Identification of Fertile Translations in Medical Comparable Corpora :
a Morpho-Compositional Approach

Compositional translation relies on the principle of compositionality which states that “the meaning of the whole is a function of the meaning of the parts” [Keenan and Faltz, 1985, 24-25]. Applied to bilingual lexicon extraction, compositional translation ($\mathcal{CT}$) consists in decomposing the source term into atomic components ($\mathcal{D}$), translating these components into the target language ($\mathcal{T}$), recomposing the translated components into target terms ($\mathcal{R}$) and finally filtering the generated translations with a selection function ($\mathcal{S}$):

$\mathcal{CT}\text{(``ab'')}$
$~=\mathcal{S(R(T(D(}\text{``ab''}))))$
$~=\mathcal{S(R(T(}\text{\{a, b\}})))$
$~=\mathcal{S(R(}\{\mathcal{T}\text{(a)}\times\mathcal{T}\text{(b)}\}))$
$~=\mathcal{S(R(}\text{\{{a, b}\}}))$
$~=\mathcal{S(}\text{\{{a, b}\}, \{{b, a}\}})$
$~=\text{``{ba}''}$
 

where “ab” is a source term composed of a and b, “ba” is a target term composed of b and a and there exists a bilingual resource linking a to a and b to b.

Existing implementations differ on the kind of atomic components they use for translation.

Lexical compositional translation [Grefenstette, 1999, Baldwin and Tanaka, 2004, Robitaille et al., 2006, Morin and Daille, 2010] deals with multi-word term to multi-word term alignment and uses lexical words³³3as opposed to grammatical words: preposition, determiners, etc. as atomic components : rate of evaporation is translated into French taux d’évaporation by translating rate as taux and evaporation as évaporation using dictionary lookup. Recomposition may be done by permutating the translated components [Morin and Daille, 2010] or with translation patterns [Baldwin and Tanaka, 2004].

Sublexical compositional translation deals with single-word term translation. The atomic components are subparts of the source single-word term. Cartoni (2009) translates neologisms created by prefixation with a special formalism called Bilingual Lexeme Formation Rules. Atomic components are the prefix and the lexical base: Italian neologism anticonstituzionale ’anticonstitution’ is translated into French anticonstitution by translating the prefix anti- as anti- and the lexical base constituzionale as constitution. Weller et al. (2011) translate two types of single-word term. German single-word term formed by the concatenation of two neoclassical roots are decomposed into these two roots, then the roots are translated into target language roots and recomposed into an English or French single-word term, e.g. Kalori₁metrie₂ is translated as calori₁metry₂. German Noun₁-Noun₂ compounds are translated into French and English Noun₁Noun₂ or Noun₁ Prep Noun₂ multi-word term, e.g. Elektronen_N1-mikroskop_N2 is translated as electron_N1 microscope_N2.

Compositional translation faces four main challenges which are (i) morphosyntactic variation: source and target terms’ morphosyntactic structures are different: anti-cancer${}_{\textsc{Noun}}$ $\rightarrow$ anti-cancéreux${}_{\textsc{Adj}}$ ’anti-cancerous’ ; (ii) lexical variation: source and target terms contain semantically related - but not equivalent - words: machine translation $\rightarrow$ traduction automatique ’automatic translation’ ; (iii) fertility: the target term has more content words than the source term: isothermal snowpack $\rightarrow$ manteau neigeux isotherme ’isothermal snow mantel’ ; (iv) terminological variation: a source term can be translated as different target terms: oophorectomy $\rightarrow$ ovariectomie ’oophorectomy’, ablation des ovaires ’removal of the ovaries’.

Solutions to morphosyntactic, lexical and to some extent terminological variation have been proposed in the form of thesaurus lookup [Robitaille et al., 2006], morphological derivation rules [Morin and Daille, 2010], morphological variant dictionaries [Cartoni, 2009] or morphosyntactic translation patterns [Baldwin and Tanaka, 2004, Weller et al., 2011]. Fertility has been addressed by Weller et al. (2011) for the specific case of German Noun-Noun compounds.

Morpho-compositional translation (morpho-compositional translation) relies on the following assumptions:

Lexical subcompositionality. The lexical items which compose a multi-word term or a single-word term may be split into semantically-atomic components. These components may be either free (i.e. they can occur in texts as autonomous lexical items like toxicity in cardiotoxicity) or bound (i.e. they cannot occur as autonomous lexical items, in that case they correspond to bound morphemes like -cardio- in cardiotoxicity).

