Icon11 Marathi Kridant
Icon11 Marathi Kridant
Ganesh Bhosale, Subodh Kembhavi, Archana Amberkar, Supriya Mhatre, Lata Popale,
Pushpak Bhattacharyya
Department of Computer Science and Engineering,
IIT Bombay
1
http://sampark.iiit.ac.in
2.3 Kridanta forms in other Agglutinative
Krudanta Example Aspect Languages
Type padutunnappudoo is a participial form (here, an
adverb derived from a verb), which is similar to the
Ne वाचयासाठ Perfective Marathi Krudanta forms derived by affixing the
derivational morpheme ‘taanaa’ to the verbal base.
{vaachaNyasaThI}
Turkish:
hazirlanmis plan
{for reading}
prepare-past plan
The plan which has been prepared
La वाचयावर Perfective Here, hazirlanmis is the participial form derived
from a verb.
{vaachalyavar}
{after reading} 3 Challenges in Marathi Morphological
Processing
Tana वाचताना Durative
Marathi is a morphologically rich language with a
{vaachataanaa} high level of agglutination. Inflections can get at-
tached at various positions. We mention some of
{while reading} these challenges.
Lela वाचलेले पुःतक Perfective 3.1 High level of inflection and agglutinative
morphemes
{vaachalele pus- There are many examples of Krudanta forms with
tak} more than two suffixes attached to the root after
inflection. Stacking of affixes is common. Consid-
{book which
is/was read} er in this example, मारणाढयानीदे खील {maarNaa-
ryaneedeKeel} {the killers also} {मार + णारा + नी
Nara वाचणारा Stative
+ दे खील} the root is मार {maar} {kill} as a verb
{vaachaNaaraa} attached with three suffixes णाढया, नी and दे खील
respectively. Here both the suffixes णाढया and नी
{the one who are followed by another morpheme. Morphological
reads} analysis with affix stripping (Eryiğit and Adalı,
2004) fails to process such a Krudanta form.
Va वाचावे Inceptive
3.2 Complex Forms of Krudanta
{vaachave}
As Krudanta are made up of verbal roots, they
sometimes encode ‘aspect’ feature. Hence it is very
{must read}
difficult for the traditional affix stripping approach
without a proper rule to process such complex
Oon वाचून {vaachun} Completive forms. But, maintaining many such rules for each
possible grammatical form is not feasible. For ex-
{having read} ample in केलेयानेस# ु ा {kelelyanesudhaa} {the
person who did as well}, कर is the root and is af-
Table 1: Krudanta Types in Marathi
fixed with the Krudanta morpheme ‘lela’, which
denotes ‘perfective’ aspect. The emergent form is 4 Morphological Processing of Marathi
then affixed with two more suffixes ने and सु#ा keeping krudanta in focus
respectively. We have developed a Morphological Analyzer for
Marathi using a finite state machine approach. For
3.3 Ambiguity in Finite and Non-Finite that purpose, we are using SFST (Stuttgart Finite
forms State Transducer) tool 2which is used for the im-
plementation of computational morphology. This
In Marathi, there are pairs of morphemes that have approach builds on the work of Bapat et al. (2010).
similar phonological and orthographical shape and There are three important components in the archi-
that bring about similar changes in the phonologi- tecture of the morphological analyzer as shown in
cal and orthographical shape of the base word they Figure 2.
are affixed to. As a result, the final forms are pho-
nologically and orthographically similar, but have Inflector: The inflector takes as input lexicon
two different meanings. For example, there can be and suffix replacement rules (SRR). The format of
two Marathi morphemes, which are represented by an SRR is (Paradigm, Suffix, Ultimate Insertion,
the letter ‘त’ {t}, one of them being an inflectional Ultimate Deletion, Penultimate Insertion, Penulti-
morpheme denoting habitual past and the other a mate Deletion, and morph features). Consider, for
derivational morpheme denoting progressive as- example, मुले {mule} {boys} where to extract the
pect. Thus, when attached to a verbal root like root मूल we used the SRR rule below <मूल>< , ◌े,
‘$फर’ {fir} meaning ‘to wander’, these two suffix-
◌ू, ◌ु><noun,n,pl,,,,,d>. This means the following:
es produce two similar forms- $फरत {phirat}, one
as a first step, we remove ◌े from ultimate position
of which means ‘they used to wander’, while the
other means ‘wandering’. In such cases, the Mor- of the word and then we replace ◌ू for ◌ु in the
phological Analyzer should be able to produce penultimate position to get the root word मूल. The
both the analyses. After that, we need to disambi-
Inflector creates a flat file where it stores all the
guate these analyses, by using statistical HMM
inflected forms of words present in the lexicon
along with distinguished morph features (Hakkani,
along with the associated grammatical attributes
2002).
and labels. We call this flat file as a dictionary of
inflected forms.
3.4 Two consecutive Krudantas
Morphological Recognizer: The Morphological
Consider, for example, the individual Krudantas
Recognizer breaks a given word into morpheme
चालून {chalun} and येताना {yetaanaa}, which segments. Morphotactics are modeled using the
mean ‘after walking’ and ‘while coming’ respec- FSM which carried out the segmentation with the
tively. However, when they are used together help of generated ‘dictionary of inflected forms’.
चालून येताना ‘chalun yetaanaa’, the meaning is
Morphological Parser: These segments along
{while coming walking}. A sequence of Krudantas
with their labels are input to the Morphological
may have a meaning different from the constituent
Parser which uses the ‘dictionary of inflected
Krudantas. In case of two such consecutive Kru-
forms’ to extract the root and also gives the mor-
dantas, we get the chunk type as VGNF (Verb
phosyntactic properties of each morpheme.
Group Non-Finite). In the above example, the in-
dividual processing of each Krudanta gives an ac-
curate translation, but for two consecutive
Krudantas the translation process fails because of
the incorrect Vibhakti (suffix generation after head
computation in the chunk).
2
http://www.ims.uni-
stuttgart.de/projekte/gramotron/SOFTWARE/SFST.html
Figure 2: Architecture of the Marathi Morphologi-
cal Analyzer Figure 4: FSM rules for Krudanta
We show an example of Krudanta in figure 3. No- To process this Krudanta word the following FSM
tice how the large number of morphemes get iso- rule is used. [VERBS -> Nara_DF -> SSY]. In
lated through the above mentioned stages. processing of गाणारे दे'खल {gaanaaredekhil}
{singers also} (which is the plural form of the
गाणारादे 'खल) following FSM rule is used.
[VERBS -> Nara_OF -> SSY] where VERBS file
contains verb roots like गा, Nara_DF file contains
suffix णारा, Nara_OF file contains suffix णारे with
other inflected forms of suffix ‘णारा’, and SSY file
contains suffixes like दे 'खल.