<?xml version="1.0" standalone="yes"?>
<Paper uid="W98-1010">
  <Title>A Morphological Analyzer for Akkadian Verbal Forms with a Model of Phonetic Transformations</Title>
  <Section position="3" start_page="0" end_page="73" type="metho">
    <SectionTitle>
2 The Akkadian Language
</SectionTitle>
    <Paragraph position="0"> Akkadian is a dead language of the Semitic family. It was used as a native language in Mesopotamia and as a written language in a wider area, including the entire near-east: Egypt, Syria, Palestine, Anatolia, and Persia.</Paragraph>
    <Paragraph position="1"> The name comes from the city of Akkad, once the center of one of the oldest empires in the region. Akkadian was later the language of the Babylonian and Assyrian empires. The oldest documents written in Akkadian date back to approximately 2500 B.C., whereas the most recent ones are from the first century B.C. During the 2500 years of life, the language changed and so has traditionally been divided in several dialects using temporal and spatial criteria (Old Akkadian, Old Babylonian, Medium Babylonian, New Babylonian, Late Babylonian, Old Assyrian, Medium Assyrian, Late Assyrian).</Paragraph>
    <Paragraph position="2"> Akkadian is written using Cuneiform signs.</Paragraph>
    <Paragraph position="3"> Most texts are written on clay tablets, though  a few are written on stone or metal. There are numerous documents, mostly in museum and university reserves, and new ones are discovered every year. The writing system was inherited from the Sumerian language, which was spoken and written in Mesopotamia before Akkadian appeared. Figure 1 shows an example of an Akkadian phrase in Cuneiform. The system combines logograms, syllabograms and determinatives. Most signs have both a logographic value and several syllabographic values, so the system is ambiguous. The phonetic value of a sign is a syllable made of either a single vowel (such as a or i), or a consonant followed by a vowel (nu, ta), or a vowel followed by a consonant (ak, im), or a consonant-vowel-consonant pattern (til, nin). Words are decomposed into syllables, which are written by a sign. For instance, the word iprus (he separated) can be decomposed into ip-ru-us. Note that a single u appears in two adjacent syllables. This system cannot write consonants which are not just before or just after a vowel. For example, prus is not writable. A vowel must be added somewhere. Akkadian is the only Semitic language where all vowels are written.</Paragraph>
    <Paragraph position="4"> As Cuneiform is not very convenient for the modern writer, people interested in Akkadian represent texts with modern writing systems.</Paragraph>
    <Paragraph position="5"> The one we use for this paper and the morphological analyzer is the transcription in an extended Roman alphabet. This system is also used in the grammars, to describe the language.</Paragraph>
  </Section>
  <Section position="4" start_page="73" end_page="75" type="metho">
    <SectionTitle>
3 Akkadian Verbs
</SectionTitle>
    <Paragraph position="0"> Akkadian verbs take a great variety of forms, which differ in five ways: * adjunction of a prefix * adjunction of a suffix * adjunction of an infix syllable or consonant within the root * doubling of the second root consonant * change in the vocalization of the root consonants null The factors of verbal forms are: the root of the verb; the characteristic vowels of the verb; the mode; the aspect; the gender; the number; the stem; the regional factor; the temporal factor. We now detail these factors and give some examples of their influence on morphology. As in the other Semitic languages, verbal roots are composed of consonants. In most cases, there are three consonants in a root, though occasionally there are four. The root consonants appear in all the forms of a verb. Here are some of the numerous forms of parasum, to separate: parasum, iprus, niprusam, paris, uptanarras, putarris, pitarras, ipparrasam, u~apris, pursa, parsaku, parsatina. One can check that every form contains the root consonants p, r, and s in this order. In the following, we will call radicals individual root consonants. null The Akkadian consonants are: b, g, d, w, z, h, t., j, k, l, m, n, s, p, .s, q, r, ~, t, aleph. Aleph is usually denoted by a single quote, but we prefer to write it a in this article to avoid any confusion. null Each Akkadian verb has two characteristic vowels. They are used to vocalize one or another radical in some forms. These vowels, as well as the root, comprise a lexical piece of information. There are 4 vowels in Akkadian: a, e, i, and u. Each vowel may be either long or short. Figure 2 shows examples of variations in forms due to the vowels.</Paragraph>
    <Paragraph position="1"> There are three modes in Akkadian:  phasize the verb's meaning Some examples are given in figure 3.</Paragraph>