Irrelevance of the bound/free feature in translation. Translation occurs regardless of the components’ degree of freedom: -cardio- may be translated as cœur ’heart’ as in cardiotoxicity $\rightarrow$ toxicité pour le cœur ’toxicity to the heart’.

Irrelevance of the bound/free feature in allomorphy. Allomorphy happens regardless of the components’ degree of freedom: -cardio-, cœur ’heart’, cardiaque ’cardiac’ are possible instantiations of the same abstract component and may lead to terminological variation as in cardiotoxicity $\rightarrow$ cardiotoxicité ’cardiotoxicity’, toxicité pour le cœur ’toxicity to the heart’, toxicité cardiaque ’cardiac toxicity’.

Like other sublexical approaches, the main idea behind morpho-compositional translation is to go beyond the word level and work with subword components. In our case, these components are morpheme-like items which either (i) bear referential lexical meaning like confixes⁴⁴4we use the term confix as a synonym of neoclassical roots (Latin or Ancient Greek root words). (-cyto-, -bio-, -ectomy-) and autonomous lexical items (cancer, toxicity) or (ii) can substantially change the meaning of a word, especially prefixes (anti-, post-, co-…) and some suffixes (-less, -like…). Unlike other approaches, morpho-compositional translation is not limited to small set of source-to-target structure equivalences. It takes as input a single morphologically constructed word unit which can be the result of prefixation ’pretreatment’, confixation ’densitometry’, suffixation ’childless’, compounding ’anastrozole-associated’ or any combinations of the four. It outputs a list of $n$ words who may or may not be morphologically constructed. For instance, postoophorectomy may be translated as postovariectomie ’postoophorectomy’, après l’ovariectomie ’after the oophorectomy’ or après l’ablation des ovaires ’after the removal of the ovaries’.

As an example, we show the translation of the adjective cytotoxic into French using a toy dataset. Let $Comp^{l}_{type}$ be a list of components in language $l$ where $type$ equals $pref$ for prefixes, $conf$ for confixes, $suff$ for suffixes and $free$ for free lexical units ; $Trans$ be the translation table which maps source and target components ; $Var^{l}$ be a table mapping related lexical units in language $l$ ; $Stop^{l}$ a list of stopwords in language $l$ ; $Corpus^{l}$ a lemmatized, pos-tagged corpus in language $l$: $Comp^{en}_{conf}=\{\text{-cyto-}\}$ ; $Comp^{en}_{free}=\{\text{cytotoxic, cytotoxicity, toxic}\}$ ; $Comp^{fr}_{conf}=\{\text{-cyto-}\}$ ; $Comp^{fr}_{free}=\{\text{cellule, toxique}\}$ ; $Trans=\{\{\text{-cyto-}\rightarrow\text{-cyto-},\text{cellule}\},$       $\{\text{toxic}\rightarrow\text{toxique}\}\}$ ; $Var^{en}=\{\text{cytoxic}\rightarrow\text{cytoxicity}\}$ ; $Stop^{fr}=\{\text{pour},\text{le}\}$ ; $Corpus^{fr}=\text{``le/{det} cytotoxicit\'{e}/{n} \^{e}tre/{aux} le/{det}}$ propriété/n de/prep ce/det qui/pro être/aux toxique/a pour/prep le/det cellule/n ./pun” ; ’The cytotoxicity is the property of what is toxic to the cells.’

Morpho-compositional translation takes as input a source language single-word term and outputs zero or several target language single-word terms or multi-word terms. It is the result of the sequential application of four functions to the input single-word term: decomposition ($\mathcal{D}$), translation ($\mathcal{T}$), recomposition ($\mathcal{R}$) and selection ($\mathcal{S}$).

Our corpus is composed of specialized texts from the medical domain dealing with breast cancer. We define specialized texts as texts being produced by domain experts and directed towards either an expert or a non-expert readership [Bowker and Pearson, 2002]. The texts were manually collected from scientific papers portals and from information websites targeted to breast cancer patients and their relatives. Each corpus has approximately 400k words (cf. table 1). All the texts were pos-tagged and lemmatized using the linguistic analysis suite Xelda⁵⁵5http://www.temis.com. We also computed the comparability of the corpora. We used the comparability measure defined by [Bo and Gaussier, 2010] which indicates, given a bilingual dictionary, the expectation of finding for each source word of the source corpus its translation in the target corpus and vice-versa. The English-French corpus’ comparability is 0.71 and the English-German corpus’ comparability is 0.45. The difference in comparability can be explained by the fact that German texts on breast cancer were hard to find (especially scientific papers): we had to collect texts in which breast cancer was not the main topic.

