Extraction of restricted lexical combinations by detecting non-compositionality of multiword expressions

EXTRACTION OF RESTRICTED LEXICAL COMBINATIONS BY DETECTING NON-COMPOSITIONALITY OF MULTIWORD EXPRESSIONS EXTRAÇÃO DE COMBINAÇÕES LEXICAIS RESTRITAS PELA DETEÇÃO DA NÃO COMPOSIONALIDADE DE EXPRESSÕES PLURIVERBAIS 1 Joana Veloso UNIVERSIDADE DO MINHO, PORTUGAL [email protected] Alberto Simões UNIVERSIDADE DO MINHO, PORTUGAL [email protected] Álvaro Iriarte S. UNIVERSIDADE DO MINHO, PORTUGAL [email protected] In this article an evaluation of a method for extracting restricted lexical combinations from parallel corpora by detecting non-compositionality of multiword expressions in translation will be presented. This method presupposes that by finding sequences of words whose translation does not follow a simple word-to-word conversion of the component words, a collocation is probably present. Word bigrams are used. Keywords: Collocations, Translation, Parallel Corpora Neste artigo apresentamos uma avaliação sobre um método para extrair combinações lexicais restritas a partir de corpora paralelos, pela deteção da não composicionalidade de expressões pluriverbais na tradução. Este método baseia-se na presunção de que, encontrando sequências de palavras cuja tradução não siga a tradução palavra por palavra dos seus componentes, é provável estar-se perante uma colocação. São usadas palavras brigrama Palavras-chave: Colocações, Tradução, Corpora Paralelos 0. Introduction Collocations can be defined in very different ways according to different authors, and the same sequence of words can be considered or not a collocation by different researchers, even when using a similar definition. A great number of linguists define “collocation” in terms of frequency. That is the criterion used by many authors within the computational linguistics community for corpus-based automatic collocation extraction. However, other linguists (especially lexicologists and lexicographers) believe that the fact that two lexemes frequently co-occur in context is not a 1 We are grateful to the anonymous reviewers for their valuable comments. good enough reason to be considered a collocation. Mel’čuk et al. (1995: 51) state: «Kjellmer 1994 illustre un essai d’extration automatique de collocations d’un corpus informatisé sans intervention décisionnelle d’un lexicologue; ce dictionnaire est rempli d’expressions comme Mr. Smith, was a member, the abilities, a bad thing, etc. qui n’ont rien à voir avec les collocations. » Coseriu denies that a lexeme combination is a “lexical solidarity” only because words are often combined. For Coseriu, a lexical solidarity is explained by lexical restrictions based on the linguistic content of a lexeme that drive it to combine with other lexemes. «… la probabilidad estadística de las combinaciones no tiene prácticamente nada que ver con las solidaridades y no es prueba de su existencia: cavallo bianco es, probablemente, más frecuente que cavallo sauro; pero, en el primer caso, la probabilidad de la combinación depende de la realidad extralinguística; en el segundo, en cambio, está dada lingüísticamente, por el contenido de sauro.» (Coseriu, 1977: 160). There are software tools that allow the automatic extraction of frequent lexical combinations, however, these tools do not distinguish between collocations, idioms, free phrases, compound names, or multiword technical terms. It should be noted that “restricted lexical combinations” (especially collocations) with “frequent combinations” of two or more lexemes must not be confused. It is true that a collocation is also a frequent combination of two or more lexemes but frequent lexical combinations are not always collocations. A lexical combination such as ler um livro is a frequent combination, but it is a completely free phrase, a combination of words formed according to the syntactic rules (the verb ler can be combined with everything capable of being read). However, the adjectives in lexical combinations such as atividade febril, mudança radical, vontade louca, ódio mortal, amor cego are fixed, although they have the same meaning (intense, a lot). These lexical combinations are frequent because the choice of adjective is not free, that is, the fact that these two lexemes frequently appear combined is not the cause but the consequence of them being a collocation (Alonso Ramos, 1993). However, from a lexicographical point of view, these frequent lexical combinations, although they are not collocations in the narrow sense that we use, should have special treatment in dictionaries and lexical databases. In this regard, Cowie´s distinction between restricted collocations