In this paper, we describe an ongoing project of developing a treebank for Amharic. The main objective of developing the treebank is to use it as an input for the development of a parser. Morphologically-rich Languages like Arabic, Amharic and other Semitic languages present challenges to the state-of-art in parsing. In such language morphemes play important functions in both morphology and syntax. In addition to the existence of high lexical variations due the morphology, Amharic has a number of clitics which are not indicated with any special marker in the orthography. Considering the status of Amharic resources and challenges to the existing approach to parsing, we suggest to develop a treebank where clitics are separated manually from content words and annotated semi-automatically for part-of-speech, morphological features and syntactic relations.
A treebank in general can be viewed as a
linguistically annotated corpus that includes
grammatical analysis beyond part-of-speech level
Parsing, on the other hand, is a process of recognizing input sentences and identifying units like subject, verb, and object. It is used to determine the interaction among these grammatical functions. Deep parsing, sometimes referred to as deep processing, is a process whereby rich linguistic resource is used to give a detailed (or deep) syntactic as well as semantic analysis of input sentences. Such systems are basic components and language specific resources for any Natural Language Processing applications which require deeper understanding of Natural Languages.
The structure of the paper is as follows: after we give a brief discussion on the motivation for the development of the treebank in Section 2, the challenges of Amharic to treebank development will be discussed in Section 3. Section 4 deals with the existing resources and their limitation. Section 5 is devoted to the proposed solution and finally the paper will concluded in Section 6. 2.
The main objective of the project is to develop a
parser for Amharic sentences. For parser
development there are two major approaches.
These are grammar-driven (or rule-based) and
data-driven (or statistical-based)
Motivated by the hypothesis that humans
recognize patterns and phrases that have occurred
in past experience
Both methods have their own shortcomings
and benefits. Grammar-driven methods are
known to be linguistically precise, but have a
problem of robustness and ambiguity. On the
other hand, data-driven methods are good for
developing wide-coverage parsers rapidly.
However, the accuracy of the parser depends on
the magnitude of the training data and the
existence of accurate language specific resources.
As data-driven methods depend on the size of
corpus, it is also subject to problems of
robustness, the ability of a system to give a
certain analysis for a new or unseen input
sentence
On the other hand, in choosing which
approaches to follow for the development of
parsing system, we need to consider the status of
the language. In general, languages can be highly
resourced or less resourced. Such distinction is
based on the availability of tools and electronic
data prepared in a language. Regarding Amharic
there are initiatives to develop a large corpus. It
is also stated that language processing research
for Amharic has shown some progress in recent
years in both corpus and basic Natural Language
Processing (NLP) tools development
In developing a treebank for languages which are morphologically rich and less resourced, there are a number of challenging issues which should be given due attention. In the following sections we provide a summary of the major challenges we have noticed so far.
Amharic being one of the morphologically rich
languages presents a challenge to the area of NLP
in general and to parsing in particular
An orthographic word in such a language
which is delimited by white space, may be a
combination of one or more function words and
inflectional morphemes. For instance, an
Amharic orthographic word like ከየቤቱና
/kəjjəbetunna/ “from each house and”, includes
the preposition ከ- /kə-/ “from”, reduced form of
the distributive marker እየ- /ɨjjə/ “each”, ቤት /bet/
“house”, the definite marker -ኡ /-u / “the” and the
conjunction -ና /-nna/ “and”. The clitics like
preposition, conjunction, auxiliaries, etc. have
syntactic roles that indicate grammatical relations
with the content words
State-of-the-art in parsing technology has
been using data-driven systems. However, due to
high morphological variations that exist in
morphologically rich languages, it is impractical
to observe all such variations in a given annotated
data. As a result, data-driven systems do not
guarantee recognition of all morphological
variations
The Amharic script is called Ge’ez or Ethiopic where a consonant and a vowel are represented by one symbol. In most literature, the writing system is considered as alpha-syllabic. But some people argue that there may not be a one-to-one relationship between a grapheme and a syllable. In other words, each symbol in Amharic orthography represents a CV syllable. However, a syllable in Amharic may have CVC structure. Problems related to the writing system which could be worthy considering in the development of the treebank are:
A. Gemination or consonantal length - in Amharic gemination is phonemic in the sense that it could bring meaning change. However, the writing system does not handle this feature. In some cases, gemination can be used to convey grammatical information. Such case is very common in the construction of relativized verbs.
B. Compound words - As can be observed in other languages, compound words are written in three ways; with a space between the combined words (ቡና ቤት /bunna bet/ “bar”፣ አየር መንገድ /ʔajər məngəd/ “airline”), separated by a hyphen (ስነ-ስርዓት /sɨnə-sɨrʔat/ “procedures”፣ ስነጥበብ /sɨnə-t‘ɨbəb/ “art”) or written as a single word (ቤተክርስቲያን /bətəkɨrsɨtijan/ “church”፣ መስሪያቤቶች /məsrijabetoʧʧ/ “offices”).