    <Paragraph position="2"> The Akkadian language uses five aspects. As in other Semitic languages, aspects do not encode temporal information, but the status of the action.</Paragraph>
    <Paragraph position="3"> * the imperfect denotes an action that is not accomplished. It may or may not have begun yet. It is usually translated into English by the present perfect or future tenses. * the perfect is employed for an action just finished.</Paragraph>
    <Paragraph position="4">  root vowel 1 vowel 2 infinitive imperfect preterit prs a u parasum iparras iprus sbt a a .sabatum isabbat isbat pqd i i paqadum ipaqq/d ipq/d  * the preterit is the aspect of completed actions. null * the stative designates an atemporal state or the lasting effects of an action.</Paragraph>
    <Paragraph position="5"> * the imperative has the same use as in English. null Some examples are given in figure 4.</Paragraph>
    <Paragraph position="6"> As shown by the example in figure 4, gender and number are factors of the verbal form. The differences are in prefixes and suffixes. Sometimes (e.g., stative, imperative), the absence of vocalization of the third radical imposes the vocalization of the second. For instance, the stative third person masculine should be Spars. But it is impossible to write this within the syllabic framework of Akkadian Cuneiform. A vowel u is therefore added.</Paragraph>
    <Paragraph position="7"> Some verbal forms are not conjugated at any aspect: the infinitive and the participle, which act as nouns and are declined as such. The verbal adjective acts as an adjective and it is declined as well. Infinitives, participles, and verbal adjectives do not exist in the subjunctive and they have the same stems (presented below) as the conjugated forms.</Paragraph>
    <Paragraph position="8"> Verbs are conjugated in several subsystems called stems. They are distinguished by prefixes, infixes, and reduplication of the second radical. There are 12 different stems classified in 5 stem groups: * stem I is the basic one. The other stems may be described as a transformation of this one. Example: iprus, he separated (root prs, preterit).</Paragraph>
    <Paragraph position="9"> * stem II (or, D-stem) is characterized by the reduplication of the second radical and by the prefix vowel u. Example: uparris (root prs, preterit).</Paragraph>
    <Paragraph position="10"> * stem III (or, S-stem): the ~ consonant prefixes the root and the prefix vowel u is used. Semantically, it is a causative form. Example: ugapris (root prs, preterit).</Paragraph>
    <Paragraph position="11"> * stem III/II (or, SD-stem): there are both a SS prefix and reduplication of the second radical. It is semantically equivalent to the stem III. Example: u~parris.</Paragraph>
    <Paragraph position="12"> * stem IV (or, N-stem): the root is prefixed by an n. Example: ipparis. In this example, the n prefixed has changed into a p by an assimilation process. This is not a special case: the prefixed n almost always assimilates to the first radical.</Paragraph>
    <Paragraph position="13"> Within each stem group, stems are distinguished by the presence or the absence of an infix. The infix is added after the first radical for stems of groups I and II, and after the infixed g or n for groups III, III/II, and IV. The notation for a stem is made by adding an index to the stem group number.</Paragraph>
    <Paragraph position="14"> * no infix (index 1): all the groups have a stem without infix.</Paragraph>
    <Paragraph position="15"> * infix T (index 2): stem groups I, II, and III have a stem with a t infixed.</Paragraph>
    <Paragraph position="16">  pers. imperfect preterit perfect imperative stative 3 sing. iparras iprus iptaras * -- * paris 2 sing. masc. taparras taprus taptaras purus parsata 2 masc. pl. taparrasa taprusa taptarsa pursa parsatunu 2. fern. pl. taparrasa taprusa taptarsa pursa parsatina  * infix Tn (index 3): all stem groups except the III/II group have a stem with a tn infix. Each stem has a specific semantics that combines with the semantics of the root to give a meaning to a verbal form. The table of figure 5 summarizes the semantics of the stems.</Paragraph>
    <Paragraph position="17"> Though the language evolved during its 2500 year lifetime, verbal forms did not change dramatically. One typical change was the disappearance of the final m from the infinitive. In old Akkadian, the infinitive of the verb prs was parisum whereas in later states of the language it became parasu.</Paragraph>
    <Paragraph position="18"> Akkadian also varied slightly between northern and southern Mesopotamia. The imperfect subjunctive for parasum is given in the figure 6 for Babylonian and Assyrian dialects.</Paragraph>
    <Paragraph position="19"> The combination of all the factors gives a great number of different forms (more than 1000 for each verb).</Paragraph>
