Unicode, Ancient Languages and the WWW

Carl-Martin Bunz, M.A.

Universität des Saarlandes, Saarbrücken, Germany

Prof. Dr. Jost Gippert

Johann Wolfgang Goethe-Universität, Frankfurt/M., Germany

0. Herodotus, the famous Greek historian of the fifth century B.C., reports how the Greeks took over and adapted the Semitic alphabet:

Herodot 5,58

"These Phoenicians who had arrived together with Cadmus -- of whom the Gephyraeans were a part -- after having settled in this country, imported many kinds of learning to the Greeks and especially characters which, as I think, the Greeks had not before -- first those (characters) that all the Phoenicians used. Afterwards, as time went on, simultaneously with the change of language, they altered also the shape of the characters. At this time, in many of the places around them the Ionian Greeks were dwelling who, after having been taught by the Phoenicians, took over the characters, modified few of them in shape and used them. In doing so they spoke of them by the name of Phoenician -- as indeed was but just, seeing that the Phoenicians had introduced (the characters)."

We shall not discuss the historical correctness of Herodotus' record: His account of the transfer of a Semitic script to early Greece is essentially right. The import of this writing technique -- which was not the first one used for the notation of Greek -- was to entail enormous consequences; if it had not taken place, the world would probably not have seen the development of the Latin alphabet and its wide proliferation. By these few sentences Herodotus describes the process of script adaptation which many language communities before and after the historian's time performed. There is an ethnic group that settles in an foreign country; they bring with them their language and their writing system, using a certain script. Later on, they adopt a different language and adjust the repertoire of symbols their script disposes of. Further, there is another language community dwelling in the neighbourhood that is taught this script by those who first practised this method of writing, and again adjustments of the symbols used are necessary. But the character set, although modified by the transfer from one community to the other, is given the name of those who first introduced and taught its usage.

Herodotus, however, does not explain the crucial point of the change the Semitic inventory of consonant symbols underwent when it was adapted to the notation of Greek. The Semitic script units operate as symbols of syllabic classes, whereas in Greek the first attempt was made to represent in writing what we today call phonemes, when the adaptors defined vowel symbols.

Writing now has a long history, its dissemination and diversification exactly in the way Herodotus describes is considerable seeing that there are but a few original notation principles of language data, i.e. ideographic, syllabic and alphabetic writing. Language communities all over the world developed writing systems making use of one or more of these principles. Nevertheless, in most areas one writing system prevailed.

Those language communities of the past that used writing recorded what was important for them in daily life, but also contents they regarded as worth conserving for later generations, such as legal and religious texts. The material they wrote on, the climate and the political history of the region decided upon the preservation of these documents. The textual monuments which survived since man first developed writing at about 3000 B.C. constitute in their entirety the world's memory of civilization: Sumerian agricultural records and religious poetry, a Maya calendar, reports of Hittite kings, the liturgical texts of the Zarathustrians, etc. etc.

Historical and comparative linguistics provides the scientific basis for a linguistic analysis and understanding of these texts. Furthermore comparative linguistics interprets the linguistic data and determines the genetic relationship between the languages as well as between different historical stages of one language. Like this, comparative linguistics traces the history of human verbal expression.

A 16-bit encoding space is necessary at least for parallel processing and transfer of linguistic data using distinct encodings for different scripts. We do hope that, within the WWW, comparative linguistics will soon be able to supply the ancient texts in electronic form -- alongside with comprehensive philological and cultural information -- so that the scientific and scholarly community as well as everyone else interested in the matter can explore the memory of the world, provided a worldwide character encoding system exists that supports historical data as well as national and commercial standards.

Unicode, the Worldwide Character Encoding, primarily represents a standard designed as to process the scripts of the major modern languages of the world, on the basis of the norms national ISO commissions elaborated. Like this, the architecture of the encoding is based on synchronic scripts. On the other hand, since the publication of Unicode Standard Version 1.0 the designers of the encoding maintain that historic scripts will be included as well. But as soon as historic scripts are considered the historic development of the scripts has to be taken into account as well. This is to say that the most important structural principle of the Unicode architecture, the Character/Glyph Operational Model, must be open for a historic interpretation too. Before having reconsidered the methodical impact that the encoding of historic scripts bears, it is not very reasonable to simply integrate historic scripts into the patterns the Unicode encoding strategy has established up to now.

First we have to see that administering the world's written memory of civilization does in no way mean to establish norms or standards the establishment of which is often influenced by extra-linguistic factors. Rather, we aim at preserving all the information that the ancient records provide: When processing historic documents, we must never modify the script and text data at our convenience. What we intend to demonstrate in this paper, is precisely the challenge Unicode faces as soon as historic scripts will be taken into consideration. The flow chart in Table 1 may give an idea of the constellation.

