambiguity of words is mainly due to the in-

herent semantics of their morphological bases. For

instance, montre means a watch and some inﬂectional

forms of the verb show; while poste means post ofﬁce

and employment position. The French language has

been the object of several NLP research studies and is

provided with several resources (dictionaries, lexica,

terminologies, etc.) and tools.

2.2 Rationale of the Study

Given the differences that exist between the two stud-

ied languages, it is difﬁcult to apply the same methods

for processing data from each. For this reason, the

French data are processed with supervised methods

and a sophisticated set of linguistic features involving

several resources and NLP tools (section 4.1), while

the Xhosa data are processed with unsupervised meth-

ods and poorer linguistic data (section 4.2). In Table

1, we summarize the rationale of the study: in Xhosa,

the unsupervised methods are applied to the surface

forms of words which linguistic description relies on

the contexts of these words; while in French, the su-

pervised methods are applied to the lemmas of words

which are provided with a rich linguistic description.

Table 1: Rationale of the methods for Xhosa and French.

Xhosa French

Methods unsupervised supervised

Units forms lemmas

Features contexts rich set of features

2.3 Objectives

We have several objectives: (1) analyze how the dis-

tinction between technical and understandable words

can be done automatically; (2) work with multilingual

data from the medical ﬁeld; (3) apply different meth-

ods to the different languages but obtain comparable

results; (4) study how and to which extent we can

TheComprehensionofMedicalWords-Cross-lingualExperimentsinFrenchandXhosa

335

enrich the linguistic description of the less resourced

language.

3 MATERIAL

We use two datasets related to the medical ﬁeld lan-

guage, one in French and one in Xhosa. We also use

two lexica (English and bilingual Xhosa-English).

3.1 Building and Preprocessing the

French Linguistic Data

In French, the data are obtained from the medical ter-

minology Snomed International (C

e, 1996), in its

version currently distributed by ASIP Sant

. We use

this terminology because its aim is to describe the en-

tire medical ﬁeld: this provides us with the possibility

to study the main medical notions extensively. Be-

sides, the terms recorded in Snomed Int are usually

extracted from real clinical documents and often cor-

respond to real expressions used in the language, and

with which the patients may be faced.

Snomed contains 151,104 terms structured into

eleven semantic axes. Among these, we study

ﬁve taxonomies related to the main medical notions

(disorders, abnormalities, procedures, functions, and

anatomy). The corresponding terms (104,649) are to-

kenized and segmented into words. We obtain 29,641

unique words: they correspond to our main material

and we use them in order to study how the complex-

ity of specialized words is felt by the speakers and

how it can be described and processed with auto-

matic approaches. The preprocessing of words con-

sists of automatic assignment to a syntactic category

(Schmid, 1994; Namer, 2000), such as noun (ag-

nosie, ad

enome), adjective (bacteroide, D-d

ependant,

abact

erien), verb (alimenter, baver, changer), adverb

(essentiellement, crescendo, facilement), preposition

or determinant; and their lemmatization that provides

canonical form of words (singular for nouns, mascu-

line singular for adjectives, inﬁnitive for verbs). In

the following, all the experiments in French are per-

formed on lemmas. These linguistic data are also

annotated according to the aims of our study: three

speakers without any medical training, considered as

laymen, are involved. They are asked to analyze the

29,641 words and to assign them to one of the three

categories:

1. I can understand the word;

2. I am not sure about its meaning;

http://esante.gouv.fr/asip-sante

3. I cannot understand the word.

The assumption is that the words that cannot be un-

derstood by the annotators should be considered as

semantically difﬁcult. These manual annotations cor-

respond to the reference data.

3.2 Building and Preprocessing the

Xhosa Linguistic Data

In Xhosa, the linguistic data are mainly obtained from

brochures created for patients (prevention, pathology

and treatments for HIV, rape and other medical con-

ditions). They are freely available and are collected

from several websites

. These documents are con-

verted from PDF to text format to make NLP process-

ing possible. The ﬁnal corpus contains 34 documents

and over 206,000 tokens of words. This corpus corre-

sponds to the main material exploited in Xhosa. The

corpus is further preprocessed:

• tokenizing the punctuation, i.e. separation of dots,

commas, which is the ﬁrst step of normalization;

• recognition and ﬁltering out sentences in English,

as in the exploited documents some sentences in

English may occur.

