Automatinis švietimo ir mokslo terminų nustatymas lingvistiniais metodais

Automatic Identification of Education and Science Terms Using Linguistic Methods ERIKA RIMKUTĖ Vytautas Magnus University KEYWORDS: corpus, education and science terms, linguistic methods, lemma form, grammatical form INTRODUCTION Corpus linguistics and computational linguistics in Lithuania are now in their second decade—the beginnings of these branches of science can be traced to the establishment of the Vytautas Magnus University Centre of Computational Linguistics (see http://tekstynas.vdu.lt)) in 1994. Several corpora and automatic language analysis and synthesis programs have been prepared1. Nevertheless, the principles of corpus linguistics are still applied only to a limited extent in terminology and terminography. The 2010 article Possibilities for the Automatic Identification of Lithuanian Terms (Rimkutė 2010) reviewed the extent to which corpora are used in compiling various terminological works in Lithuania. Lithuanian terms can be found in several international databases and terminology banks; a considerable number of electronic terminology collections or databases have also been compiled2, but all Lithuanian electronic terminology collections are essentially based on terminology dictionaries. Many Lithuanian dictionaries have been compiled on the basis of relatively small amounts of text and linguistic intuition; therefore, it can be said that prescriptive approaches to term extraction and description still prevail in Lithuania. This article aims to present the results of a project3 designed to identify and describe education and science terms, relating to the automatic identification of terms— 1 A comprehensive inventory of Lithuanian-language resources prepared before 2011 can be found at http://sruoga.vdu.lt/lituanistiniai-skaitmeniniai-istekliai/lituanistiniu-istekliu-sarasas. 2 Links are provided in the aforementioned article. 3 This is a project funded by the Research Council of Lithuania Automatic Identification of Education and Science Terms (ŠIMTAI 2) (contract no. LIT-2-44). The main objectives of this project are: 1) to develop and test a methodology for the automatic recognition of terms in a Lithuanian-language corpus as a new instrument for terminology management; 2) to create a specialized corpus of education and science texts; 3) to compile, based on the specialized corpus, an explanatory dictionary of education and science terms and an ontology for this field. 54 Erika Rimkutė Automatinis švietimo ir mokslo terminų nustatymas lingvistiniais metodais and their description. The idea for a project devoted to the automatic identification of terms arose from the observation that existing standardizing sources4 are rather incomplete, focus on the most general terms of education policy, and scarcely cover the field of science policy at all; they do not keep pace with new terms emerging from practical work, and terminologists do not manage to standardize them in time. Traditional standardization is nevertheless rather subjective, so usage and standards need to be grounded in corpus-linguistic methods. This article therefore discusses a still relatively new method of term extraction and analysis in Lithuania, which could substantially facilitate the work of terminologists and make it more objective. AUTOMATIC TERM IDENTIFICATION To identify and define the terms of a particular field automatically, one field was selected—science and education. A corpus of science and education (hereafter SE) comprising 4 million words was compiled, focusing on two main topics: 1) scientific research and 2) education. Within scientific research, the greatest attention was paid to research policy and higher education policy. The education component included post-secondary education, with the greatest emphasis on higher education and continuing education and training. Vocational training was also included as an additional field. The corpus was designed to cover texts of various genres and types, for example: laws and other political documents issued by the highest government institutions (resolutions and statements of the Seimas of the Republic of Lithuania, the Government of the Republic of Lithuania, and the President of the Republic of Lithuania; rulings of the Constitutional Court, etc.); strategic documents; implementing legislation; documents of institutions and projects implementing science and education policy; documents of science and higher education institutions; informational publications, studies of the state of the field, evaluations, feasibility studies; press releases; and discussion materials. Given a representative, well-balanced corpus of the field under study5, terms can be identified automatically. For this purpose, the methods most commonly used- 4 Examples include the Terminology Bank of the Republic of Lithuania, L. Jovaiša’s Encyclopedic Dictionary of Educational Science Terms (2007), and the TESE thesaurus [accessed 2012-05-20], available online at: http://eacea.ec.europa.eu/education/eurydice/documents/tese/pdf/teselt_005_alphabetic.pdf. 5 Corpus