We hope to show up the difficulties in all their complexity and intertwinement that will arise when Unicode is applied to historical data and to historical and comparative linguistics dealing with them. Our presentation is meant to contribute to a discussion which in our view is really necessary.

1. Let us enter again into the sphere of Herodotus, i.e. the emergence and further development of the Greek script. More than 1000 years after its development from the Semitic basis, the Greek script was borrowed by the Georgians and Armenians. Table 2 gives a synopsis of these scripts; they are encoded in Unicode in the character blocks U+0530 U+058F and U+10A0 U+10FF respectively.

1.1.1. Like the Greeks of the archaic period, the inventor(s) of the Georgian Asomtavruli alphabet on the one hand followed exactly the order of the model alphabet so as to keep the numerical values of the characters, assigning to each graphical unit a similar phonetic value in Georgian -- with a few exceptions where the Greek value had no correspondence in Georgian. On the other hand he established an appendix of characters that represent phonetic units alien to Greek, just as the Greeks themselves did before. In this appendix, the sequence of characters reflects a specific articulatory organization. Like this, the Asomtavruli script represents an alphabet which, although considerably differing in shape, is mapped perfectly onto the model alphabet, with a number of additional units (cf. Table 2, fourth column).

In ancient Georgia, writing was practised exclusively for clerical purposes. Asomtavruli as a monumental script was designed for inscriptions first (cf. Table 3), but later on it was used in manuscripts too (cp. e.g. the lower text of the palimpsest manuscript reproduced in Table 4). By the eighth century A.D. the scribes began to style larger letters to mark the beginning of paragraphs, and so they introduced the minuscule-majuscule distinction into the writing system. Consequently, in the course of the ninth century, a half-cursive script variant was derived from the Asomtavruli minuscule characters, later named Nusxa-Xucuri (cp. the third column in Table 2 and the upper text of the palimpsest in Table 4). In the tenth century A.D. an other important change happened in Georgian writing: now Mxedruli, the proper cursive script variant, was developed especially for secular use (cf. the second column in Table 2). At that time, the Georgian script system had attained its greatest complexity: In manuscripts we find Asomtavruli characters operating as uppercase letters alongside with Nusxa-Xucuri characters, while "colophon" notes were often written in Mxedruli; but only in the 20th century Asomtavruli letters were used as Mxedruli "majuscules". This attempt of calquing European writing usage was introduced by Georgian scholars only and it had no impact on everyday's use.

From all this we conclude that the encoding of Georgian as present in the Unicode block U+10A0 -> U+10FF (cf. Figure 1) is not sufficient when handling historic texts written in that language, the block containing Asomtavruli (U+10A0 -> U+10C5) and Mxedruli (U+10D0 -> U+10F6) characters only. The script variant Nusxa-Xucuri must be encoded as well because, as was shown above, the three scripts were clearly distinguished in their application during the history of Georgian writing, existing side by side for a long period. Working with the Unicode Georgian character block in its present form would mean, e.g., to encode an ancient inscription (like the one printed in Table 3) or an ancient manuscript with the Mxedruli characters, the Asomtavruli script units being glyph representations of these code elements; the Asomtavruli characters of the Unicode block would have to be reserved for majuscule forms of Asomtavruli letters instead (cf. Figures 3, 4, 5). The situation becomes even more complex when we want to encode a document that contains all three scripts, as described above. This is why we propose to establish three character blocks for Asomtavruli, Nusxa-Xucuri and Mxedruli respectively, each block comprising (Asomtavruli) majuscule characters.

1.1.2. The Armenian script had also been drawn up on the model of the Greek alphabet and so it can be mapped onto the original to a certain extent too, although the shapes of the matching Armenian letters are altered to an even higher degree than in the Asomtavruli alphabet. But the designer of the Armenian script did not aim at aligning the numerical values of the characters with those of the Greek patterns (rather he even aimed at establishing differences as to the Greek series). Interspersing all the additional characters between the characters inherited from the Greek stock, he established a new sequence so that a mapping of the complete Armenian alphabet on Greek is not possible. Like this, a sharp difference arose between the Armenian and Georgian scripts in that the latter only could be treated as glyph variants of the Greek alphabet in Unicode. In this sense Georgian behaves just as Coptic does which has in fact been integrated as an extension of Greek (U+0370 U+03FF).