3.3 Lexica

The lexicon of English words is built using the Robert

& Collins dictionary. It contains 398,229 entries. It is

used to ﬁlter out sentences in English. A bilingual

Xhosa-English lexicon is built to help interpret the re-

sults generated from the Xhosa corpus. It is built us-

ing the data available on different websites

, and from

the inventory of Xhosa plant names (Dold and Cocks,

1999). It is also completed with word pairs

while

analyzing the clusters. The resulting lexicon contains

6,594 entries, including agglutinated forms.

4 METHODS

The proposed method is composed of three steps: (1)

processing of the French data with supervised meth-

ods and choice of the descriptors suitable for the au-

tomatic distinction between understandable and non-

understandable words; (2) exploitation of the Xhosa

corpus with unsupervised methods and building of

clusters of words; (3) interpretation of clusters.

www.tac.org.za, www.doh.gov.za, www.capetown.gov.

www.travlang.com, sabelo.tripod.com/dictionary.htm,

http://www.dicts.info

http://mokennon.albion.edu/

HEALTHINF2014-InternationalConferenceonHealthInformatics

336

4.1 Processing the French Data

In French, the task is addressed as classiﬁcation prob-

lem for the automatic distinction between understand-

able and non-understandable words. Supervised ma-

chine learning method is applied: the annotations pro-

duced by the annotators are used as training data for

the creation of speciﬁc models, and as reference data

for evaluation. This process relies on speciﬁc set

of features (section 4.1.1), machine learning (section

4.1.2), and evaluation protocol (section 4.1.3).

4.1.1 Generation of the Features

We exploit 24 features computed automatically.

These features can be grouped into ten classes.

Syntactic Categories. Syntactic categories and

lemmas are computed by TreeTagger (Schmid, 1994)

and checked by Flemm (Namer, 2000). The syntactic

categories are assigned to words within the context

of their terms. If a given word receives more than

one category, the most frequent one is kept. Among

the main categories we ﬁnd nouns, adjectives, proper

names, verbs and abbreviations.

Presence of Words in Reference Lexica. We

exploit two reference lexica of the French language:

TLFi (TLFi, 2001) and lexique.org

. TLFi is a dic-

tionary of the French language covering XIX and

XX centuries, with almost 100,000 entries. lex-

ique.org is a lexicon created for psycholinguistic ex-

periments, with over 135,000 entries, among which

almost 35,000 lemmas. These two lexica are ex-

pected to represent the common lexical competence

of speakers and we suppose that those words that are

present in these lexica should be more familiar.

Frequency of Words through a Mon-

specialized Search Engine. We query a non-

specialized search engine in order to know its

frequency attested on the web. Those words that are

more frequent are expected to be easier to understand.

Frequency of Words in Medical Terminology.

We compute the frequency of words in the medi-

cal terminology, Snomed International. Similarly, we

suppose that words that are more frequent there can

be less difﬁcult to understand by layman speakers.

Number and Types of Semantic Categories as-

sociated to Words. We exploit information on the se-

mantic categories of Snomed International: we expect

that words that belong to several categories may con-

vey more fundamental medical notions and be better

known by speakers.

Length of Words in Number of their Charac-

ters and Syllables. We compute the number of char-

http://www.lexique.org/

acters and syllables, and expect that longer words are

potentially more difﬁcult to understand, because they

can correspond to lexically complex lexemes.

Number of Bases and Afﬁxes. Lemmas are an-

alyzed by the morphological analyzer D

erif (Namer,

2009). It performs their decomposition into bases and

afﬁxes known in its database. Here again, we expect

that morphologically more complex lemmas may cor-

respond to semantically more complex lexemes.

Initial and Final Substrings of the Words. We

compute the initial and ﬁnal substrings (three to ﬁve

characters). We expect that these substrings may be

evocative of the bases or afﬁxes positioned at the be-

ginning and, especially, at the end of words. The main

motivation is that ﬁnal substrings correspond to the

semantic base of compounds, often Latin of Greek

components.

Number and Percentage of Consonants, Vowels

and other Characters. We also compute the number

and the percentage of consonants, vowels and other

characters (for instance, hyphen, apostrophe, comas

such as they occur in names of chemical products).