We tested our algorithm on a set of source terms extracted from the English texts. The extraction was done in a semi-supervised manner. Step 1: We wrote a short seed list of English bound morphemes. We automatically extracted from the English texts all the words that contained these morphemes. For example, we extracted the words postchemotherapy and poster because they contained the string post- which corresponds to a bound morpheme of English. Step 2: The extracted words were sorted : those which were not morphologically constructed were eliminated (like poster), and those which were morphologically constructed were kept (like postchemotherapy). The morphologically constructed words were manually split into morphemes. For example, postchemotherapy was split into post-, -chemo- and therapy. Step 3: If some bound morphemes which were not in the initial seed list were found when we split the words during step 2, we started the whole process again, using the new bound morphemes to extract new morphologically constructed words. We also added hyphenated terms like ER-positive to our list of source terms.

We obtained a set 2025 English terms with this procedure. For our experiments, we excluded from this set the source terms which had a translation in the general language dictionary and whose translation was present in the target texts. The final test set for English-to-French experiments contains 1839 morphologically constructed source terms. The test set for English-to-German contains 1824 source terms.

Tables 2 and 3 show the size of the resources we used for translation.

General language dictionary We used the general language dictionary which is part of the linguistic analysis suite Xelda.

Domain-specific dictionary We built this resource automatically by extracting pairs of cognates from the comparable corpora. We used the same technique as [Hauer and Kondrak, 2011]: a SVM classifier trained on examples taken from online dictionaries⁶⁶6http://www.dicts.info/uddl.php.

Morpheme translation table To our knowledge, there exists no publicly available morphology-based bilingual dictionary. Consequently, we asked translators to create an ad hoc morpheme translation table for our experiment. This morpheme translation table links the English bound morphemes contained in the source terms to their French or German equivalents. The equivalents can be bound morphemes or lexical items.

In order to handle the variation phenomena described in section 2.3, we used a dictionary of synonyms and lists of morphologically related words. The dictionary of synonyms is the one part of the Xelda linguistic analyzer. The lists of morphologically related words were built by stemming the words of the comparable corpora and the entries of the bilingual dictionary with a simple stemming algorithm [Porter, 1980].

The decomposition function uses the entries of the bound morphemes translation table (242 entries) and a list of 85k lexical items composed of the entries of the general language dictionary and English words extracted from the Leipzig Corpus [Quasthoff et al., 2006] which is a general language corpus.

As explained in section 2.2, compositional translation consists in generating candidate translations. These candidate translations can be filtered out with a classifier [Baldwin and Tanaka, 2004], by keeping only the translations which occur in the target texts of the corpus [Weller et al., 2011, Morin and Daille, 2010] or by using a search engine [Robitaille et al., 2006]. Unlike alignment evaluation in parallel texts, there is no reference alignmens to which the selected translations can be compared and we cannot use standard evaluation metrics like AER [Och and Ney, 2000]. It is also difficult to find reference lexicons in specific domains since the goal of the extraction process is to create such lexicons. Furthermore, we also wish to evaluate if the algorithm can identify non-canonical translations which, by definition, can not be found in a reference lexicon. Usually, the candidate translations are annotated manually as correct or incorrect by native speakers. Baldwin and Takana (2004) use two standards for evaluation: gold-standard, silver-standard. Gold-standard is the set of candidate translations which correspond to canonical, reference translations. Silver-standard corresponds to the gold-standard translations plus the translations which “capture the basic semantics of the source language expression and allow the source language expression to be recovered with reasonable confidence” (op. cit.).

The first evaluation metric is the precision $P$ which is the number of correct candidate translations $|Corr|$ over the total number of generated candidate translations $|A|$: $P=\frac{|Corr|}{|A|}$. In addition to precision, we propose to indicate the coverage $C$ of the lexicon, i.e. the proportion of source terms (st) which obtained at least one candidate translation regardless of its accuracy:

$$C=\frac{\sum^{|\textsc{st}|}_{i=1}\alpha(\textsc{st}_{i})}{|\textsc{st}|}$$

were $\alpha(\textsc{st}_{i})$ returns $1$ if $|A(\textsc{st}_{i})|\geq 1$ else $0$. As augmenting coverage tends to lower precision, we also compute $OQ$, the overall quality of the lexicon, to get an idea of the coverage/precision tradeoff: $OQ=P\times C$.