1.2. Up to this point, we have considered the processing of original scripts. Now we should like to enter into the issue of transliteration and transcription. In order to avoid misunderstandings, let us first state how we interpret what we shall call `transliterative processing' and `transcriptive processing' with respect to Unicode characters. This presupposes that we define clearly what we are doing when handling linguistic material.

Linguistic transliteration and transcription are strategies that aim at rendering language data by mostly Latin letter forms in order to achieve a common notational level which enables analysis and comparison of the data. Greek letters and modifications of Latin letters are also used. Most of these signs are modified by diacritics marking additional distinctions the simple letter forms are not able to convey. The difference between both strategies can then be described as follows:

Transliteration is a method of miming original scripts with the help of symbols as described above. This miming pertains to the graphical units the script uses, without regard of their functional (i.e. phonetic) value. In certain cases, miming may even extend to iconic imitation when e.g. the form of a diacritic mimes a graphical element that belongs to the original script.

Transcription is a method of rendering the original script after a graphemic and phonemic analysis of the linguistic data involved. That is, transcription presupposes a functional interpretation of the material. Therefore, transcription normally uses fewer graphical units than transliteration and is totally independant from graphical shapes of the original script.

Both strategies are indispensable in linguistic work. Especially in the case of historic languages, transliteration can be used to render the linguistic material as soon as the original script has been segmented -- it does not presuppose, however, that the writing system in concern has been fully understood. We represent linguistic data of dead languages in transliteration in order to provide the data in a shape that is open for an interpretation aiming at establishing the functional value of the graphical units as well as their phonemic and morphemic values.

Transcription in its turn is a vehicle for transporting linguistic interpretations no matter whether the scholarly community or part of this community agree upon this interpretation.

Phonetic transcription as normalized by the IPA must be kept apart from transliteration and transcription as practised in historical linguistics, because the application of the well-defined IPA symbols would demand an insight into the phonic reality of the ancient languages involved. Although linguists often aim at establishing the phonetic system of ancient languages in great detail, they will hardly be able to be as precise as linguists dealing with modern languages. Historic transcription, however, aims at a level of phonic characterization that leaves room for further investigation. Furthermore, the repertoire of IPA characters would not suffice to mark the distinctions a specific transcription needs. If we applied IPA diacritics in historical linguistics, we should introduce articulatory distinctions nobody can witness to -- and like this we should leave the area of scientific work. Within historical linguistics, phonetic notation plays an important role only when the development of word forms in language change has to be analyzed and demonstrated.

It is self-evident that both methods, transliteration and transcription, have to be encoded unambiguously when linguistic data are stored and linguistic investigations are carried through electronically. Now we try to transfer this constellation into the context of Unicode architecture:

A transcriptional system to represent a given writing system constitues a writing system of its own right. In a Unicode conforming environment, a transcription character set is made up by code elements, i.e. Unicode characters, which are found mostly in the Latin Character Blocks of the encoding, occasionally in other Character Blocks (Greek, Cyrillic). In analogy with the processing of other latin-based writing systems, the Unicode characters assume their specific semantics only within the transcriptional system in question.

As for transliteration, there are two principally different ways it can be handled using Unicode:

(a) Either a transliterational system of a given writing system is treated in the way that it constitutes a writing system of its own right, exactly as a transcriptional system does. A transliteration character set would then be made up by code elements, i.e. Unicode characters (cf. above on transcription).

(b) Or transliteration, as opposed to transcription, can be considered as a glyph representation of the original script.

Linguists will certainly favour the second solution which reflects the nature of transliteration in an ideal manner. But such a method presupposes that the character blocks of the original scripts do in fact exhibit all units required (cf. Figure 8).

After this theoretical excursus, let us return to Georgian and Armenian. We consider Georgian first: In Table 2, the column exhibiting a Latin-based rendering of the Georgian script units is labeled `Transcription' because the choice of a definite Latin character and a definite diacritic presupposes a phonemic analysis of the language. When we transcribe, e.g., the affricates written , , , by the symbols c, , , , a phonemic classification of the sounds is implied, the diacritics representing articulatory features in a systematic manner. Thus, the rendering is wholly independent of the shape of the original script units. The most effective processing of Georgian transcription would be based on an exact correspondence of one original script unit to one transcription unit each. Other solutions are more complex as they require a one-to-n mapping.

Armenian is usually rendered by a transcription similar to the one used for Georgian, but it differs in the method how certain articulatory features are marked by diacritics. In transcribing Georgian stops and affricates, we mark the absence of aspiration, or the optional presence of glottalization respectively (p ­ , t ­ , k ­ , c ­ , ­ , etc.); while in Armenian, where the same phonemic functionality prevails, the presence of aspiration is marked ( ­ p, ­ t, ­ k, ­ c, ­ , etc.). When preparing Georgian and Armenian transcription with the view to a Unicode conforming character processing, we should not, however, level these transcriptional systems, because the difference of notation facilitates the distinction of the language material concerned.