representativeness and size are among the most complex issues in corpus linguistics. It may be assumed that the SE corpus represents the selected field fairly well. Its size is more open to debate: for example, the English–Italian legal corpus Bononia Legal Corpus (BOLC) comprises approximately 18 million words (Rossini Favretti et al. 2001). Terminologija | 2012 | 19 55 three types of methods: statistical, linguistic, and hybrid models combining statistical and linguistic methods. Several experiments were conducted to identify the most suitable methods for term identification. The following statistical methods were tested: 1) the WordSmith Tools program function Keyword Clusters; 2) the KEA machine-learning method; 3) a collocation extraction method using the LICE tool (see Grigon worth et al. 2011 for more details). These studies led to the conclusion that linguistic methods are the most suitable for Lithuanian. The characteristics of these methods are briefly explained below. The main advantage of statistical term-identification systems is that they do not require large databases containing examples of terms supplied by humans. In addition, they are language-independent. The basic operating principle of statistical systems is that words frequently used together are related and can be treated as potential terms. Linguistic methods are usually based on morphological tags and word order. Morphologically annotated corpora are required to apply linguistic methods. When linguistic term-identification methods are used, the grammatical forms that terms may take are usually identified. The characteristics of each language must be taken into account when identifying grammatical forms. For example, for analytic languages it is important to identify the part of speech of terms. The part of speech of Lithuanian terms is also very important, but it is necessary to know which other grammatical categories must be taken into account for automatic term identification. It has been observed that number and case are also important categories in Lithuanian, whereas gender is not particularly significant for automatic term identification6. Tyrinėtojų (pvz., Bowker 2002) nustatyta, kad beveik neįmanoma pritaikyti kitoms kalboms sukurtų terminus nustatančių įrankių, ypač tų, kurie naudoja lingvistinius metodus. Pavyzdžiui, tipiška anglų kalbos terminų struktūra yra bdv. + dkt., dkt. + dkt., o prancūzų kalbai būdingi modeliai yra dkt. + bdv., dkt. + prl. + dkt. Taigi norint taikyti lingvistinius terminų atpažinimo metodus, reikia sukurti programą, kurioje būtų atsižvelgta į tam tikros kalbos gramatinę sistemą. The remainder of this article presents attempts to identify Lithuanian terms automatically using linguistic methods. It should be noted 6 Gender becomes relevant when case syncretism occurs, for example, mokslo darbuotojų (‘researchers’): the automatic morphological analysis program recognizes the form darbuotojų as the genitive plural of both a masculine and a feminine noun, and therefore gives both mokslo darbuotojas and mokslo darbuotoja as the citation forms. The masculine form should be given as the term, that is, mokslo darbuotojas. 56 Erika Rimkutė Automatinis švietimo ir mokslo terminų nustatymas lingvistiniais metodais attention, that a program created at the Vytautas Magnus University Centre of Computational Linguistics was used to identify possible terms (hereinafter PTs)7. This program used linguistic rules specifying which combinations of parts of speech or word forms should be analyzed. As mentioned, morphologically annotated corpora are needed when using linguistic methods, so the collected Education and Science corpus was first morphologically annotated using a morphological annotator8 developed at the Vytautas Magnus University Centre of Computational Linguistics. An example of automatically morphologically annotated text is presented here: <word=”PROJECTS” lemma=”project” type=”noun masc.gen.pl K”>9 <space> <word=”FUNDING” lemma=”funding” type=”noun masc.gen.sg K”> <space> <word=”CONDITIONS” lemma=”condition” type=”noun fem.gen.pl K”> <space> <word=”DESCRIPTION” lemma=”description” type=”noun masc.gen.sg V”> <p> <word=”I” lemma=”I” type=”Roman numeral”> <sep=”.”> <space> <word=”GENERAL” lemma=”general” type=”adjective positive indeclinable pronoun fem.sg.gen”> <space> <word=”PROVISIONS” lemma=”provision” type=”noun fem.gen.sg K”>10 <p> <number=”1”> <sep=””> <space> <number=”2007”> <sep=”–”> 7 Programs that use various methods to automatically identify terms most often recognize possible terms, or candidate terms (see Zeller 2005), although the words or combinations identified automatically often include completely meaningless word sequences or meaningful non-terminological word combinations, such as collocations. The actual terms are usually identified from these words and combinations by an expert in the relevant field. 8 For more details, see http://tekstynas.vdu.lt/page.xhtml?id=morphological-annotator-how-to-use. 9 Word indicates the specific word form used in the text, lemma denotes the headword form (lemma), and type provides detailed morphological information. 