Classical Readability Scores. We apply two clas-

sical readability measures: Flesch (Flesch, 1948) and

its variant Flesch-Kincaid (Kincaid et al., 1975). Typ-

ically used for evaluating the difﬁculty level of texts,

they exploit surface characteristics of words (number

of characters and/or syllables) and normalize these

values with speciﬁc coefﬁcients. Longer words are

considered to be more difﬁcult to understand.

4.1.2 Machine Learning System

Table 2: Number (and percentage) of words assigned to ref-

erence categories within the majority set.

Categories Number %

1. I can understand 7,655 27

2. I am not sure 597 2

3. I cannot understand 20,511 71

Total annotations 28,763 100

Machine learning is used in order to classify the data

and to distinguish between comprehensible words

among laymen, and also to study the importance of

various features for the task. The machine learning

exploits an annotated dataset, that is described with

suitable features such as those presented above. On

the basis of such features, the algorithms can detect

the regularities within the training dataset to gener-

ate a model and apply the generated model to process

new unseen data. We apply several algorithms avail-

able in WEKA (Witten and Frank, 2005).

The annotations, provided by the three annotators,

constitute our reference data. We use here the dataset

TheComprehensionofMedicalWords-Cross-lingualExperimentsinFrenchandXhosa

337

majority (Table 2) that contains the annotations for

which we can compute the majority agreement of the

annotators i.e., at least two of the annotators agree.

This dataset contains 28,763 words (out of 29,641),

among which 71% are assigned to I cannot under-

stand, 27% to I can understand and only 2% to I am

not sure: the non-comprehensible words are the most

frequent. According to the Fleiss’ Kappa (Fleiss and

Cohen, 1973), suitable for processing the data pro-

vided by more that two annotators, the inter-annotator

agreement shows substantial agreement (Landis and

Koch, 1977), with the score 0.73. This corresponds to

a very good agreement level, especially when work-

ing with linguistic data, for which the agreement is

usually difﬁcult to obtain, as it greatly depends on the

individual linguistic feeling of the speakers.

4.1.3 Evaluation

The success of the machine learning algorithms is

evaluated with three classical measures: recall R

(how exhaustive are the results?), precision P (how

correct are the results?), and F-measure F (harmonic

mean of P and R). We perform a ten-fold cross-

validation. In the perspective of our work, these mea-

sures help evaluate the suitability of the methodol-

ogy to the distinction between words understandable

or not by layman speakers and the relevance of the

chosen features to the aimed task. The baseline cor-

responds to the assignment of words to the biggest

category, e.g., I cannot understand, which represents

71%. We can also compute the gain, which is the ef-

fective improvement of performance P given the base-

line BL (Rittman, 2008):

P−BL

1−BL

4.2 Processing the Xhosa Data

The Xhosa corpus is processed with distributional

methods (Harris, 1968; Brown et al., 1992): they

aim at grouping words that share the same or simi-

lar contexts. As an example, symptom and pain may

be grouped together because they share several com-

mon contexts as they appear in the neighborhood of

words such as relieve, appear or treatment. It is as-

sumed that such groups of words also have some se-

mantic relations among them, although these relations

are not semantically typed. The context is deﬁned as

co-occurrence window (n words before and/or after a

given word). This context is exploited to compute the

association strength between words within the win-

dow: it is usually based on frequencies or Mutual In-

formation. A similarity measure (i.e., Jaccard or Co-

sine) is computed and allows to group words together

(Curran, 2004). We apply an implementation of the

Brown algorithm

with the following parameters:

• corpus content with and without punctuation;

• English text ﬁltered out or not;

• normalization to lower-cased characters or not;

• setting up the minimal number of occurrences of

words, within the interval [1, 2, 3];

• setting up the number of clusters to be generated

within the interval [50, 100, 150, 200, 250...1000,

1500, 2000, 2500].