and open collocations is useful (Cowie et al.,1984). It will, therefore, be appropriate, in lexicographical practice, to recognize the existence of more or less free/restricted lexical combinations and to register usual or frequent free combinations in dictionaries (Ettinger, 1982; Corpas, 1995). As we know, stability and semantic specialization are the main features of this type of non-free phrases. Clearly syntactic constraints will be higher for idioms than for collocations. For example, the idiom perder a cabeça has more syntactic constraints than the collocation prestar atenção, which allows some changes (see Aguilar-Amat, 1993: 67-68). However, from a lexicographical point of view, it is impossible to establish general syntactic rules for all idioms and collocations. The criteria that allow us to determine if a lexical combination was lexicalized cannot be morphological or syntactic in nature. This has more to do with the consensus and the memory of the linguistic community that uses it: «Le critère ultime de définition d’une unité lexicale est bien ici, par excellence, le consensus de la communauté linguistique […], non pas comme en syntaxe ou en morphologie par la reconnaissance d’une bonne formation mais sur la base de la mémorisation.» (Paillard, 1997: 66). Ultimately, when it comes to distinguishing between free phrases and collocations, in addition to the more or less intuitive perceptions of speakers, the use of a foreign language serves to illustrate that choosing a collocative for a base of collocation is not free (Caldern, 1994: 80; Tomaszczyk, 1983: 45). Therefore, we are of the opinion that the use of a foreign language can be useful to extract collocations from parallel corpora. For the purpose of this work, our definition of collocation is: if any of the member words or the complete sequence inherit a different meaning of its/their usual sense when used in conjunction with another word, then this sequence of words is considered to be a collocation. Our assumption is that, given a multiword expression tA and its translation tB, tA is a collocation if tB includes words which are not direct translations of any of tA words. To test this assumption, we chose a mid-sized corpus, the European Central Bank corpus from the Per-Fide project (Araújo, Almeida, Simões & Dias, 2010). For the languages, we chose the Spanish/Portuguese language pair. The main reason for this choice is the proximity of the languages, and the bilingual translation dictionaries we had available. We expect to make further tests on this hypothesis with other language pairs in the future, namely including Germanic languages. As for the text to be analyzed, we selected Spanish word pairs in which one of the words is an adjective and the other is a noun (in any order). For this purpose, the FreeLing (Simões & Carvalho, 2012; Padró & Stanilovsky, 2012) morphological analyzer was used. For each word pair, each possible word translation was looked up in the translation segment. This was possible with the help of a Spanish/Portuguese translation dictionary Apertium (Corbí-Bellot et al., 2005). If any of the possible translations for both words occur together (in any order) in the Portuguese part, that word pair is discarded. On the other hand, if only one of the words has a translation in the Portuguese part, the Spanish segment is stored, together with a snippet of the Portuguese translation. The next section explains this algorithm in more detail. The word pairs obtained, together with the respective translation, were manually evaluated for whether they are, or not, restricted combinations. The evaluation section will discuss the details on the manual assessment of the obtained results, explaining the main problems found as well as the future enhancements to the proposed algorithm. Finally, a set of conclusions is drawn from the obtained results. 1. Related Work The task of identifying restricted lexical combinations, as we will state in this article, is not new. It is a relevant procedure for different tasks on Natural Language Processing (NLP) like, for example, Machine Translation (MT), where idiomatic expressions cannot be translated literally. Even collocations need to be translated with caution. For example, computer graphics cannot be translated into Portuguese as gráficos de computador, not because it is a wrong translation, but because the term that was coined in the Portuguese community was computação gráfica (*graphics’ computation). The easiest way to detect sequences of words likely to be considered as a collocation or, at least, as a compound term, is to use the Mutual Information (MI) or Pointwise Mutual Information (PMI) metrics. As this is just an association measure, authors use these indicators together with other techniques, like the usage of patterns (Guinovart & Simões, 2009). However, by themselves, these two measures are not enough for the extraction of collocations. A prior study (Pavel 2005) presents a vast amount of measures that can be used to detect collocations. Nevertheless, most of them perform badly by themselves, and as presented below, new approaches have been used. Probably the bigger challenge is to detect idiomatic expressions, mostly when they can also have a literal meaning (like break the ice that can be considered literally or not). Li & Sporleder (2009) present a set of different properties that can be extracted from texts in order to detect if these expressions are being used literally. Properties are very diverse, from the usage of prepositions before or after the expression, to graphs of cohesion between the different sentence components. These properties are then used in a Support Vector Machine algorithm. These same authors (Li & Sporleder, 2010) also worked on the use of Gaussian Mixture Models for this same task. Their evaluation points to 92% precision for the detection of literal expressions, but only 42% to detect non-literal (idiomatic) expressions. Muzny & Zettlemoyer (2013) also use classification techniques to distinguish between idiomatic and non-idiomatic expressions. For that, they trained a binary perceptron based on two types of features: lexical features, like the usage of capital letters, and graph features, using relations information obtained from WordNet and Wiktionary. The perceptron was training on Wiktionary labeled data, and used the non-labeled data for test purposes. The results go up to 65% of precision, and recall over 52%. The most relevant study found using multilingual information for the detection of collocation extraction (Seretan & Wehrli, 2006), does not use translation information, but only a parser able to process text in different languages. The extraction method, itself, does not take any real advantage of parallel corpora. Our approach understands the translation as a function that can, somehow, associate “semantic” to each word. Therefore, if the translation (“semantic”) is not compositional, we have a candidate collocation. 2. The Hypothesis The hypothesis we are testing is: if a sequence with two words, an adjective and a noun, is translated by two other words, and only one of them is a translation of the original words found in a translation dictionary, then we have a candidate collocation. This can be better explained using mathematical syntax. Let us define the T function, that translates Spanish words into Portuguese, and the concatenation operator (a dot), which joins two words. The translation of two words wa and wb is considered to be compositional if (1) T(wa . wb) = T(wa) . T(wb) or even, T(wa . wb) = T(wb) . T(wa) Therefore, we are looking for a pair of words (wa , wb) in which one of them is an adjective and the other a noun, and whose translation does not follow the equation presented above (1). That is, we want to find wa and wb where (2) T(wa . wb) = T(wa) . wc ∧ T(wa . wb) = wc . T(wb) with T(wa) ≠ wc ∧ T(wb) ≠ wc. The extraction algorithm used is very simple, and its main purpose is to test the hypothesis that the collocation extraction based on non-translation composition is possible. The algorithm starts by iterating over each translation unit in the parallel corpus. A translation unit is composed by a segment SSP for the Spanish language, and a segment SPT for the Portuguese language. Then, each possible bigram from the segment SSP is analyzed using the FreeLing morphological analyzer, looking for a sequence in which one of the words is a noun and the other an adjective. Note that, although FreeLing has modules to do part-of-speech tagging they were not used. Nevertheless, we are aware of the problems this approach arises, and we will discuss them later. When such a pair of words is found, their possible translation sets are computed. Note that each word can have more than one translation, and, therefore, we need to construct a set of translations. This translation was done using the Apertium translation dictionary. Then, these translation sets are searched in the target language segment SPT. If any of the words from both translation sets occur together, the word pair is discarded. On the other hand, if one of the words has a translation in the target segment, but the other does not, the Spanish word pair is saved for manual assessment. Together with the word pair, a segment of Portuguese words in the vicinity of the found translation is seized and also stored. This list was then assessed manually. 