C. Syntactic words – words which are separated by white-space (semicolon in old documents) may be coupled with functional words like a preposition, conjunction or auxiliaries. Thus, an orthographic word may be a phrase (ለሰው /ləsəw/ “to human”), a clause (የሚገኙትና /jəmmigəɲɲutɨnna/“and those that are found/available), or even a sentence (አልመጣችም /ʔalmət‘t‘aʧʧɨm/ “She did not come.”)
All such features need to be addressed in processing Amharic texts. Some of the above problems call for standardization efforts to be made whereas others are due to the decision to write what is in mind and what is actually produced.
In recent years, language processing research on
Amharic has grown. This is partly because of the
existence of a reasonable-size Part of Speech
(POS) tagged corpus and the development of
Morphological Analyzer. The tagged corpus is
news corpus from Walta Information Center
(WIC). It is manually tagged by the staff member
of the Ethiopian Languages Research Center
(ELRC). It consists of 210,000 tokens collected
from 1065 news documents
Another important resource is a
morphological analyzer called HornMorpho
In the previous sections we have shown that the existing corpora and tool cannot be used for our purpose due to their limitation of scope or focus. For our purpose, we want to analyze both lexical and function words. Thus, we propose the development of a treebank where both content and function words are separated. In other words, we propose to separate function words or clitics from their phonological host. Even though, the distinction between clitics and affixes are debatable, for our purpose, the following list elements are considered as clitics.
1. Prepositions 2. The Possessive marker or pronominal genitive markers 3. Definite marker 4. Accusative marker 5. Conjunction 6. Negation 7. Auxiliaries 8. Relative pronouns 9. Nominal clause marker 10. Subject and object pronominal
agreement markers
The above elements should be separated from content words. To do so, we have collected five thousand sentences from different sources which include grammar books, biographies, news, fictions, science books, law and religions. All the collected sentences were manually checked for spelling errors. These sentences will then be annotated at different levels. Before the annotation, we will decompose words into smaller meaningful units without loss of their basic meaning. As it is indicated above, in Amharic writing system, those listed function words are written together with content words. Thus, we should segment the two. Such segmentation will be done following a guideline which we have prepared.
The guideline gives what should be considered in the manual segmentation. For instance, complex word in the text, that is a combination of a content word and one or more clitics should be embraced by a bracket. This helps to keep track of the input word which is segmented. When a complex word is segmented, the elements in the orthographic words may not always be the same. They may be modified or reduced in some way. For instance, the word ወደሚገኝ /wədəjəmmigəɲɲ/ “to which that is found” will be segmented into ወደ_የ_እም_ይ_ተገኝ wədə_jə_ɨmm_jɨ_təgəɲɲ. From this example we noticed that the orthographic word is a reduced form. When a preposition precedes the complimentizer የ jə (relative marker), the form will be reduced into ም mm. Thus, we need to keep the input orthographic word using the bracket and show the components that make up the form in the segmentation.
The guidline also provides on how to check wheather a certain form is a clitics or part of the content word. Clitics that we have listed above, in most cases are short forms which may be part of the word. In such cases, they will not be segmented. For instance, the form ከ /kə/ “from” is a preposition. However, it can be part of a content word as in ከበደ /kəbbədə/ “became heavy” or “a personal name”. In such cases, the from ከ/kə/ should not be segemented. This indicates that we cannot apply a certain rule or write a regular expression to automatically segment clitics. Separating clitics, we can say that, requires knoweldge of exisiting words in the laguage.
In addition, the form of the clitics can be changed due to phonological process. This change can also be observed in the orthography. For instance, the prepostion ለ lə “to/for” is attached to a content word that begines with the vowel like ኣሸናፊነት /aʃʃənnafinnət/ “winning”, the form of the preposition will be changed.As a result, the form becomes ላሸናፊነት /laʃʃənnafinnət/ “ for a winning”. If we consider all the variations a clitic may have, it will be problematic to handle all variations. Thus, the guideline suggests to restore to the orginal form in the segmentation. Accordingly, ላሸናፊነት will be segmented into ለ_ኣሸናፊነት.