10 It should be noted that morphological annotation is performed automatically, so errors may occur; for example, in this example, an incorrect form is indicated because of case syncretism: the feminine singular genitive is indicated instead of the feminine plural nominative. For more details on inaccuracies arising from the specific characteristics of the morphological annotator, see the section Problems of Automatic Term Identification. Terminologija | 2012 | 19 57 <number=”2013”> <space> <word=”m” lemma=”m” type=”abbreviation” <sep=”.”> <space> The operating principle of programs that automatically identify PTs can be described briefly and simply as follows: morphologically tagged text is analyzed, and chains of related words are sought. The first criterion is to select uninterrupted word combinations according to the marked boundaries of the text and sentences. Some text boundaries are marked as punctuation marks; others may be marked as HTML tags, for example, the paragraph tag <p>, the tag for a space between words <space>, and so on. Next, certain linguistic rules must be prepared for implementation in the program. The following basic rules were prepared and applied when identifying education and science terms11: 1. Only those combinations that contain at least one noun are analyzed. If a PT consists of two or more words, the other parts of speech in that combination may be adjectives or participles. 2. Only those combinations whose final word must be a noun are analyzed12. 3. In two-word PTs, proper nouns are not analyzed, since they most often occur in the names of persons, companies, institutions, etc., which are not considered terms of the field (when identifying longer terms, proper nouns are also included in the PT term lists). 4. Remove numbers written in digits from the compounds. 5. Do not analyze words or compounds that contain at least one word not recognized by the morphological tagger. These are most often foreign-language insertions, misspelled words, or shortened word forms13. Applying these rules identified candidate terms (CTs), from which an expert14 selected education and science (ES) terms. These terms were analyzed in greater detail, and their structure is described in the section Structure of Education and Science Terms. 11 The rules are universal and are essentially suitable for automatically identifying terms in any subject area. 12 Other term structures are also possible, e.g., a noun with a prepositional infinitive, or a noun with a prepositional construction. There are few terms of this type, so they were not analyzed. Future research should also analyze term structures not included in this article. 13 Applying certain restrictions when selecting candidate terms may result in the loss of valuable information. This study did not calculate how many actual terms went unrecognized. When using automatic term recognition programs, one must accept that some information will be lost; however, automatically identified information will be easier to process, and the resulting compounds will be grammatically more regular. 14 The expert on this project is Lithuanian philologist Giedrius Viliūnas, who has extensive pedagogical, administrative, and managerial experience in education and science. 58 Erika Rimkutė Automatinis švietimo ir mokslo terminų nustatymas lingvistiniais metodais TERM SELECTION Applying the method, program, and rules described in the previous section identified approximately 11,000 candidate terms (CTs), consisting of between one and fifteen words15. Faced with such a large volume of data, several selection criteria were applied. The first was term length: considering general trends in term structure, it was decided to analyze terms of no more than seven words. The second criterion was term frequency. The analysis focused on one-word CTs occurring at least 20 times in the ES corpus; two-word CTs occurring at least 10 times; three-word CTs occurring at least 8 times; four-word CTs occurring at least 6 times; five-word CTs occurring at least 4 times; six-word CTs occurring at least 3 times; and seven-word CTs occurring at least twice. When automatically calculating the frequency of candidate terms, the words’ headword (dictionary) forms, i.e., lemmas, were used. The terms identified using these selection criteria were analyzed in greater detail (their structure is presented in the next section). After an expert reviewed the candidate terms identified using the criteria mentioned above, the actual education and science terms were selected. Their distribution by length is shown in Table 1. Table 1. Distribution of education and science terms by length Number of words comprising the terms Number of terms Percentage of terms 1 155 19,87 2 474 60,77 3 125 16,03 4 22 2,82 5 4 0,51 Total 780 100,00 15 Here are some of the longest terms found: repayment of a state-supported loan to the borrower by a decision of a commission formed by the fund director in the cases specified in the paragraph (15 words; numbers, i.e. the number of the paragraph, are omitted here); at the borrower's address