In the case of Georgian and Armenian, transliteration in the sense defined above bears little importance because the script units quite perfectly match the phonemic units modern linguistic analysis determines. A transliteration of Georgian like the one printed in Table 3, indicates the spacial distribution of the characters on the monument; there is no need, however, to use symbols different from the transcription characters. Regarding oriental languages, this seems to be exceptional as we will see later on.

2. At a certain time during the Sasanian empire (224-651 A.D.), Zoroastrian priests felt the need to fix the recitation of their sacred texts, until then handed orally, by writing, in a manner that would preserve as many articulatory details of the pronunciation as possible. With this aim in view, a phonetic alphabet was designed and the corpus of Zoroastrian sacred texts, known as Avesta, was written down in this alphabet.

2.1. In Unicode Technical Report 3 (Exploratory Proposals), p.70-74, an integrative encoding proposal is exhibited for both Pahlavi and Avestan scripts (cf. Figure 9). Indeed, the inventor of the Avestan repertoire of phonetic symbols exploited the character stock of the so-called Pahlavi script, i.e. the Middle Persian Book Script currently in use at his time. But the Pahlavi characters could serve as a sort of socket symbol set only, given that most of them were multi-functional to a great extent (e.g. there was just one character denoting the sounds n, r, w and the end of words); for Avestan, however, each script unit was to be assigned a precise phonetic value. This is why, in Unicode, Avestan must in any case be encoded in a separate character set, all the more since most Avestan manuscripts offer Avestan and Pahlavi (and even Sanskrit) texts side by side (cf. Table 7).

Table 5 shows the development of the Avestan alphabet. It makes evident that the designer of these characters had been guided by his keen phonetic observation, quite similar to the modern phoneticians who created the IPA notational standard.

2.2. According to the structure of the script, in analyzing the Avestan language, linguistics starts from an unambiguous transliteration system the units of which mime the original characters on the basis of the Latin and Greek scripts. The choice of Latin or Greek basic characters and diacritics mirrors both the phonetic values (p ­ , f etc.) and the graphic derivation ( -> , , /y; k ­ , ; etc.) of the original script units ( -> , , ; -> ,). In two cases (, ), the form of the diacritic mark even imitates the distinctive graphic element of the original script (, ). When Avestan transliteration is to be processed using Unicode, the transliterative processing strategy (b) as defined above (p. 6) would be the most favourable: the rendering engine should then be able to choose the transliteration symbols among the glyph inventory of each code element of the Avestan character set. Method (a), on the contrary, would provide transliteration as plain text, and we would have to add 10 precomposed combinations to the Latin character blocks (, , , , , , , , , ) in order to establish one-to-one correspondences. But we think that a concept which shifts transliteration to the level of rendering proves more suitable: Since transliteration symbols serve to represent the original script, it seems quite natural to encode the original script units only.

Linguistic transcription of Avestan texts represents the phonemic interpretation achieved by internal and external reconstruction. This should of course be handled on the level of plain text because transcription, as stated above, behaves like any other writing system which uses primarily Latin characters (e.g. Czech, Polish, Icelandic, etc.). Table 6 illustrates the different layers of text handling in historical linguistics applied to the example of Avestan: (1) digitizing of the original script document, (2) abstraction of representative graphical forms of the original script units (i.e. creation of a computer font of the original script), (3) transliteration, (4) transcription, (5) translation.

3. For the notation of another Old Iranian language, viz. Old Persian, a specific cuneiform syllabary had been established at the time of Darius the Great (522­486 B.C.). The creation took place in the context of cuneiform writing still prevalent in Ancient Mesopotamia at that time, but the syllabary essentially differs from the arrangement of the older cuneiform systems (cf. the synopsis in Table 8). As far as the shape of the script units are concerned, Old Persian cannot participate in a cuneiform `unification' which proposes itself when these scripts are integrated in Unicode, the Old Persian characters consisting in individual constellations of cuneiform elements. Additionally, the Old Persian writing system principally operates with two sets of `consonant characters' and `vowel characters' respectively; only a few symbols are apparently defined as to involve the vowels i or u, the other `consonant characters' being understood as either bearing a vowel (mostly a) or no vowel at all. In some way, this reflects the structure of the phonemic system of the language. But as far as we can guess, the writing system was elaborated during its application, so that the existence of the characters di, mi, vi, ku, gu, tu, du, nu, mu, ru cannot be motivated from a systematic point of view.