The manual segmentation is important to solve some ambiguities observed in the arthography. For instance, some verbs which are relativized can be in active or passive form. This ambiguity occurs because when the relative marker is attached to a passive verb, the passive marker ‘ተ‘ /tə/ will get assimilated to the consonant that begins the word. Thus, we cannot tell whether a relatived verb is an active or passive from the orthography unless we consider the context or the pronouciation. For instance, the word, የሚበላ can read as /jəmmibəlla/ “the one who is eating” as an active form or read as /jəmmibbəlla/ “the one who is being eaten” as a passive form. In the morpholigical anlysis of HornMopho, this is handled by giving both analysis. The following figure shows the analysis of HornMorpho.
We notice from figure 1 above that the expression የሚበላ can have two citation forms በላ /bəlla/ “eat” for active and ተበላ /təbəlla/ “ being eaten” for passive. The possible interpritation of the expression is given under “grammar” part of the analysis. In our manual anotation, since the segmentation is done for a give sentence which is the context, this expression will be segmented as either as የ_እም_ይ_በል_ኣ or as የ_እም_ይ_ተ_በል_ኣ depending on the contex. The manual segmentation is therefore important to the development of a morphological analyser with a disambugation module for the future.
This stage is the basic and fundamental step where the data is given to three annotators who are linguists and have better understanding of the language for the manual segmentation. Interannotators agreement will be checked. After we have reached above 95% inter-annotator agreement, we will assign them a separate data for clitic segmentation. The result of this level will be a corpus of clitics separated from lexical words. It will help us to develop a tokenizer which is a basic tool for the language.
After the segmentation, the corpus will be annotated for POS tag and morphological features. We have compiled 56 POS tag sets based on morphosyntactic properties words. Table 1 summarizes the POS tag sets.
The above tags will be revised based on the feedback we will get from the annotators. The list is subjected for modification. In addition, we have listed possible morphological features which words in Amharic can represent. Table 2 lists the morphological features that can be annotated in the treebank.
Basic Categories Inflection
Gender Number
Case
Type masculine feminine common singular plural dual collective nominative
Tags masc fem com sing plur dual coll nom
Nominal Definite verb form
Tense Aspect Voice Person Negative Agreement
Gender Number accusative genitive definite infinitive gerund indicative jussive question negative past present future imperfect perfective prospective progressive active passive reciprocal causative first second third positive /affirmative negative subject object dative applicative masculine feminine singular acc gen def inf ger ind jus que neg past pres fut imp perf pro prog act pass rcp cau 1 2 3 pos neg subj obj dat app masc fem sing plu
Therefore, the segmented sentences will be annotated for both POS tag and morphological features. This could be done in a semi-automatic way. That means, some of the data like 100 sentences will be manually annotated and then the machine learns the tag and morphological features out of these seed sentences. Then other set of sample sentences will be given to the Verb system to annotate for both type of information. The result will be manually checked and corrected by the annotators. The system again learns from the corrections. In other words, it will be done in iterative ways i.e. manually annotation, training, manual correction, retraining, and annotation (Judge et al., 2006). The result will be used to develop an automatic POS tagger and morphological analyzer.
Finally, we plan to annotate the sentences
with grammatical relations using a universal
dependency framework
Category Relation Description
Nominal
Dependency verbal Dependency phrases and clauses adj pred app spec cpnd subj pass obj impv pro gen link conj sub cond rslt conc temp loc caus adjective possessive construction predicate apposition specification subject of a verb passive subject object of a verb imperative prohibition prepositional phrase PP attachment coordinating conjunction subordinate clause condition result concessive temporal local causal
Using the above relations, sentences will be semi-automatically annotated for syntactic relations following the same procedures we follow in the above annotations. Consequently, put all the activities together we will have a treebank annotated for all the three information: In figure 2 we observe a dependency tree for an Amharic sentence. We many notice that complex expressions are embraced by a bracket and their segmentation are indicated following the bracket. When we want to retrieve the orthography we can consider the expression in bracket and when we want to represent the syntactic roles played by the clitics we can consider the segmentation. Furthermore, the morphological features are also indicated together with their tags if a given token has morphological features. We have produced the above kind of representation for some couple of sentences. However, the manual segmentation of the remaining sentences is in progress.
We have described an ongoing project that aims at developing a treebank for Amharic. As the language is less resourced and morphologicallyrich, we suggest the annotation of the treebank to have three tiers: POS tag, morphological features and syntactic relations. Before the annotation is done, orthographic word needs to be segmented if it has clitics. We suggested that the minimal unit for our analysis should be syntactic words, i.e. both content word and functional words.
This project is partially funded by the Director for Research of AAU under Adaptive ProblemSolving Research grant and NORHED fund under Linguistic Capacity Building-Tools for inclusive development of Ethiopia, (http://www.hf.uio.no/iln/english/research/projec ts/linguistic-capacity-building-tools-for-theinclu/). We want to thank both institutions for their support. We also would to extend our appreciations to the three anonymous reviewers for their feedback that greatly improved the article.