4.3 Interpreting the Xhosa Results

The distributional methods generate clusters which

words share common contexts and possibly common

semantics. We expect that the clusters we generate

with these methods may contain: (1) words belonging

to the same syntactic classes (i.e., verbs, nouns); (2)

words playing similar syntactic roles (i.e., preposi-

tional phrases); (3) words with similar semantics (i.e.,

general language words, pathologies, treatments); (4)

or even words that represent similar comprehension

levels (i.e., easy to understand, difﬁcult to under-

stand). This last possibility, unsupervised automatic

distinction between words that represent similar com-

prehension levels, would make a direct parallel be-

tween French and Xhosa data. However, given the

difference of the source data (simple contexts and fre-

quencies with Xhosa data, and a set of 24 sophisti-

cated features with French data), we cannot expect to

achieve this possibility with the currently exploited

methods and resources.

For interpreting the Xhosa results, the generated

clusters are annotated using the bilingual lexicon.

Two kinds of annotations are performed:

1. direct: those cluster words that exist in the lexicon

are provided with their English translations;

2. indirect: the words that are not recorded in the

lexicon are checked for a surface likeness with

words from the lexicon. For this, the number of

character deletions, insertions and reversals from

a given cluster word to obtain a given lexicon

word are counted (Levenshtein, 1966). For in-

stance, ewonke is not part of the lexicon, but

wonke, konke and zonke are. The cost to transform

ewonke into these lexicon words is 1, 2 and 2, re-

spectively. We have indeed to delete one character

e to obtain wonke, and we have to delete one char-

acter e and replace one character w→k to obtain

konke. The maximal cost is set to 2, which means

that the three candidate translations detected for

http://cs.stanford.edu/ pliang/software/

HEALTHINF2014-InternationalConferenceonHealthInformatics

338

ewonke are acceptable. The meaning of the three

candidate translations (e.g., all), can be then trans-

posed onto the ewonke word.

5 RESULTS AND DISCUSSION

5.1 Relevant Features for French

Table 3: Performance and gain obtained for F by J48.

P R F BL gain

Majority 0.876 0.889 0.881 0.71 0.16

Performance of the J48 algorithm is among the

best obtained: we use it to present the results. Be-

sides, this algorithm provides the decision tree, which

allows analyzing the features exploited for the classi-

ﬁcation. Performance of J48 on the majority dataset is

indicated in Table 3: P 0.876, R 0.889, and F 0.881.

The gain we obtain is 0.16 point by comparison with

the baseline. These are good results and indicate that

the chosen features are relevant to the purpose of the

study. A more detailed analysis of the inﬂuence of the

individual features indicate that:

• with the syntactic categories alone we obtain P

and R between 0.65 and 0.7;

• semantic axes of SNOMED Int decrease precision

(be frequent in Snomed Int does not mean to be

easier to understand) but improve recall;

• presence of words in the reference lexica is bene-

ﬁcial to both precision and recall;

• the frequencies of the lexemes on the general

search engine are often beneﬁcial;

• the sufﬁxes with the three- and four-character

length (i.e., omie, phie,

emie) have a positive im-

pact, but the sufﬁxes with the ﬁve-character length

negatively impact the results;

• among the features taht negatively impact the re-

sults, we ﬁnd also readability scores (especially

the Flesch-Kincaid score) and number and per-

centage of consonants;

• the remaining features have no or very small im-

pact on the performance.

Notice that features, such as frequency on the search

engine or presence in the reference lexica, proved to

be efﬁcient in SemEval contest (Specia et al., 2012).

5.2 Generation of Clusters for Xhosa

In Table 4, we give quantitative information on clus-

ters generated with the Xhosa data while the number

Table 4: Quantitative information on clusters generated

with the Xhosa data (minimal frequency set to 2).

Nb clusters 1000 1500 2000 2500

Min/cluster 1 1 1 1

Max/cluster 60 60 54 115

Average 17.89 11.93 8.95 7.16

of clusters is set to 1000, 1500, 2000 or 2500, and

the minimal frequency of words is set to 2. The to-

tal number of words is then 17,890 (punctuation tok-

enized and removed, English text kept, lower-cased

data). We indicate minimal, maximal and average

number of words per cluster. We can observe that,

logically, the average number of words per cluster is

going decreasing with a higher number of clusters.