3. Assessments and Evaluation The assessment was performed manually using online resources as reference, such as IATE (InterActive Terminology for Europe) (Johnson & Macphail, 2000), and both paper and online Spanish-Portuguese and Portuguese-Spanish dictionaries. A Linguistics MSc student classified each word pair manually into one of the following classes: ● Error: used for all entries whose Spanish and Portuguese segments are not related with each other. This happens mainly because the application was not able to find the sequence of words that include the translation of the selected pair of words, or because the original corpus had alignment errors; ● Free combination: the pair of words is correctly translated, but it is not a restricted lexical combination (accordingly with the criterion we defined earlier). This happens mostly when a possible word translation is not included in the used translation dictionary; ● Restricted combination: the pair of words is correctly translated, and it corresponds to a collocation or a quasi-phraseme. When in doubt about a combination being considered restricted or free, we took the decision to consider it as a free combination. This means that our evaluation is less favorable to our hypothesis. From here on, we will discuss each class, providing real examples for each of them. 3.1. Errors The errors found are from very different kinds, such us from alignment problems, some minor bugs in the algorithm implementation, or the lack of translations from the translation dictionary: ● The use of a morphological analyzer instead of a part-of-speech tagger leads to some examples with verbs misclassified as nouns and/or adjectives. Nevertheless, considering that our hypothesis is the existence of a sequence with a noun and an adjective, the examples classified as a result of this problem are irrelevant for proving it. Table 1 2 shows some of these situations. Table 1: Examples of extractions where a verbal form was mistakenly interpreted as a noun. anexo figura presente anexo figura um modelo conjunto presente Se o conjunto apresentar informe figura este relatrio consta de o certificado falla verificao de o certificado falhar ● The algorithm, when searching the set of words in the context of the found translation broke the segment, losing the interesting part of the translation. This turned the assessment impossible. This was a problem inherited from the bad segmentations performed by other tools like the segmenter, tokenizer and the sentence-aligner. For example, the alignment for “actividades pesqueras” computed by the algorithm was “definitiva de as actividades de *”. Given the missing word (marked by the asterisk) this segment could not be classified correctly, and therefore, fell in the error class. Just like with the case above, we do not have any detail on the validity (or not) of the hypothesis. Table 2 shows further examples of this segmentation problem. All these cases can be safely ignored for the hypothesis test. Between parentheses we show the missing words. This seemed to be a problem on the corpus segmentation and alignment process. Table 2: Examples of truncated segments. 2 In these tables, the left column is the Spanish extracted pair, and to the right is the segment extracted from the Portuguese side. In italics we give emphasis to the translation of the terms from the first column. tabaco crudo Sector de o tabaco em (cru) derechos humanos promoo de os direitos de (o homem) productos pesqueros mercado de os produtos de (pesca) Seguridad Alimentaria Europeia para a Segurana de (a alimentação) Seguridad Alimentaria Europeia para a Segurana de (a alimentar) ● The algorithm implementation is not prepared to find all occurrences of the word translations. This means that, if two similar pairs occur (like “cantidad superior” and “calidad superior”) the algorithm will use the first translation pair twice (aligning “qualidade superior” with “cantidad superior” and not the correct “calidad superior”). This is, indeed, a bug introduced by our implementation, but when it was detected it was too late to perform a complete new extraction and restart the manual evaluation. Therefore, they were ignored for our hypothesis test. Table 3 shows some of these examples. In italics, on the right, the aligned segment. Table 3: Examples of misaligned segments. Comunidad Econmica que institui a Comunidade Europeia cantidad superior em uma quantidade inferior tiempo completo de trabalho a tempo parcial tercera columna referidas em a coluna 2 navegacin marítima afectos a a navegao area ● The translation, sometimes, uses a pronoun