<?xml version="1.0" standalone="yes"?>
<Paper uid="W93-0304">
  <Title>Towards a cross-linguistic tagset</Title>
  <Section position="1" start_page="0" end_page="36" type="abstr">
    <SectionTitle>
Abstract
</SectionTitle>
    <Paragraph position="0"> With the spread of large quantities of corpus data, the need has arisen to develop some standard not only for the format of'interchange of text (an issue which has already been taken up by the Text Encoding Inititiave), but also for any information added in some subsequent stage of (linguistic) enrichment. The research community has much to gain by such standardization since it will enable researchers to e~ectively access and therefore make optimal use of the results of previous work on a corpus.</Paragraph>
    <Paragraph position="1"> This paper provides some direction of thought as to the development of a standardized tagset.</Paragraph>
    <Paragraph position="2"> We focus on a minimal tagset, i.e. a tagset conraining information about wordclasses. We investigate what criteria should be met by such a tagset. On the basis of an investigation and comparison of ten different tagse~s that have been used over the years for the (wordclass) tagging of corpora, we arrive at a proposal for a cross-linguistic minima\] tagset for Germanic languages I .</Paragraph>
    <Section position="1" start_page="0" end_page="30" type="sub_section">
      <SectionTitle>
Part I
Introduction
</SectionTitle>
      <Paragraph position="0"> The last few years there has been an increasing interest in the use of corpus data, especially by those working in the field of natural language processing (NLP). This development can in part be ascribed to the fact that the scale on which such data are becoming available has increased: as publishing houses, industry, etc.</Paragraph>
      <Paragraph position="1"> switched to an electronic format for the in- null 1. In this paper the focus is on British English. Amerlean English, Dutc21 and German.</Paragraph>
      <Paragraph position="2">  terchange of texts this meant a dramatic increase in the amount of text that was readily available to anyone interested, while developments in hardware and in software have made it possible to manipulate large quantities of data (more) effectively.</Paragraph>
      <Paragraph position="3"> Earlier corpus-based approaches proved to be laborious undertakings. Before the 'real' work could begin one would have to go through the painstaking process of designing a corpus, gaining permission from publishers to use (part(s) of) texts and somehow making the texts computer-readable. The size of the corpora was very much determined by such factors as cost (in terms of time and money to be invested) and availability of data. Corpora generally were compiled for particular research purposes, such as the investigation of a particular variety of the language. Typical examples here are the Brown Corpus and the Lancaster-Oslo/Bergen (LOB) Corpus which were both compiled with the intention of representing a cross-section of the language (American English and British English respectively). Compilers of corpora would each adopt his/her own conventions for representing the textual data. In a similar fashion the further processing of the data, the enrichment of (basically) raw text with some kind of linguistic information (wordclass information, syntactic, semantic and/or even pragmatic information) would largely depend on the time and money available, and also on the particular interests of the researchers involved.</Paragraph>
      <Paragraph position="4"> Recently, with the increase in the amount of data that are becoming available, the attention of corpus compilers has been drawn to the need for a common interchange format for texts in order to make these data more readily  accessible for third parties. The Text Encoding Inititiative (TEI) has undertaken to develop a standard for the marking up of texts that is based on the Standard Generalized Markup Language (SGML). So far we have not yet reached the point where any serious attempts are being made to standardize the linguistic information that is being added in successive stages of linguistic enrichment. Instead, researchers from different backgrounds and with different beliefs are working their own turf exploring a variety of methods for the enrichment of corpora, ranging from purely stochastic to strictly rule-based approaches, which seem to be in competition with each other. In view of the current state of the art in corpus analysis, the amount of work that has already been done in the area of tagging corpora for wordclass information and the experience that has been gained in the process, it would appear that by now the time has come to start thinking about developing some sort of standard for the encoding of wordclass information.</Paragraph>
      <Paragraph position="5"> In the remainder of this paper we focus on the design of a minimal tagset, i.e. a tagset containing information about wordclasses, that will provide a common basis for the wordclass tagging of texts written in Germanic languages.</Paragraph>
      <Paragraph position="6"> We first compare a number of tagsets that have been or are being used for the tagging of wordclass information in prominent corpora.</Paragraph>
      <Paragraph position="7"> On the basis of this comparison we arrive at a number of criteria that a standardised (minireal) tagset should meet. Finally put forward a proposal for a basic tagset that may be applied cross-linguistically.</Paragraph>