Clusters containing only one word are frequent. Ex-

ceptionally, we can also generate clusters with a large

number of words (i.e., 115 with 2500 clusters). In

the following, we indicate and discuss the results for

1500 clusters. These provide a good compromise of

the size of clusters and their content: they are neither

too exclusive nor too inclusive. The cluster words are

mapped to the lexicon. With 1500 clusters, we obtain

5,873 direct and 12,017 indirect mappings. Among

the general observations, we can notice that:

• when the English sentences are kept, the English

words are usually clustered together in separate

clusters, which shows the efﬁciency of unsuper-

vised approaches for language recognition. To

give an idea of the English words in the cor-

pus: when all words are taken into account (mini-

mal frequency 1, with upper-cased characters), we

ﬁnd 50,421 words among which 8,401 are in En-

glish; with the minimal frequency set to 2, we ﬁnd

19,405 words among which 3,402 are in English;

• when upper and lower-cased words are used to-

gether, the same words written differently (i.e.,

iya and Iya) are usually grouped together, which

means that their contexts are similar and that their

regularity can be observable in both cases.

Concerning more semantic aspects of the clusters,

the results indicate that:

• a lot of clusters gather verbal forms, among which

those dedicated to verbs with meaning to go/come

or to get/give are very frequent, i.e.:

– ayiyi (it doesn’t go) and iya (it goes), or ndiza

(I come), sisiya (we go) and uza (he comes);

– bakunike (they gave it to you), amnike (they

gave for him), lunike (give it), zinike (give for

them) and afumane (they got);

• cluster words may share common semantics, like:

TheComprehensionofMedicalWords-Cross-lingualExperimentsinFrenchandXhosa

339

– the notion of improvement, with words such

as kuyanceda (it helps), kuthenjelwe (you can

hope); or kuncede (it helps you), kuxhaswe

(you support), unyangwe (he or she was cured),

kuthintelwe (prevent), uhlale (he or she lived);

– the notion of movement inside, with words such

as kungena (it goes into), ebunzimeni (they are

in the mass), zifakwa (they are put in);

– the notion of interaction, with words such as

bajongane (look at each other), kuxoxwe (dis-

cuss it), and ukuqondwa (understand);

• moreover, some clusters contain medical no-

tions, such as macrophage and ziingxaki (they

are the problems), zigulane (sick each other) and

ngozi (danger); kunengozi (it has the danger)

and mfuneko (it’s a requirement/need); agcine

(they saved), ndincede (help me) and kunyanga (it

cures). Sometimes, these are related to the im-

provement notion.

The currently obtained results on Xhosa are in ac-

cordance with our expectations on their content: we

can automatically detect and group together words

with general, grammatical and medical notions. The

current method cannot distinguish between compre-

hensible and non-comprehensible words. This means

that such distinction cannot rely only on the contexts

of words, but requires additional information such as

those used in the study on the French data. For this

reason, work done on the Xhosa data should be con-

sidered as preliminary. Still, the English words are

usually clustered separately: typically, such words

can represent difﬁcult notions for Xhosa-speaking

community (Levin, 2006b; Schlemmer and Mash,

2006). Notice that these results are similar to what we

observe on French data: words borrowed from Latin

and Greek, or words that are morphologically com-

plex and that contain Latin or Greek bases, are usu-

ally felt to be non-understandable by French speak-

ers. Further observation of the difﬁculty of the Xhosa

words will require additional methods and rely on a

speciﬁc judgment of native speakers.

5.3 Comparison with Existing Studies

Some ﬁndings of the proposed work can be compared

with existing studies: relevance of some features used

in French (frequency, presence in the reference lexica)

to the readability diagnosis (Specia et al., 2012). Still,

existing work is usually applied to English data, while

processing of French and of less resourced languages

(like Xhosa) is rare or even non-existing. Moreover,

usually data from the general language are processed

and little interest is paid to medical language. For

such reasons, it remains difﬁcult to fully compare the

proposed study with existing work. We can neverthe-

less rely on the SemEval contest and on the work done

in French to improve the method applied to Xhosa and

to design a similar method for this language.