of residence specified in the state-supported loan agreement concluded by the credit institution (13 words); the procedure for payment of university students' tuition fees for the academic year is established by order of the rector (11 words). It is clear that these text fragments cannot be considered potential terms. Terminologija | 2012 | 19 59 The table data show that about 97% of all education and science (ES) terms consist of terms made up of 1–3 words. No six- or seven-word GT education and science terms were found, so they were not analyzed further. STRUCTURE OF EDUCATION AND SCIENCE TERMS This section describes in detail the structure of ES terms identified automatically and selected by an expert, and presents their grammatical models. The terms are described from the shortest, i.e. one-word terms, to the longest, i.e. five-word terms. 1. One-word terms. A total of 7635 one-word GTs were identified, of which 1311 occurred more than 20 times in the corpus. Only these words were analyzed in greater detail below. Of the 1311 words, 155 can be considered ES terms, i.e. about 20%. The ES terms occurring most frequently in the corpus are: student (420816), university (3959), studies (1562), science (1508), field (1403). 2. Two-word terms. In the corpus of 4 million words, 145,468 two-word GTs were found. As noted earlier, only those two-word GTs with a frequency of at least 10 occurrences were analyzed further—4889 such terms were found. After reviewing the GTs, only 474 two-word terms remained in the final list of ES terms, and they account for the majority of all terms analyzed—more than two-thirds (see Table 1). Two-word terms can be divided into three types according to their grammatical forms. 1) noun.sg.nom. + noun.sg.nom. type: there are 250 ES terms, i.e. 52.7% of all two-word terms. The most frequent ES terms of this structure occurring in the corpus are study programme (1432), cycle of studies (306), field of study (303), researcher (289), quality of studies (253). 2) adjective + noun.17 These combinations make up 200 ES terms (42.2% of two-word terms). The most frequent ES terms are higher education institution (2160), scientific research (918), higher education (730), distance learning (453), vocational training (403). Adjective + noun type combinations are considered GTs when the adjective agrees with the noun in gender, number, and case. Otherwise, the com- 16 The frequency of the term in the ES corpus of 4 million words is given in parentheses. 17 Cases are not specified in the models for two-word terms, because adjectives and participles agree with nouns, while nouns are usually in the singular nominative case. 60 Erika Rimkutė Automatinis švietimo ir mokslo terminų nustatymas lingvistiniais metodais grammatically incorrect¹⁸ combinations, e.g., closer universities, most important law, most important learning, best year, or the combination is correct, although often incomplete, and these are not education and science (E&S) terms, e.g., largest in the world, most dynamic in the world, favorable to students. 3) adv. + noun combinations classified as E&S terms number only 24, i.e., 5%. The most frequent E&S terms are: final thesis (162), elective subject (109), adult¹⁹ education (65), feedback (63), applied research (45). 3. Three-word terms. A total of 119,881 three-word GTs were identified. A total of 2,438 combinations occurring at least 8 times in the corpus were analyzed in detail. A total of 125 E&S terms were found, i.e., 16.03% of all terms. The most frequent E&S terms are: higher education institution (251), vocational training institution (205), state-supported loan (149), general education school (147), vocational training institution (77). Three-word E&S terms follow seven grammatical patterns. Among three-word E&S terms, combinations of the type adj. gen. + noun gen. + noun²⁰ are the most frequent (they account for 48% of all three-word terms), e.g.: higher education institution (251), vocational training institution (205), general education school (147). The second most frequent among three-word E&S terms are combinations of the type noun gen. + noun gen. + noun (24.8%), e.g.: study field regulations (66), description of study fields (60), state research institute (50). The third most frequent structural pattern for three-word E&S terms is adj. nom. + adj. nom. + noun (12%), e.g.: national complex program (69), fundamental scientific research (61), initial vocational training (52). The following structural patterns for three-word E&S terms can be illustrated with only a few examples each: noun gen. + participle nom. + noun (4%), e.g.: state-supported loan (149), state-recognized qualification (33), state-funded student (21); noun gen. + adj. nom. + noun (4%), e.g.: foreign higher education institution (41), college academic council (32), doctoral student's supervisor (20); participle nom./gen. + noun gen. + noun (3.2%), e.g.: final qualification assessment (36), tuition-free study place (13), applied research (13); adj. nom. + ¹⁸ In this work, grammatically correct combinations are considered to be those in which dependent words agree with the head word and the syntactic relations between the components of the phrase are clear. ¹⁹ Participles that have become nouns are also included among the participles. ²⁰ The case of the final noun is not specified because it is most often nominative. Terminologija | 2012 | 19 61 participle, active voice, present tense (or adjective, if the participle is substantivized) + noun (2.4%), e.g.: competitive subject of instruction (15), non-formal adult education (13), scientific research work (12); participle, active voice, present tense + adverb, present tense + noun (1.6%), e.g.: applied scientific activity (20), commissioned scientific research (8). 