    </Section>
    <Section position="2" start_page="30" end_page="33" type="sub_section">
      <SectionTitle>
Part II
</SectionTitle>
      <Paragraph position="0"> Tagsets: a comparison In order to give the reader some idea of the kind of (wordclass) information covered in various tagsets, a comparison is made of the tagsets employed in ten corpora 2 the Brown Corpus (Ku~era and Francis, 1967), the Lancaster-Oslo/Bergen Corpus (3ohansson et al., 1978), the SUSANNE Corpus (Sampson, 2. Information regarding a detailed characterization of the corpora including size, design, research context and method of processing can be found in the corresponding literature forthcoming), the Penn Treebank (Santorini, 1991), the IBM-Lancaster Corpus (Black et al., 1993), the British National Corpus (Leech, 1993), the ICE Corpus (Greenbaum, 1991), the Tosca Corpus (Oostdijk, 1991), the Dutch Eindhoven Corpus (uit den Boogaart, 1975), and the German IDS-Mannheim Corpus (Neumann, 1987).</Paragraph>
      <Paragraph position="1"> As a first step in making a comparison of the ten tagsets employed in the wordclsss tagging of the above corpora we start by distinguishing the wordclass categories that these tagsets have in common. There are nine categories that are (in one fashion or another) included in each of the tagsets: noun, pronoun, article/determiner, adjective, adverb, preposition, conjunction, verbal form, and interjection. Apart from these nine categories we distinguish one other category, which by definition is open-ended since it is intended to cover all tags that do not fit into any of the other categories. We shall refer to this category as Vopen category&amp;quot;. Next, we make an inventory of the different tags employed by each of the tagsets. As a result we obtain lists 3 of tags for each of the wordclass categories we distinguished. For example, Table 1 lists the various tags as they occur in the ten tagsets under consideration for the tagging of conjunctions (incl. connectives).</Paragraph>
      <Paragraph position="2"> There appears to be a great deal of overlap: while actual tags may differ (cf. CO, CI, 70 for conjunctions in general, in the ICE, Brown and Eindhoven tagsets respectively), there seems to be some consensus as to the kind of informarion one wants to encode or tag. Before going into this, however, we take a closer look at both the format and the nature of the tags.</Paragraph>
      <Paragraph position="3"> Starting with the oldest tagset included in our comparison, the Brown tagset, we see that this is in a sense a very 'fiat' coding scheme. It was developed with the intention of encoding genera\] wordclass information. The tags consist of character sequences which encode the wordclass category of a word and occasionally extended information such as form, aspect or case. Thus for example we find JJ for adjective and CC for coordinating conjunction, but also VB for base form of verb, VBD for past tense of verb, VBG for present participle, etc.</Paragraph>
      <Paragraph position="4"> The LOB tagset, but also the tagsets 3. These lists are rather sizeable and have therefore not been included. They are available, however, from the author (via e-mail).</Paragraph>
      <Paragraph position="5">  subord, conj. matrix sent. wordorder introductory part of conjunctive groups  of Penn Treebank, IBM-Lancaster and the British National Corpus very much follow the Brown tagset. Of these tagsets, the LOB tagset is of course closest to the Brown tagset, since the corpus was compiled with the intention of comparing American and British English from the same year (1961).</Paragraph>
      <Paragraph position="6"> The Penn Treebank tagset appears to be a reduced version of the Brown/LOB tagset.</Paragraph>
      <Paragraph position="7"> The reduction becomes manifest in that some tags are less detailed. For example, where Brown/LOB distinguish between possessive pronoun, persona\] pronoun, and reflexive pronoun and with each of these has a further subclassification (e.g. PPIA for personal pronoun, first person singular nominative; PP1AS for persona\] pronoun, first person plural nominative; PPIO for persona\] pronoun, first per-son singular accusative; etc.), Penn Treebank only has one tag, PP, to cover all persona\], possessive, and reflexive pronouns. All in all the derivation of the Penn Treebank from the Brown/LOB tagset is fairly straightforward. A minor deviation occurs in the tagging of genitive markers in the case of nouns. Here according to the Penn Treebank tagging scheme genitive markers are assigned their own tag, which is not the case with the Brown/LOB tagging scheme.</Paragraph>