    <Paragraph position="20"> There is one more factor of verbal forms that we deliberately separate from the others: it is the phonetic factor. We have already seen an effect of this factor in an example: the form IV.1 was ipparis instead of ~inparis. The n is assimilated to the following p. There are other examples of assimilation (ex: .tt &gt; .t.t). There are also other transformations such as dissimilation, contraction, mutation, etc.</Paragraph>
    <Paragraph position="21"> Phonetic transformations are almost systematic for a subset of consonants called weak consonants. The weak consonants in Akkadian are: ~, w, j. They usually do not appear at all in actual forms. Sometimes, there are traces of these consonants: there is either another consonant or a vowel that comes from the transformations occurring in the context of the weak consonant. For instance, the following transformations occur: aw &gt; u, ay &gt; i, *nw &gt; nn. Sometimes, there is no trace whatsoever: *wis.i :&gt; s.i, *irnnuw &gt; imnu. Sometimes, however, the weak consonants remain: wagabu (but the form agabu is also attested).</Paragraph>
    <Paragraph position="22"> N is a semi-weak consonant: it assimilates easily, but it does not disappear.</Paragraph>
    <Paragraph position="23"> A verb with a weak consonant in its root is called a weak verb. Forms of these verbs are difficult to recognize because all the radicals are not actually in the form. For instance, the reduplication of the second radical is important to identify the stem II and the imperfect aspect. How do we recognize this form and this aspect when the second radical is weak? When the first radical is weak, it is sometimes difficult to find the relevant entry in a dictionary. Some verbs are doubly weak and there is even one verb with all three of its radical weak. Weak verbs are not rare. For instance, in 17 forms collected in a text fragment 1, there are 9 weak forms.</Paragraph>
    <Paragraph position="24"> Here are some examples of weak verbs forms compared with their supposed original form: *j~ip &gt; e.sip, *iwa~ab &gt; u~ab, *banaju &gt; banu.</Paragraph>
  </Section>
  <Section position="5" start_page="75" end_page="79" type="metho">
    <SectionTitle>
4 Morphological analyzer
</SectionTitle>
    <Paragraph position="0"> Recognizing Akkadian verbal forms is certainly the most difficult part of Akkadian morphology.</Paragraph>
    <Paragraph position="1"> We attacked this issue first and the result is a morphological analyzer for the verbal forms.</Paragraph>
    <Paragraph position="2"> The aim of this work is to provide some help to the Akkadian learner. This aid is twofold: first, the analyzer can help students learn strong verb conjugation; second, it helps generate hypotheses about weak forms.</Paragraph>
    <Paragraph position="3"> There are some restrictive hypotheses: * forms are free of suffixes such as pronouns or enclitic particles. Such suffixes are quite frequent in texts.</Paragraph>
    <Paragraph position="4"> * the analyzer is designed to analyze the Old Babylonian dialect. It should also work for some forms of other dialects, but not all of them. Most grammars describe this dialect ICodex Hammurapi, items 228 to 233</Paragraph>
    <Section position="1" start_page="75" end_page="75" type="sub_section">
      <SectionTitle>
Babylonian Assyrian Babylonian Assyrian
</SectionTitle>
      <Paragraph position="0"> male 3 iparrasu iparrasuni iparrasu iparrasuni femel 3 taparrasu taparrasuni iparrasa iparrasani  first and the other ones by the difference to this basic dialect. We have used a corpus for this dialect, namely the Hammurabi code (Szlechter, 1977).</Paragraph>
      <Paragraph position="1"> * the length of vowels is not taken into account. Each vowel may be short or long, but the length is not always explicit in writing. null The analyzer has two levels. The first describes the complete paradigm for strong verbs without any transformation. The second describes transformations that may apply on a given form. The two-level approach of morphology is classical (Sproat, 1992). We adopted a simple model where the two levels are sequential processes with no strong interaction.</Paragraph>
    </Section>
    <Section position="2" start_page="75" end_page="77" type="sub_section">
      <SectionTitle>
4.1 Strong verb paradigm
</SectionTitle>
      <Paragraph position="0"> Conceptually, the first level of the analyzer is a finite language. We have a finite number of parameters: the root is a consonant triple and there are only a finite number of consonants. Within this domain, not all triples are confirmed roots. Each of the parameters discussed in the previous section ranges over a finite domain. If we consider all the combinations of these parameters, they are finite in number.</Paragraph>
      <Paragraph position="1"> Though finite, the language is quite large.</Paragraph>