6 CONCLUSIONS AND FUTURE

WORK

We proposed experiments in French and Xhosa lan-

guages with the main objective to detect comprehen-

sible and non-comprehensible words from the medi-

cal ﬁeld automatically. The material, resources and

methods used in both languages are different (non su-

pervised clusters of words in Xhosa, supervised cate-

gorization of French words), which logically lead to

different results. Nevertheless, we can do similar ob-

servations in both languages concerning the detection

of borrowed and foreign words (e.g., Latin and Greek

words and morphological components in French, En-

glish words in Xhosa), which appear to be difﬁcult

to understand for native speakers without training in

medicine. It appears that the distinction between

comprehensible and non-comprehensible words can-

not rely only on the contexts of words: additional in-

formation and exploitation of supervised methods are

necessary. The results obtained for Xhosa are prelim-

inary from this point of view.

Future studies should help explicate these avenues

and analyze more complete data. For instance, we

plan to study the content of clusters and to address the

understandability level of the Xhosa words with the

native speakers. We also plan to use more sophisti-

cated clustering approaches and to apply a supervised

approach for processing the Xhosa data. For perform-

ing this last issue, we need to compute a reasonable

set of suitable features, some of which are suggested

by the work performed in French and in SemEval con-

test: frequency of words in the studied corpus or on

the web, their frequency in a reference corpus or in

a more technical corpus if available, morphological

analysis of words (Bosch et al., 2008; Pretorius and

Bosch, 2009) and their lexical complexity, complexity

of words computed at the character level (i.e., number

of characters and syllables, readability scores adapted

to the Xhosa language), common contexts of words.

ACKNOWLEDGEMENTS

Experiments on the French data have been done in

part within the MESHS project COMETE.

HEALTHINF2014-InternationalConferenceonHealthInformatics

340

REFERENCES

Allwood, J., Gr

onqvist, L., and Hendrikse, A. (2003). De-

veloping a tag set and tagger for the african languages

of South Africa with special reference to Xhosa.

Southern African Linguistics and Applied Language

Studies, 21(4):223–237.

AMA (1999). Health literacy: report of the council on sci-

entiﬁc affairs. Ad hoc committee on health literacy

for the council on scientiﬁc affairs, American Medi-

cal Association. JAMA, 281(6):552–7.

Berland, G., Elliott, M., Morales, L., Algazy, J., Kravitz,

R., Broder, M., Kanouse, D., Munoz, J., Puyol, J.,

Lara, M., Watkins, K., Yang, H., and McGlynn, E.

(2001). Health information on the internet. accessi-

bility, quality, and readability in english ans spanish.

JAMA, 285(20):2612–2621.

Bosch, S., Pretorius, L., and Fleisch, A. (2008). Experi-

mental bootstrapping of morphological analysers for

Nguni languages. Nordic Journal of African Studies,

17(2):66–88.

Brown, P., deSouza, P., Mercer, R., Della Pietra, V., and

Lai, J. (1992). Class-based n-gram models of natural

language. Computational Linguistics, 18(4):467–479.

e, R. (1996). R

epertoire d’anatomopathologie de la

SNOMED internationale, v3.4. Universit

e de Sher-

brooke, Sherbrooke, Qu

ebec.

Curran, J. R. (2004). From distributional to semantic simi-

larity. PhD thesis, University of Edinburgh.

Del

eger, L. and Zweigenbaum, P. (2008). Paraphrase acqui-

sition from comparable medical corpora of specialized

and lay texts. In AMIA 2008, pages 146–50.

Dold, A. and Cocks, M. (1999). A preliminary list of Xhosa

plant names from the Eastern Cape, South Africa.

Bothalia, 29:267–292.

Elhadad, N. and Sutaria, K. (2007). Mining a lexicon of

technical terms and lay equivalents. In BioNLP.

Fleiss, J. and Cohen, J. (1973). The equivalence of weighted

kappa and the intraclass correlation coefﬁcient as

measures of reliability. Educational and Psycholog-

ical Measurement, 33:613–619.

Flesch, R. (1948). A new readability yardstick. Journal of

Applied Psychology, 23:221–233.

Flores, G., Abreu, M., Olivar, M., and Kastner, B. (1998).

Access barriers to health care for latino children. Arch

Pediatr Adolesc Med, 152:1119–1125.

Franc¸ois, T. and Fairon, C. (2013). Les apports du TAL

a la lisibilit

e du franc¸ais langue

etrang

ere. TAL,

54(1):171–202.

Harris, Z. S. (1968). Mathematical Structures of Language.