      <Paragraph position="8"> The various tagging schemes that have emerged from the cooperation between IBM and the Lancaster-based Unit for Computer  Research on the English Language 4 all can be placed within what might be referred to as the Brown/LOB tradition in tagging. For our discussion we single out one particular tagging scheme, called CLAWS2a. This tagging scheme was developed to encode the information needed at wordclass level in order to be able to parse computer manuals. Although this tagging scheme clearly shows some resemblance to the Brown/LOB tagging scheme, it stands out in that with certain categories (such as verb and noun) there is a lot of extra detail that we do not find elsewhere. For example, tags such as noun o\[sty\]e and noun oPS organization are unique for this tagset. When we look at the format of the tags we find that there may be up to five characters per tag, while the leftmost characters relate the more general wordclass information and the characters that occur towards the right specify further detail. Compound lexical items can be encoded by means of tags that extend over more than one word. For example, while II would be the tag for simple preposition, the complex preposition in spite of is tagged I131 I132 I133. The tagset that has been developed within the framework of the British National Corpus (BNC) contains only some 60 different tags s The reason for this must be sought in the fact that the BNC intends to incorporate a very large amount of text (approx. 100 million words). The grammatical (i.e. wordclass) tagging of the corpus is to be carried out automaticaily. For reasons of efflclency and also to increase the success rate of the tagging it was decided to have a rather small tagset. A comparison of the BNC tagging scheme to the one used in the IBM-Lancaster Corpus shows that the two tagging schemes are closely relate&amp; Both can be characterized as 'fiat' tagging schemes. The major difference between the two, apart from their sizes, appears to be that the BNC uses more mnemonic abbreviations for otherwise similar tags 6 ; for example general adjectives get the label AJO instead of  JJ.</Paragraph>
      <Paragraph position="9"> 4. Black et al. (1993).</Paragraph>
      <Paragraph position="10"> ~. The British National Corpus is a joint undertaking by Oxford University Press, Longman, and V~'.&amp; R. Chambers, the universities of Lancaster and Oxford, and the British Library.</Paragraph>
      <Paragraph position="11"> 6. This was probably done in the hope to improve the  readlbilhy of the various tags. For example, the tag JJ for adjective is replaced by the tag AJO for general ~ljective.</Paragraph>
      <Paragraph position="12"> So far we have been looking at the tagging schemes employed by five English language corpora. Turning away from those for a moment and shifting our attention to the German and the Dutch corpora, we see that the tagging schemes employed in these corpora do not differ all that much from what we have already seen with the English corpora. Again 'fiat' tagging schemes are found, while the wordclass categories that are distinguished largely coincide with the English wordclasses. As with the tagging schemes for the English language corpora, the tagging schemes for the German and Dutch corpora each have their own degree of detail.</Paragraph>
      <Paragraph position="13"> The 12-million-word Mannheim Corpus was tagged automatically for both wordclass information and syntactic information by means of the SATAN parser ~. Neither level of analysis includes much detail. The wordclass tagging is rather rudimentary and includes only the most basic wordclass information needed for the syntactic analysis. Only occasionally information is added about prepositions that occur as collocates of other words and about the case that prepositions require for their complements. null The tagging of the relatively small Eindhoven Corpus is rather detailed. The tags consist of three digit codes, where the first digit indicates the wordclass, the second digit supplies information on the subclass, and the third digit carries additional information of various kinds, such as verb aspect, person, number, etc.</Paragraph>
      <Paragraph position="14"> Returning now to the tagging schemes eraployed for the tagging of English language corpora, we find that in the case of the three remaining tagging schemes under consideration - the SUSANNE tagging scheme, the ICE tagset, and the Tosca tagset - they have opted for more hierarchically structured tagging schemes. Each of these tagging schemes encodes highly detailed wordclass information in a systematic fashion by introducing some sort of 'additional feature(s)' slots in their tags. The SUSANNE tagging scheme distinguishes as many as 352 distinct tags. Closer examination of these tags however learns us 7. The Mannheim Corpus has been compiled by the Irmtitut f/h&amp;quot; Deutsche Sprache (ID5), the German national institute for research on the German language. The SATAN parser was developed at the University of 5aarbr,~cken for the purpose of machine translation.  that at the basis of these 352 items there are some 70 major wordclasses, while the addition of more featurelike information such as number, person, case, etc. leads to a total of 352 tags. Something similar appears to be the case when we consider the Tosca tagset and its less detailed derivative, the ICE tagset. Again tags for major wordclass categories have been extended so as to include additional feature in= formation.</Paragraph>