4. Four-word terms. A total of 77,961 four-word GTs were found, with 760 combinations occurring more than 7 times. Of these, the expert identified 22 education and science (ES) terms (2.82% of all terms). The most frequent combinations in this term group are institution of science and studies (568), science and technology park (62), general national comprehensive programme (40), vocational training in the labour market (40). The four-word terms used in the ES corpus represent nine different grammatical patterns. Since there are few occurrences of four-word terms, each pattern is illustrated by just a few examples or only one. Listed here, grouped by frequency, are all the structural patterns of four-word ES terms, with one example illustrating each pattern: adjective, present tense + noun, nominative case + noun, nominative case + noun (22.7% of all four-word terms): international scientific database (13); adjective, present tense + adjective, nominative case + noun, nominative case + noun (18.2%): general university education subject (8); adjective, nominative case + noun, nominative case + participle, active voice, present tense + noun (13.6%): student with special needs (8); noun, nominative case + conjunction + noun, nominative case + noun (13.6%): institution of science and studies (568); adjective, nominative case + noun, nominative case + adjective, present tense + noun (9.1%): high-level scientific research (18); participle, active voice, present tense + adjective, present tense + noun, nominative case + noun (9.1%): recognized international database (11); adjective, present tense + adjective, present tense + adjective, present tense + noun (4.5%): general national comprehensive programme (40); noun, nominative case + adjective, nominative case + noun, nominative case + noun (4.5%): state research institution (31); noun, nominative case + noun, nominative case + noun, nominative case + noun (4.5%): field of technology studies (28); 5. Five-word terms. A total of 49,768 five-word GTs were identified; 601 combinations occurring more than 5 times were found, of which 4 were ES terms, i.e. 0.51% of all terms. It is difficult to generalize about their structure from such a small number of terms, so they were not classified structurally. All five-word terms occurring in the ES corpus are: research and experimental development (124), research and technological development (67), research and technology development (13), high-level research centre (7). 62 Erika Rimkutė Automatinis švietimo ir mokslo terminų nustatymas lingvistiniais metodais 6. Six-word and seven-word terms. It has already been mentioned that no educational science (ES) terms were found among the six-word and seven-word GT terms, although a total of 30,298 six-word GTs were identified (of these, 750 combinations were found to occur more than 3 times) and 17,815 seven-word GTs (of these, 2,608 combinations were found to occur more than 2 times). COMPARISON OF EDUCATION AND SCIENCE TERMS WITH THE TERMS IN THE TERM BANK OF THE REPUBLIC OF LITHUANIA The terms used in the education and science corpus can be compared with the approved education and science terms included in the Term Bank of the Republic of Lithuania21, of which there are 74. The structure of these terms is highly varied, ranging from single words to combinations consisting of 18 words. Only one very long term, consisting of 15 words, and one consisting of 18 words were found. The longest term listed is the non-formal continuing vocational training programme for farmers, forest owners, members of local action groups, other rural residents and their consultants. It is debatable whether the name of a programme should be considered a term. A comparison of the structure of education and science terms in the education and science corpus and the Term Bank shows that the most frequent terms in both sources are those consisting of one to three words. However, one-to-three-word terms make up 86.5% of the Term Bank, whereas terms of the same length used in the education and science corpus account for 97% (see Table 1). The Term Bank contains somewhat more four-word terms than the corpus presented here; no five-word terms are used; there are only one or two terms consisting of 6, 7, 8, or 9 words in the Term Bank; there are no terms consisting of 10 words, but the previously mentioned terms consisting of 15 and 18 words are included (see Table 2). It is doubtful whether several terms included in the Term Bank should be classified as belonging to the field of education and science, e.g., Lithuanian, World Lithuanian Community, re-emigration, ethnicity. The resulting list of terms shows that the field of education and science is not fully covered: for example, tractor driver is listed as a vocational education term, but combine harvester operator, cook, carpenter and so on are not; terms reflecting science policy are particularly lacking. Thus, the terms identified in the study discussed in this article should at least partly fill the gap in the terminology of education and science. 