      <Paragraph position="2"> Enumerating all the forms is not tractable, so a grammar must be written. The natural way to describe such a language is probably a Finite State Automaton (FSA). Conceptually, our grammar of Akkadian verbal forms may be seen as an FSA, but formally, it is a Prolog Definite Clause Grammar (DCG). There are several reasons for this. First, it is a concise way to describe the FSA. A single DCG rule may implement a number of FSA transitions. Second, it gives procedures to use the FSA either for parsing or generation. Third, Prolog is convenient for computations with partial information.</Paragraph>
      <Paragraph position="3"> This grammar is a mid-size grammar, with 162 rules in the current version. A form is described in several slices. At first, we attempted to divide forms into three parts: the prefix, the root, and the suffix. It was just too difficult to design these three parts, so we split the forms in smaller slices. There are now 9 parts: * the personal prefix, which depends mainly on the number and gender of the subject.</Paragraph>
      <Paragraph position="4"> * the stem prefix, which depends on the stem.</Paragraph>
      <Paragraph position="5"> * the infix, which is placed before the first  radical if there is a stem prefix * the first radical. This consonant never varies, but its vocalization does, depending on many factors, including the aspect, the stem, and the infix.</Paragraph>
      <Paragraph position="6"> * the infix, which is placed after the first radical whenever there is no stem prefix.</Paragraph>
      <Paragraph position="7">  the subject's gender and number, or on the verb's mode and aspect.</Paragraph>
      <Paragraph position="8"> Each of these parts of a form is described using a proper non-terminal. The experiment proved that this slicing is tractable, but we believe that it is not optimal. For instance, the description of the third radical is trivial, whereas the second radical with its vowel is complex (33 rules).</Paragraph>
      <Paragraph position="9"> The grammar in its current state implements many, but not all, of the verbal forms. The infinitive, participle, and verbal adjective are not fully implemented. More precisely, the declension, which is the nominal declension for the first two, and the adjectival declension for the latter, are not described. The other forms may all be generated by the grammar.</Paragraph>
      <Paragraph position="10"> This grammar has been carefully tested. It is written in pure Prolog, so it is reversible, and the grammar may be used either for parsing or generation.</Paragraph>
      <Paragraph position="11"> Currently, our grammar does not use any dictionary because we do not have any Akkadian dictionary or any Semitic root dictionary in an electronic form. We did not want to rely on non-existing resources, but the results are not as satisfactory as they would have been with a good lexical source. In parsing mode, the grammar does not actually recognize verbal forms, but gives a possible interpretation of the form. The proposed root has to be checked in a dictionary. null For a delimited corpus such as Hammurabi code, we can make a comprehensive dictionary of verbs. It is easy to interface our grammar with this lexical information. We have not yet tested whether this greatly enhances performance. null</Paragraph>
    </Section>
    <Section position="3" start_page="77" end_page="78" type="sub_section">
      <SectionTitle>
4.2 Phonetic transformations
</SectionTitle>
      <Paragraph position="0"> The second level of the morphological analyzer describes the transformations that may apply on a given form. This level is not a grammar, as we are not trying to recognize a language, but to rewrite words. There is one word in input and one or several in output.</Paragraph>
      <Paragraph position="1"> The focus of our work, designing a model of the transformations due to phonetic phenomena, is quite difficult. We started with a set of rewrite rules (given in (Ryckmans, 1960)) that we completed with other rules when required by a weak form from our corpus. These rules are simple and somehow context-free. The same rules apply on the beginning and ending of verbs. The same transformations apply on infixed t and on radical t. Neither the length of vowels nor tonic accent is taken into account. The model is therefore simplistic and it overgenerates: a rule may be applied even to some contexts where it should not.</Paragraph>
      <Paragraph position="2"> Furthermore, it is very influenced by the set of weak forms that we have considered. Somehow, one can say that the set of rules is sufficient to give the good interpretation of all these forms, among other interpretations that are not all satisfactory. We cannot predict how the set of rules will act on other weak forms. It is likely that several other rules will be added to handle cases not yet encountered; we must consider a large set of examples.</Paragraph>