      <Paragraph position="15"> In summary, having compared the different tagging schemes that have been or are being employed in the tagging of wordclass information in ten different corpora, we find that * there appears to be some consensus as to what to tag as far as wordclass information is concerned; * even the format of the tags in different tagging schemes does not differ all that much; * the kind of information does not vary too much from one language to another, in other words it appears feasible to have a tagset that could be applied crosslinguistically. null As to the last point, we must observe that the idea of having a tagset that can be applied cross-linguistically is not at all new. For example, the tagset employed in the multi-lingua/ ESPRIT-860 project \[Kesselheim, 1986\] was intended to comprise all the tags that would be required for the tagging of wordclass informarion in each of the EC languages. Since this meant that not only Germanic languages were included (Dutch, English, German), but also Romance languages (French, Italian, Spanish) and Greek, the tagset could be expected to be truly cross-linguistic. Unfortunately, however, examination of the tagset shows that it is more of a collection of various tags that any of the languages might at some stage require, rather than a thoroughly designed minimal interlingual (or cross-linguistic) tagset. It is therefore all the more surprising to find that certain wordclasses in German, Dutch and English are not accounted for s.</Paragraph>
      <Paragraph position="16"> 8. For example, there appears to be no wordclnss tag for interjections for German and Dutch, nor is there a tag for particl~ in English or in Dutch.</Paragraph>
    </Section>
    <Section position="3" start_page="33" end_page="34" type="sub_section">
      <SectionTitle>
Part III
</SectionTitle>
      <Paragraph position="0"> Criteria for a minimal tagset Having compared the tagsets of ten prominent corpora we now arrive at a point where we should reflect a little on the criteria that will have to apply to a tagset to make it cross-linguistic and generally applicable. If we take a closer look at the results of our study described above we can conclude that there are mainly three types of criteria in relation to the development of a tagset; criteria of a linguistic nature, criteria for the data format of the labels and terminological criteria for the label names. Therefore the following points deserve special attention: * coverage of major word classes * addition of relevant feature information * format of the tags . representation of the labels Let us begin with the linguistic criteria. The major word classes should minimally be covered by the tagset. During our projection of the different tagsets we came to the conclusion that there are 12 main word classes 9 that can be distinguished. As we said earlier, it became obvious that across the different tagsets variation as far as the main classes are concerned was not very large. The same was true even for the two non-English tagsets.</Paragraph>
      <Paragraph position="1"> The specification of feature information, however, turned out to be more problematic than the determination of word classes, because the tagsets differed strongly in their degree of detail and/or their research context. From a linguistic point of view, there is a great variety of feature information. First of all, there are subclassiflcations of the major word classes, such as the distinction between common and proper nouns or the different types of pronoun. Furthermore, in relation to verbs, we can think of additional feature information about the degree of subcategorisation (transitivity). Although this appeared not to be a very common feature in the tagsets examined, it can be of great importance to grammar-based syntactic corpus analysis 1deg. Another televant linguistic criterion is to facilitate the inclusion of morphological information, at least 9. see following section 10. see Oostdljk (1991)  for verbs, nouns and adjectives. Features containing this information are for example: number (singular, plural), person (first, second, third), gender (masculine, feminine, neuter), degree of comparison (positive, comparative, superlative) and, especially for the analysis of German texts, case information (nominative, genitive, dative, accustive). From the tagsets under survey it appeared that morphological information is desirable, but there seems to be no agreement on what sort of information should be added. This leads us to a following criterion that is important for the development of our tagset: the format of the tags.</Paragraph>
      <Paragraph position="2"> From our study it has become clear that there is some degree of agreement on major word class information and to some extent also on their subclassification. However, with respect to additional feature information we can see a sort of growing disagreement as the detail of specification increases. This has to do with the various purposes people may have in relation to the enrichment of corpora. On the other hand language-specific items play a role in this context too, for example prepositions in postposition in Dutch n. So the format of the tags has to account for two major aspects: * hierarchical structuring of information * flexibility in relation to the encoding of special features and/or language specific items Hierarchical structuring implies the ordering of information within a specific label. From left to right the degree of detail wiLl increase, so that at the beginning we will find the major word class label, followed by a subclassiflcation, and then several additional features. With ttexibi\]itY= we mean that the label format should be open so that no researcher is limited by our tagset in adding special features that, according to his opinion and/or research aims, are useful or even essential. On the other hand researchers who want to make use of basic word classes only, need only concern themselves with parts of the tags and can ignore the other parts.</Paragraph>