Compositional-translation methods give better results when they are applied to general language texts rather than domain-specific texts. This is due to the fact that the translations of the components can be easily found in dictionaries since they belong to the general language and it is also easier to collect large corpora. Working with general language texts, Baldwin and Takana (2004) were able to generate candidate translations for 92% of their source terms and they report 43% (gold-standard) to 84% (silver standard) of correct translations. The size of their corpus exceeds 80M words for each language. Cartoni (2009) works on the translation of prefixed Italian neologisms into French. He considers that the generated neologisms have a “confirmed existence” if they occur more than five times on Internet. He finds that between 42% and 94% of the generated neologisms fall into that category. Regarding domain-specific translation, Robitaille et al. (2006) use a search engine to build corpus from the web and incrementally collect translation pairs. They start with a list of 9.6 pairs (on average) with a precision of 92% and end up with a final output of 19.6 pairs on average with a precision of 81%. Morin and Daille (2009) could generate candidate translations for 15% of their source terms and they report 88% of correct alignments. The size of their corpus is 700k words per language. Weller et al. (2011) were able to generate 8% of correct French translations and 18% of correct English translations for their 2000 German compounds. Their corpus contains approximately 1.5M words per language.

We ran the morpho-compositional translation prototype on the set of source terms described in section 4.2. The output candidate translations were manually annotated by two translators. Like Baldwin and Takana (2004), we used three annotation values: canonical translation, recoverable translation and incorrect. In our case, recoverable translations correspond paraphrastic and morphological translation variants. For example, the canonical translation for post-menauposal is post-ménopausique. Recoverable translations are post-ménopause ’post-menopause’ and après la ménopause ’after the menopause’. Fertile translations can be canonical translations if a non-fertile translation would have been more awkward. For example, the canonical translation for oestrogen-sensitive is sensible aux œstrogènes ’sensitive to oestrogens’. A non-fertile translation would sound very unnatural. We computed inter-annotator agreement on a set of 100 randomly selected candidate translations. We used the Kappa statistics [Carletta, 1996] and obtained a high agreement (0.77 for English to German translations and 0.71 for English to French).

First, we tested the impact of the linguistic resources described in section 4.3 (B for Baseline dictionaries, D for Domain-specific dictionary, S for Synonyms, M for Morphologically related words). We also tested a simple Prefix+lemma translation (Pref) in similar vein to the work of Cartoni (2009) to serve as a line of comparison with our method. The results are given in tables 4 and 5. The best results in terms of overall quality are obtained with the combination of the baseline and domain-specific dictionaries (BD). Morphologically related words and synonyms increase coverage to the cost of precision. Regarding English-to-French translations, we were able the generate translations for 26% of the source terms. The gold-standard precision is 60% and the silver standard precision is 67%. Regarding English-to-German translations, we were able the generate translations for 26% of the source terms. The gold-standard precision is 39% and the silver-standard precision is 43%. The prefix+lemma translation method has a very high precision (between 84% and 76%) but produces very few translations (between 1% and 2%). Coverage and precision scores compare well with other approaches knowing that we have very small domain-specific corpora (400k words per language) and that our approach deals with a large number of morphological constructions. The lower quality of the German translations can be explained by the fact that the English-German corpus is much less comparable than the English-French corpus (0.45 vs. 0.71).

We also tested the impact of the fertile translations on the quality of the lexicon. Tables  6 and  7 show the evaluation scores with and without fertile translations. As expected, fertile translations enables us to increase the size of the lexicon but they are less accurate than non-fertile translations. Fertile translations increase the overall quality of the English-French lexicon by 4% to 5%. This is not the case for English-German translations: fertile translations result in a big drop in precision. The overall quality does not significantly change. This might be partly due to the low comparability of the corpus but we think that the main reason lies in the morphological type of the languages involved in the translation. It is worth noticing that, if we consider only the non-fertile translations, the English-German lexicon has generally better scores than the English-French one. In fact, fertile variants are more natural and frequent in French than in German. English and German are Germanic languages with a tendency to build new words by agglutinating words or morphemes into one single word. Noun compounds such as oestrogen-independent or Östrogen-unabhängige are common in these two languages. Conversely, French is a Romance language which prefers to use phrases composed of two nouns and a preposition rather than a single-noun compound (oestrogen-independent would be translated as indépendant des œstrogènes ’independent to oestrogens’). It is the same with the bound morpheme/single word alternance. The term cytoprotection will be translated into German as Zellschutz whereas in French it can be translated as cytoprotection or protection de la cellule ’protection of the cell’.