      <Paragraph position="3"> Some transformations are very systematic (for instance the assimilation of the prefixed n for stem IV) while others are not (for instance, the dissimilation bb &gt; mb). At the moment, the model does not give the probability that a rule will apply (this is a difficult computation). Since the model is non-deterministic, the application of a rule is never mandatory.</Paragraph>
      <Paragraph position="4"> Intuitively speaking, rules are perceived as the formalization of a temporal evolution. The left-hand side of the rule represents the original form, and the right-hand side its form after the passage of time. But in our application, rules are used in the other direction. We have retrieved some attested forms from certain texts, and we want to deduce their original form, which is recognized by the Final State Automaton. null Going from an actual form to its possible prototype is difficult, mainly because the transfor- null mation process tends to shorten words. Consider the typical rule ij &gt; i. If you apply it backwards, you may change any i to an ij. In fact, if most ij became i, few i come from ij.</Paragraph>
      <Paragraph position="5"> Most transformation rules have this quality.</Paragraph>
      <Paragraph position="6"> Using the set of rule as a rewriting system is not adequate because it is does not converge - there is a termination problem. Even with only one rule ij &gt; i, using it backwards would produce unbounded sequences of j. This is not only a computational drawback, it is also phonetically irrelevant.</Paragraph>
      <Paragraph position="7"> Instead of a rewriting system, rules are used to define a transducer. Whenever rule composition seems possible, we just add this composition as a new rule to the set. The transducer has no loop and the transducing process terminates.</Paragraph>
      <Paragraph position="8"> Of course, it is a non-deterministic transducer.</Paragraph>
      <Paragraph position="9"> We implemented the transducer in pure Prolog so that it can also be used to generate possible forms. The results obtained in generation, however, are difficult to interpret. The transformation model is too approximative to produce actual forms.</Paragraph>
      <Paragraph position="10"> The complete code contains 47 prolog clauses.</Paragraph>
      <Paragraph position="11"> The transducer, as it is implemented now, is not very satisfactory: it is a raw and naive implementation that we used to validate our approach. It gives some interesting results that we summarize in the next subsection.</Paragraph>
      <Paragraph position="12"> The main problem encountered at the moment is efficiency. The transducer is non-deterministic and so generates many possible forms. For instance, a weak consonant may be inserted almost everywhere in a word. Prolog's procedural strategy results in enumerating all of the solutions. The transducing process is therefore exponential in the length of the input verbal form (this can be felt during experiments). Whereas the shorter forms (often the weak forms) are processed quickly, the computation of the complete set of solutions for the longer forms (up to 10 characters) may last several hours.</Paragraph>
      <Paragraph position="13"> While the quality of the results of the morphological analyzer are quite satisfactory, its efficiency is not. The first level of the analyzer, namely the finite state automaton is efficient, but the transducer is not. We view several ways to solve this problem.</Paragraph>
      <Paragraph position="14"> The first solution consists in changing the procedural way to execute the transducer, especially the way non-determinism is handled.</Paragraph>
      <Paragraph position="15"> With Prolog, the alternative solutions are found one after the other, using backtracking. An alternative solution would involve computing a single data structure to represent all the solutions, with the common parts of the different solutions shared. This would break down the complexity, since the rules encoded in the transducer apply independently on the different part of the input string. A regular expression would be the natural data structure to represent a set of strings with sharing. This form is suitable for parsing with the FSA. Parsing in this case consists in computing the intersection of two regular languages. This is a well-known operation (see for instance (Hopcroft and Ullman, 1979)).</Paragraph>
      <Paragraph position="16"> Another idea to improve efficiency is to predict where the transducer should insert weak consonants. This could be clone by a rough analysis based on consonant count.</Paragraph>
    </Section>
    <Section position="4" start_page="78" end_page="79" type="sub_section">
      <SectionTitle>
4.3 Results
</SectionTitle>