3.1. When an encoding for the Old Persian cuneiform script is planned, the first question concerns the ordering of the characters. In processing, of course, the order of code elements is irrelevant, but, on the other hand, the arrangement of any character set anticipates a certain collation. Given that no traditional order of the Old Persian `alphabet' exists, we are free to determine the sequence of the characters ourselves from a linguistic point of view. The encoding proposal in UTR 3 p. 76-79 would not be the best solution. Instead, we would like to propose an order based on the same phonetic principles as used for most syllabaries, ancient and modern (cf. Table 8).

3.2. Transliteration of Old Persian depends on the values we attribute to the `syllabic' characters. Consequently, there is a `consonantizing' and a `vocalizing' transliteration strategy (cf. Table 9). In fact both methods can be equally justified. The rendering of the `consonant characters' of the script with an appended a-vowel throughout reflects the fact that they were modelled on the basis of typical cuneiform characters, representing sequences of consonant plus vowel (CV). The consonantizing method, on the other hand, presumes that in the Old Persian writing system the characters are conceived as generic symbols that mark syllabic classes defined by the consonants involved (k for ka, ki, ku etc.). If we were to decide on the notation according to a given context, we would no longer produce transliterations but transcriptions, switching to and fro between `vocalizing' and `consonantizing' renderings of symbols like read once as k, once as ka.

Transliterative processing of Old Persian using Unicode could be a glyph representation process as considered above for Avestan, provided the rendering engine enables us to implement a one-to-n mapping of original script and glyph units and to choose between the `vocalizing' and `consonantizing' methods. Otherwise we should have to encode the transliterated text just like the transcribed text, using the characters of the Unicode Latin Character Blocks (cf. Figure 10).

4. Encoding transcriptional and transliterational materials (according to method (a), p. 6), historical linguists will always be tempted to make use of the precomposed characters already existing in Unicode. This would be an economical approach. A first evaluation made in the course of the development of transcription fonts for Indo-Europeanists (cf. http://coli.uni-sb.de/~cmb) with the view of administering the TITUS text database (cf. http://titus.uni-frankfurt.de/texte/texte.htm), showed that on the basis of existing Unicode characters, transcription and/or transliteration of texts from very many ancient Indo-European languages can be achieved in accordance with the principles established and adopted by historical linguists. An ever increasing set of additional characters (such as the one printed in Table 10) will be required though whenever we aim at representing (or rather mirroring) all aspects of human textual tradition.

4.1. A rather hybrid method would consist in using the existing precomposed combinations and to encode the missing combinations analytically (cf. Figure 11). This would establish enormous difficulties concerning sorting and other processes where overlay modules are required that define the value of character sequences containing superscripts and the like in the specific collation. Although this is a possible strategy, it would result in a huge amount of manpower to be invested before Unicode can be used for these transliteration and transcription systems. An even worse encoding strategy would consist in dispensing with precomposed Unicode characters altogether. In addition to the collation complexity referred to above, this would entail an even higher dependence on software development than the hybrid strategy discussed before.

4.2. As far as transliteration is concerned, less problems would occur if we could widely adopt method (b) consisting in a glyph mapping onto the original script units. But as we have stated above, it is necessary in this case that the original script is fully encoded with all its units. This can easily be demonstrated by one other example. If we want to represent the divergent witnesses of the Middle High German Nibelungenlied, we have to face not only special characters such as but also a nearly indefinite number of seemingly ad-hoc combinations of plain and superscript characters such as , (cp. the sample passage shown in Table 11). If we compare the structuring of the Unicode Latin Character Blocks, these combinations could easily be treated as separate units of the original script just as the additional characters of the Czech alphabet (, and the like) are treated in Unicode; for an adequate rendering, this would be the best solution. The same holds true then for their transliteration, for in the case of the Nibelungenlied manuscripts the original script and the transliteration is virtually the same because the Medieval Latin script is nothing but a style variant of the plain Latin alphabet just as the transliterational script is. Of course it would be possible to use Composed Character Sequences in such cases too; but this would require at least an additional set of superscript characters to be treated as diacritics.

5. We should be glad if our contribution might have helped to show that whenever ancient languages and historic scripts are concerned, a lot of reasoning is necessary before they can be adopted by Unicode in a way sufficient for both a reliable representation of original documents and linguistic analysis. It goes without saying that specialists from throughout the world should work together with respect to this aim.

Copyright for this text: C.M. Bunz / Jost Gippert, Frankfurt a/M, 9.3.1997. No parts of this document may be republished in any form without prior permission by the copyright holders.