Wiley, New York, NY, USA.

Kincaid, J., Fishburne, R. J., Rogers, R., and Chissom, B.

(1975). Derivation of new readability formulas (auto-

mated readability index, fog count and ﬂesch reading

ease formula) for navy enlisted personnel. Technical

report, Naval Technical Training, U. S. Naval Air Sta-

tion, Memphis, TN.

Landis, J. and Koch, G. (1977). The measurement of ob-

server agreement for categorical data. Biometrics,

33:159–174.

Levenshtein, V. I. (1966). Binary codes capable of correct-

ing deletions, insertions and reversals. Soviet physics.

Doklady, 707(10).

Levin, M. (2006a). Different use of medical terminol-

ogy and culture-speciﬁc models of disease affecting

communication between Xhosa-speaking patients and

English-speaking doctors at a South African paedi-

atric teaching hospital. S Afr Med J, 96:1080–1084.

Levin, M. (2006b). Language as a barrier to care for Xhosa-

speaking patients at a South African paediatric teach-

ing hospital. S Afr Med J, 96:1076–1079.

Messai, R., Zeng, Q., Mousseau, M., and Simonet, M.

(2006). Building a bilingual french-english patient-

oriented terminology for breast cancer. In MedNet.

Miller, T., Leroy, G., Chatterjee, S., Fan, J., and Thoms, B.

(2007). A classiﬁer to evaluate language speciﬁcity of

medical documents. In HICSS, pages 134–140.

Moropa, K. (2007). Analysing the English-Xhosa parallel

corpus of technical texts with Paraconc: a case study

of term formation processes. Southern African Lin-

guistics and Applied Language Studies, 25(1):183–

205.

Namer, F. (2000). FLEMM : un analyseur ﬂexionnel du

franc¸ais

a base de r

egles. Traitement automatique des

langues (TAL), 41(2):523–547.

Namer, F. (2009). Morphologie, Lexique et TAL :

l’analyseur D

eriF. TIC et Sciences cognitives. Her-

mes Sciences Publishing, London.

Pretorius, L. and Bosch, S. (2009). Exploiting cross-

linguistic similarities in Zulu and Xhosa computa-

tional morphology. In AFLAT, pages 96–103.

Rittman, R. (2008). Automatic discrimination of genres.

VDM, Saarbrucken, Germany.

Roux, J., Louw, P., and Niesler, T. (2004). The African

speech technology project: An assessment. In LREC,

pages 93–96.

Schlemmer, A. and Mash, B. (2006). The effects of a lan-

guage barrier in a South Africa district hospital. S Afr

Med J, 96:1084–1087.

Schmid, H. (1994). Probabilistic part-of-speech tagging us-

ing decision trees. In Proceedings of the International

Conference on New Methods in Language Processing,

pages 44–49, Manchester, UK.

Specia, L., Jauhar, S., and Mihalcea, R. (2012). Semeval-

2012 task 1: English lexical simpliﬁcation. In *SEM

2012, pages 347–355.

TLFi (2001). Tr

esor de la Langue Franc¸aise - I. IN-

aLF/ATILF. Disponible l’adresse www.tlfi.fr.

Van der Stouwe, C. (2009). A phonetic and phonological

report on the Xhosa language. Technical report. Ac-

cessed 1 October 2013, http://bit.ly/1bZwt1j.

Witten, I. and Frank, E. (2005). Data mining: Practical

machine learning tools and techniques. Morgan Kauf-

mann, San Francisco.

Woloshin, S., Bickell, N., Schwartz, L., Gany, F., and

Welch, H. (1995). Language barriers in medicine in

the united states. JAMA, 273(9):724–728.

Zeng, Q. T., Tse, T., Divita, G., Keselman, A., Crow-

ell, J., and Browne, A. C. (2006). Exploring lexical

TheComprehensionofMedicalWords-Cross-lingualExperimentsinFrenchandXhosa

341

forms: ﬁrst-generation consumer health vocabularies.

In AMIA 2006, pages 1155–1155.

Zeng, X. and Parmanto, B. (2003). Evaluation of web ac-

cessibility of consumer health information websites.

In AMIA 2003, pages 743–7.

HEALTHINF2014-InternationalConferenceonHealthInformatics

342