21 The list of approved terms in the field of education and science was obtained from the State Commission of the Lithuanian Language in 2010. The comparison of corpus and Term Bank data is based specifically on this list. Terminologija | 2012 | 19 63 Table 2. Distribution of education and science terms in the Term Bank of the Republic of Lithuania by length Number of words forming the term Number of terms Number of terms,% 1 16 21,6 2 31 41,9 3 17 23,0 4 3 4,1 5 0 0 6 1 1,4 7 1 1,4 8 1 1,4 9 2 2,7 10 0 0 11 0 0 12 0 0 13 0 0 14 0 0 15 1 1,4 18 1 1,4 Total 74 100 PROBLEMS IN AUTOMATIC TERM IDENTIFICATION When identifying terms using various methods, two main problems arise: 1) what forms constitute grammatically correct terms, which is particularly relevant when determining the headword form; 2) how to distinguish general phrases and general terms from terms belonging to the field under analysis. Scholars in other countries face similar problems (cf. Cabré Castellví et al 2001). The second problem can usually be solved only non-automatically, i.e. an expert in the field must review the GT and determine which terms belong to the field under analysis. Certain filters can also be applied, i.e. a list of words to be excluded (in English stop list) can be compiled and used to exclude words that cannot function as terms in the field under analysis. Such filtering was also applied in this study. A few examples: words or phrases whose final word denotes time were not considered terms in education and science (ES), 64 Erika Rimkutė | Automatinis švietimo ir mokslo terminų nustatymas lingvistiniais metodais i.e. a noun, e.g.: year, month, day, week, morning, evening, January, February, Monday, Tuesday, etc. Words denoting units of measure are not considered terms, e.g.: millimeter, meter, centimeter, kilometer, mile, gram, kilogram, tonne. Phrases containing word combinations characteristic of general administrative terms can be excluded, e.g.: article of a law, in a section of an article of a law, as prescribed by law, approved by a clause of an order, in accordance with the procedure laid down by legal acts, the government of the republic, the state of the republic, a clause of regulations, a clause of resolutions, an order of the minister, in accordance with the procedure laid down by the minister, approved by the minister, formed by the minister. In an effort to distinguish ES terms from general terms as accurately as possible by automatic means, two reference corpora were compiled, each the same size as the ES corpus, i.e. 4 million words. The first reference corpus consists of texts from various genres and styles; its composition is similar to that of the Corpus of the Contemporary Lithuanian Language (including journalistic, literary, administrative, and scholarly texts). The second reference corpus consists of administrative texts, but not from the ES field. The hypothesis was that the more general a term is, the more frequently it is used in the general and/or administrative corpus, whereas the more specific a term is, the less frequently it is used in the general and/or administrative corpus. This hypothesis was partly confirmed: for example, terms characteristic of the ES field, such as mobility, competence, attrition, learning outcome, were not found in the reference corpora, while terms such as student, university, textbook are frequent both in the ES corpus and in the reference corpora. This study should be continued to provide reliable data on the extent to which comparing several corpora can help distinguish general terms, at least partly automatically, from terms belonging to the field under analysis. The first problem will be discussed in greater detail below, namely how to automatically determine the correct lemma of terms (especially multiword terms) and how to bring words and phrases used in different inflectional forms in the corpus together under a single lemma. The study showed that it is not always easy to determine a term’s headword form. When a lemma is identified automatically, for example, the phrase aukštoji mokykla produces the result aukštas mokykla. This form is given because the morphological annotator specifies the masculine singular nominative form of the positive degree as the lemma for adjectives (for more detail on the annotator’s specific features, see Zinkevičius 2000). For the lemmas of multiword terms Terminologija | 2012 | 19 65 would be greatly helped by a syntactic analysis program. Unfortunately, no