      <Paragraph position="0"> We have developed and tested the morphological analyzer using verbal forms from several sources. We collected 122 forms in (Caplice and Snell, 1988), 87 strong and 35 weak. We also used 54 forms found in the Hammurabi Code, mainly in articles 185 to 233, but also from other various articles. This is only a small subset of the verbal forms occurring in the code.</Paragraph>
      <Paragraph position="1"> The first result is that all these forms are recognized by the morphological analyzer with the relevant interpretation. This is not a surprise, since we augmented the transducer in order to obtain the desired result. The point is: does the analyzer give wrong interpretations? It does indeed, sometimes, but its behavior is generally correct.</Paragraph>
      <Paragraph position="2"> First, we tested 60 strong forms. On these, only two have been interpreted as possible weak verbs: s.abat and ritgum. For s.abat, three possible roots were identified: sbt (which is the correct hypothesis), .scab and .sbc~. The interpretations given by the analyzer for the two later are the following: the form is taken as a stative, feminine, third, person, stem I, s.a~bat &gt; s.abat and s.abc~at &gt; sabat. This seems plausible. Concerning ritgum, the ambiguity comes from the t which is a radical but may be interpreted as an infix and from the m which is the mark of the ventive but can be seen as the third radi- null cal. here again, the proposed root is plausible.</Paragraph>
      <Paragraph position="3"> Surprisingly, some forms very close to the two ambiguous ones are not ambiguous.</Paragraph>
      <Paragraph position="4"> The most ambiguous form in the data we considered is iddu, for which 15 roots have been computed. The right explanation of the form is *indiju &gt; iddu. There are two transformations: assimilation of the n and contraction of iju. It is a typical example of a form that is difficult to understand for the Akkadian learner. Even if there are many hypotheses, the answer given by the system may help in such a case. The help would be much better if the system had a complete Akkadian root dictionary.</Paragraph>
      <Paragraph position="5"> The verb alakum which is sometimes said to be irregular (see, for instance, (Heise,)) is treated as the other verbs by our system. We followed the interpretation of its forms (radical c~ assimilated to the radical l) found in (Ryckmans, 1960). This gives satisfactory results. The main weakness identified so far is in the aspect discrimination between imperfect and preterit for weak verbs with a second radical weak. For these verbs, the main difference between the two aspects, namely the second radical reduplication, is not perceptible. In that case, the vowel is significant. For instance, the verb kanum has a root kwh. The preterit is *ikwun &gt; ikun and the imperfect *ikawwun &gt; ikan. The morphological analyzer proposes either preterit and imperfect as possible aspects for ikun. It is not possible to prevent the mutation of the w in u in this case, because such a mutation sometimes occurs in other contexts.</Paragraph>
      <Paragraph position="6"> Generally speaking, the morphological analyzer gives the right solution, but also proposes other ones. These other ones are often acceptable, but sometimes, as shown by the latest example, they are not.</Paragraph>
    </Section>
  </Section>
  <Section position="6" start_page="79" end_page="79" type="metho">
    <SectionTitle>
5 Conclusion
</SectionTitle>
    <Paragraph position="0"> The work we have done so far shows that many Akkadian verbal forms can be interpreted using a single conjugation paradigm and a phonetic transformation model.</Paragraph>
    <Paragraph position="1"> We think that our approach is a good one for Akkadian, due to the language peculiarities. Is this approach well-suited for other languages? We do not know.</Paragraph>
    <Paragraph position="2"> The basis for our work is that we model phonetic transformations for a language with a phonetic writing. The Akkadian writing system is phonetic and syllabic. As far as we know, it is not the case of other Semitic languages. For instance, they do not transcribe vowels. The results obtained so far show that the vocalization in Akkadian breaks down ambiguity.</Paragraph>
    <Paragraph position="3"> It is not obvious that our approach is suitable for languages other than Akkadian, for which it is quite convincing.</Paragraph>
    <Paragraph position="4"> The work described here is in progress. We have to study the work done for the morphological analysis of the other semitic languages. We also have to search for a better way to perform the second level of the analysis.</Paragraph>
    <Paragraph position="5"> The morphological analyzer presented in this paper could be enhanced by expressing the transformation rules more contextually and by coupling the two levels.</Paragraph>
  </Section>
class="xml-element"></Paper>