      <Paragraph position="3"> In our attempt to develop a cross-linguistic tagset we make use of a hierarchical data-field oriented coding scheme. The hierarchy in the labels is represented as follows: the level of detail increases from left to right and the different entries are separated by one or more 11. Hij loopt her boa in. (engl.: He runs into the forest.) unique dehmiters. This way of coding also enables researchers to convert the format of the labels in a relatively easy way for their individual needs. For example in the ICE.project, where the tags form the input of a two-level grammar, labels with the format described above can be automatically converted into two-level tags. So this way of coding seems to be attractive linguisticalJy (levels of description) as well as formally (interchangeability). Finally we have to determine how the labels should be represented. Generally we can distinguish between two ways of coding - either a completely numeric label or a mnemonic letter-digit sequence. Although for reasons of readability mnemonic labels are preferable, numeric labels can be used as well, since it is not too difficult to transform one weLl-defined form into another. The advantage of numeric labels is that they are relatively compact and therefore can be stored more efficiently.</Paragraph>
      <Paragraph position="4"> Focussing on the mnemonic way of coding, the question is: what terminology can best be used for the labels. From the tagsets examined we can conclude that there is a commonly accepted linguistic terminology with respect to word-class information, also from a cross-linguistic point of view. In order to provide codes as mnemonic as possible we are of the opinion that this terminology should also be included in our tagset.</Paragraph>
      <Paragraph position="5"> In the following section we present a first step towards a cross-linguistic tagset as a result of our comparative study and illustrate the different major word classes, their possible subclassification, additional feature information as well as the way in which this information can be coded.</Paragraph>
    </Section>
    <Section position="4" start_page="34" end_page="36" type="sub_section">
      <SectionTitle>
Part IV
</SectionTitle>
      <Paragraph position="0"> A basic tagset for Germanic languages In this section we present a basic cross-linguistic tagset for Germanic languages. Those familiar with the coding scheme adopted in the ESPRIT-S60 project will find a number of similarities with this scheme. Thus, as we mentioned above, we have adopted the idea of a hierarchical data-field oriented coding scheme. Moreover the scheme allows for over- null as well as underspecification 1~. Unlike the ES-PRIT tagging scheme, however, it is strictly datafield oriented in order to allow automatic format and addition of extra feature information. The different datafields are always separated by a # symbol, which functions as a delimiter. The hierarchical structure of the tags becomes clear when one looks at the examples we give in order to illustrate the tagging of the different wordclasses. In addition to what has been described above, the tagset also includes so-caUed ditto-tags which make it possible to account for lexical items that consist of more than one word, such as compound nouns or complex prepositions. Ditto-tags take the form of two numbers separated by a slash, where the first number indicates the current part of a compound item, while the second number indicates the total number of words that make up the compound. For example, the three word complex preposition in spite ofis tagged as follows: null in PRP#...#I/3, spite PRP#...#2/3, of PRP#...:~3/3.</Paragraph>
      <Paragraph position="1"> Our basic tagset for the encoding of word class information in corpora comprises twelve major word class categories and an additional category which is intended to accommodate language-specific items. The twelve major categories that are distinguished are listed in Table 2.</Paragraph>
      <Paragraph position="2"> The additional category we shall refer to as open. Each of the word class categories is discussed below.</Paragraph>
      <Paragraph position="3"> Perhaps the categories particle, interjection and PSormulaic expression are not so familiar to some readers. For this reason we give a more detailed description when presenting them.</Paragraph>
      <Paragraph position="4">  refers to a single plural common noun, for example house.</Paragraph>
      <Paragraph position="5"> N#com~plu~acc~fem refers to the same word class as above but contains extra case (accusative) and gender (feminine) information, for ex- null compound i/2, 2/2, etc.</Paragraph>
      <Paragraph position="6"> For example: PN#per#sg#1#nom refers to a first person singular nominative personal pronoun (/).</Paragraph>
      <Paragraph position="7"> PN#rec:#l/2# refers to the first part of a reciprocal pronoun, for example each other.</Paragraph>
      <Paragraph position="8">  In addition, features such as number, case and gender can be added.</Paragraph>
      <Paragraph position="9"> For example: ART#def#sg#acc#masc# refers to a definite singular accusative masculine article, e.g. german den.</Paragraph>
    </Section>
  </Section>
class="xml-element"></Paper>