syntactic analysis program that works in practice is currently available22. Thus, determining lemmas requires certain additional rules—in this case, it must be specified that if a lemma consists of a noun, then the adjective or participle must agree with that noun in gender; if the adjective or participle is pronominal, this category must also be retained in the lemma. Many questions arise concerning the number category. For example, the first component of the term studentų atstovybė (student representative body) must be in the plural; all components of the terms akademiniai įgūdžiai, auditorinės darbo valandos (academic skills, classroom hours) must be in the plural; whereas both components of the terms fakulteto taryba, universiteto autonomija, bendrasis priėmimas, mokslinis leidinys (faculty council, university autonomy, general admission, scholarly publication) are singular. Of course, it is easy for a person to determine where the plural form is needed and where the singular form is needed, but doing so automatically is very difficult. Indeed, the corpus shows that some words are used almost exclusively in the plural (e.g., asignavimai, duomenys, studijos, rūmai, žinios, pinigai, pareigos [appropriations, data, studies, halls, knowledge, money, positions]). Determining the correct number form is also difficult because some words are homonyms, e.g., studija (a scholarly work) and studijos (studying). For this reason, the lemmas of the following terms were determined incorrectly automatically: bakalauro studija, rezidentūros studija, magistro studija, dieninės studija, nuotolinės studija. As noted, some nouns are generally used only in the plural, although their dictionary lemma is given in the singular. This mismatch between dictionary information and actual usage resulted in the incorrect determination of the lemmas of the terms mokinio pasiekimas, švietimo resursas, praktinis gebėjimas, fizinis mokslas, akademinis užsiėmimas (they should be: mokinio pasiekimai, švietimo resursai, praktiniai gebėjimai, fiziniai mokslai, akademiniai užsiėmimai). Another problem is substantiva mobilia nouns, e.g., abiturientas – abiturientė, abonentas – abonentė (male graduate – female graduate, male subscriber – female subscriber). Since forms of both masculine and feminine gender occurred in the corpus, lemmas of both genders were specified automatically. To retain only the masculine lemma as an education and science term, a rule had to be applied whereby masculine and feminine nouns of the substantiva mobilia type were mapped to the masculine lemma. 22 Automatinės sintaksinės analizės srityje paminėtini D. Šveikauskienės 2009, E. Rimkutės 2006 darbai. Automatinės sintaksinės analizės programa kuriama VDU Kompiuterinės lingvistikos centre ir iš dalies jau taikoma analizuojant ŠM terminus. 66 Erika Rimkutė Automatinis švietimo ir mokslo terminų nustatymas lingvistiniais metodais When lemmas are determined automatically, the word order is sometimes incorrect; for example, the following headword forms were identified: tipų studija (should be studijų tipas), lygmens kvalifikacija (should be kvalifikacijų lygmuo). Because syntactic analysis was not applied when determining lemmas, modifiers do not always agree with the words they modify, e.g., tarptautinis (= tarptautinio) lygio mokslas (international-level science), or other syntactic violations occur, e.g., paskolų gyvenimo išlaidos (= paskola gyvenimo išlaidoms) (loan living expenses [loan for living expenses]). Difficulties also arise in determining term boundaries. For example, is dėstytojų ir mokslininkų kvalifikacijos bei kompetencijos atitiktis (the conformity of lecturers’ and researchers’ qualifications and competencies) one term, or should several terms be distinguished here: 1) conformity of lecturers’ qualifications and competencies, 2) conformity of researchers’ qualifications and competencies, 3) conformity of lecturers’ and researchers’ qualifications, 4) conformity of lecturers’ and researchers’ competencies, 5) lecturers’ qualifications, 6) researchers’ qualifications, 7) lecturers’ competencies, 8) teachers’ competencies, 9) conformity of qualifications, 10) competency conformity. Difficulties in identifying lemmas and, more generally, recognizing terms also arise from the specific characteristics of the morphological tagger, when an incorrect part of speech or grammatical form is selected from among several possible ones. For example, if an adjective agrees with a nearby noun in gender, case, and number, there is a high probability that the phrase will be grammatically correct. Phrases consisting of two nouns, where the first constituent is not in the genitive case, are not always grammatically correct. They need to be reviewed more carefully, and thus take more time. Consequently, whether a phrase is included among the GTs and whether it is grammatically correct depends on which part of speech has been specified. Several examples illustrating this type of error: metinėms (may be an adjective or a noun; the morphological tagger identified it as a noun) ataskaitoms; keliais (may be