to refer to a noun used on a previous sentence, while the original sentence repeats the noun. See Table 4 for some examples. Table 4: Examples of misalignments resulted from the use of pronouns. Estado membro legislao de esse Estado ciertos productos regime de esses produtos . valores límite ou de esses valores , ltimo caso . Em esse caso , segundo Estado legislao de esse Estado ; ● Some translation units were not really translated. In some cases the Portuguese version included the text in Spanish, and in some other, in English, as shown in Table 5. Some others, as shown in Table 6, include typos that, not being in the dictionary, activated our hypothesis by mistake. Table 5: Examples of translation units where at least one of the sides is untranslated. medio ambiente contaminacin del medio ambiente Autoridades nacionales Lista de las autoridades nacionales medio ambiente en el medio ambiente acuático Vivo Test In Vivo Test for Chromosomal Table 6: Examples of translations with minor typos, mistakenly extracted as collocations. presente artículo de o presnete artigo legítimo titular seu legtimo titual a ocupar zona geográfica específicas em uma zona geogrfirca presente Reglamento O presidente regulamento é proyectos transnacionales acompanhamento de os projectos trannacionais 3.2 Free Combinations The main interference with the algorithm, which could make it extract free combinations, is the lack of translations from the used translation dictionary. When a word is not found in the dictionary (either as the word not existing in the source language －Table 7; or the target language do not include the used translation －Table 8), the algorithm considers the translation to be incorrect, and therefore, it can be used for our hypothesis. A similar problem occurred with words not correctly lemmatized, and therefore, not found in the translation dictionary. Table 7: Basic examples where the dictionary lacked a entry for one of the words. Segundo resultado resultado : segundo resultado : primer trimestre em o primeiro trimestre de presente Directiva requisitos de a presente directiva primeros párrafos os dois primeiros pargrafos podem presente apndice o presente apndice , os Table 8: Basic examples where the dictionary lacked one of the synonyms. Texto pertinente EEE ) ( Texto relevante siguiente texto a seguinte redaco : Medidas vigentes Medidas existentes ltima fabricacin a ltima data de fabrico Conocimientos sucintos ; Conhecimentos sumrios de as Finally, there is yet another problem, related with the textual deixis, where there is a reference to a different position in the text that, different languages refer to in different ways. Examples are cuadro anterior/quadro acima, fórmula anterior/fórmula acima and criterios anteriores/critérios acima. As anterior and acima are not direct translations, the algorithm extracted them as restricted combinations (although they are free combinations). 3.3 Restricted Combinations and Collocations Other than the correct restricted combinations, there are two special kinds that should be mentioned: ● There are situations where the pair of words in Spanish has a single word translation in Portuguese, either because in Portuguese one of the words is usually omitted, or because there is a word with the complete meaning of the two Spanish words. This situation was named reduction and happens a few times. Examples are shown in Table 9. These were considered restricted combinations. The best examples from Table 9 are the first and the last. In Portuguese, and although there is the concept of meio ambiente, it is usually used only as ambiente. And in the case of cigarros pequeños, Portuguese as a word for that: cigarrilhas. These situations were validated manually in IATE (InterActive Terminology for Europe). Table 9: Examples of reductions: situations where two words are correctly translated by only one word. medio ambiente proteco de o ambiente , auditoría medioambientales de ecogesto e auditoria ( titular opositor parte de o titular . trabajo anual uma unidade de trabalho , cigarros pequeos cigarrilhas e cigarros , ● There is another situation with nouns (mostly geographic) that were mistakenly extracted, as shown in Table 10. These were extracted because of the way the nouns are translated. This table shows three columns. The first two are Spanish and Portuguese, and the third, a direct Portuguese translation of the Spanish term. Table 10: Examples of nouns whose translation was a problem for the algorithm. Sudeste Asiático sia de o Sudeste Sudoeste Asiático Repblica Federal originários de a Repblica Federativa República Federal continental espaol Espanha continental continental espanhol