a pronoun or a noun; selected as a noun) procentais; bendrais (may be an adjective or a noun; selected as a noun) teiginiais; uždarus (may be an adjective or a noun; selected as a noun) posėdžius. Because the part of speech is specified incorrectly and the words are not analyzed syntactically, the phrase žemiausias balas found in the morphologically annotated corpus appears at first glance to be grammatically incorrect, because Terminologija | 2012 | 19 67 the first word is assigned the following tags: adjective, superlative, non-pronominal, masculine, singular, nominative; the second: noun, feminine, plural, genitive, and the lemma is given as bala. Thus, it can be seen that in this case the noun lemma has been identified incorrectly—it should be balas. Consequently, the phrase mentioned above is grammatically correct. This section has reviewed the difficulties most commonly encountered in automatically identifying education and science terms. These problems and ways of addressing them will be discussed in other works. CONCLUSIONS The program and methodology for automatic term identification still need improvement. Nevertheless, the initial results are quite good. As the examples show, automatic analysis brings to light not only specific, technical problems, but also issues faced by terminologists, such as the choice of number category. Once the morphological tagger has been improved and an automatic syntactic analysis program applied, terms could be identified automatically more easily and reliably. Despite the difficulties encountered, the program that automatically identifies terms considerably speeds up the GT identification process. Another important advantage of the automatic term-identification process is the objectivity of data analysis. It was found that 97% of the education and science terms automatically selected and reviewed are 1–3 words long. Two-word terms predominate, accounting for almost 61% of all terms. The most common structural patterns of education and science terms are noun genitive + noun, adjective + noun, adjective genitive + noun genitive + noun, and noun genitive + noun genitive + noun (the final noun can take various cases). LITERATŪRA Bowker L. 2002: Computer-aided translation technology – a practical introduction. University of Ottawa Press. Cabré Castellví M. T., Estopà Bagot R., Palatresi J. V. 2001: Automatic term detection. A review of current systems. – Recent Advances in Computational Terminology, 53–88. Grigonytė G., Rimkutė E., Utka A., Boizou L. 2011: Experiments on Lithuanian Term Extraction. – Proceedings of the 18th Nordic Conference of Computational Linguistics NODALIDA 2011. NEALT Proceedings Series, Vol. 11, 82–89. Rimkutė E. 2006: Morfologinio daugiaprasmiškumo ribojimas kompiuteriniame tekste. Daktaro disertacija. Rimkutė E. 2010: Lietuvių kalbos terminų automatinio nustatymo galimybės. – Kalbų studijos 16, 71–78. Rossini Favretti R., Tamburini F., Martelli E. 2001: Words from Bononia Legal Corpus. – International Journal of Corpus Linguistics, special issue, 3–33. Šveikauskienė D. 2009: Kalbos vientisinių sakinių automatinė sintaksinė analizė. Daktaro disertacija. Zeller I. 2005: Automatinis terminų atpažinimas ir apdorojimas. Daktaro disertacija. Vytauto Didžiojo universitetas. Zinkevičius V. 2000: Lemuoklis – morfologinei analizei. – Darbai ir dienos 24, 245–273. 68 Erika Rimkutė Automatinis švietimo ir mokslo terminų nustatymas lingvistiniais metodais AUTOMATIC IDENTIFICATION OF SCIENCE AND EDUCATION TERMS USING LINGUISTIC METHODS The paper deals with possibilities and problems of automatic Lithuanian term extraction. Specifically, linguistic methods are discussed in the domain of Education and Science. Other researchers have shown that it is almost impossible to have language-independent term extraction tools; this is especially true for tools which are based on linguistic rules. Therefore a linguistic term extraction tool should incorporate methods that would deal with a language’s grammatical system. This paper presents a tool developed at the Centre of Computational Linguistics of Vytautas Magnus University that employs linguistic rules for extracting domain-specific terminology. In order to extract domain-specific terms automatically, some preparatory work should be completed: compilation of domain-specific corpus (a corpus of four million words has been compiled for this research), morphological annotation of the corpus, formulation of appropriate linguistic rules, and creation of methodology for filtering out irrelevant word combinations. The paper presents the linguistic rules that have been used for the extraction of Education and Science terms and the results of the extraction procedure. The identified terms are contrasted with approved terms in the Term Bank of the Republic of Lithuania. Some specific problems of automatic term extraction are discussed in the paper, e.g. number and case agreement of terms in a multi word term. Gauta 2012-05-17 Erika Rimkutė Vytauto Didžiojo universitetas K. Donelaičio g. 52, LT-44248 Kaunas E. paštas [email protected] Terminologija | 2012 | 19 6 9