| Computational Paninian Grammar framework (PG) has been successfully
applied to modern Indian languages earlier, using which anusaaraka
machine translation system has been built (Narayana, 1994).
In this paper, we show that
PG can also be applied to English resulting in an elegant
computational grammar.
First, we generalize the notion of vibhakti to include position
of the word in a sentence along with its case and associated
preposition, if any. This allows us to use the familiar PG
notions of karaka chart, karaka chart transformation, and
sharing rules (Bharati et al., 1995) to account for the
English actives and passives, lexical
control, infinitives, etc. A transformation of the karaka chart
and the vibhakti therein, very naturally accounts for what is called
movement.
Second, we introduce a new vibhakti called TOPIC position
(which corresponds to the first position in a clause) and a new
operation called join for connecting a relative clause to its
head. These two together handle long
distance dependency in relative clauses and wh-questions,
raising, tough-movement, pied-piping, etc. The karakas with TOPIC
vibhakti appear at the beginning of the clause.
This paper establishes that PG is more general than hitherto
considered, and can be used to explain not just free word order
languages but also positional languages. Further reseearch
is continuing on this and related aspects.
|
PG for Indian Languages - A Review
(TOP)
The Paninian framework considers as central
to the study of language.
When a writer (or a speaker) uses language to convey
some information to the reader (or the hearer),
he codes the information in the
language string
footnote{By string we mean any of: word, phrase,
sentence, paragraph, etc.}. Similarly, when a reader (or a hearer)
receives a language string, he extracts the information coded in it.
The computational Paninian Grammar framework (PG) is primarily
concerned with: how the information is coded and how it can
be extracted.
Two levels of representation can be readily seen in langauge use:
One, the actual language string (or sentence), two, what the speaker
has in his mind. The latter can also be called as the meaning.
Paninian framework has two other important levels: karaka level
and vibhakti level (Figure 1).
--- semantic level (what the speaker
| has in mind)
|
--- karaka level
|
|
--- vibhakti level
|
|
--- surface level (written sentence)
The surface level is the uttered
or the written sentence.
The vibhakti level is
the level at which there are local word groups based on case endings,
preposition or postposition markers.
footnote{For positional languages
such as English, it would also
include position or word order information.}
A noun group is a unit containing a noun (or a pronoun,
proper name, etc.), its vibhakti and possibly some adjectives.
Vibhakti for verbs includes the verb form and the auxiliary verbs.
At the karaka level, we have karaka relations and verb-verb relations etc.
Karaka relations are syntactico-semantic
(or semantico-syntactic)
relations between the verbs and other related constituents
(typically nouns) in a
sentence. They capture a certain level of semantics which is
close to thematic relations but different from it. But this is
the level of semantics that is important syntactically and is
reflected in the surface form of the sentence(s).
The vibhakti level abstracts away from many minor (including
orthographic and idiosyncratic) differences among languages. The
topmost level relates to what the speaker has in his mind. This
may be considered to be the ultimate meaning level. Between this
level and vibhakti level is the karaka level. It includes karaka
relations and a few additional relations such as taadaarthya (or purpose).
One can imagine several levels between the karaka and the ultimate
level, each containing more semantic information. Thus, karaka level
is one in a series of levels, but one which has relationship to
semantics on the one hand and syntax on the other.
As mentioned earlier, vibhakti for verbs
can be defined similar to that for the
nouns. A head verb may be followed by auxiliary verbs (which may
remain as separate words or may combine with the head verb). Such
information consisting of the verb ending, the auxiliary verbs, etc.
is collectively called vibhakti for the verb. The
vibhakti for a verb gives information about tense, aspect and
modality (TAM), and is, therefore, also called the TAM label. TAM
labels are purely syntactic determined from the verb form and the
auxiliary verbs.
The grammar gives the mapping between the levels. For example,
PG specifies a mapping between the karaka level and the vibhakti
level, and the vibhakti level and the surface form.
It has been shown earlier (Bharati et al., 1995) that the
Paninian grammar is particularly suited
to free word order
languages. It gives a mapping between
karaka relations and vibhakti, and uses position
information only
secondarily. As the Indian languages have (relatively) free word
order and vibhakti, they are eminently suited to be described by
Paninian Grammar.
Karaka-Vibhakti Mapping
>
(TOP)
The most important insight regarding the karaka-vibhakti
mapping is that it depends on the verb and its
tense aspect modality (TAM) label. The
mapping is represented by two structures: default karaka chart
and karaka chart transformation\index{karaka chart transformation}.
The default karaka chart for a verb or a class of verbs gives the
mapping for the TAM label known as basic.
It specifies the vibhakti permitted for the applicable karaka
relations
footnote{What karaka relations are permissible for a verb,
obviously depends on the particular verb. Not all verbs will
take all possible karaka relations.}
for the nouns etc. when the verb has the basic TAM label.
In other words, when the verb in a sentence has the basic TAM label,
then for each of the nouns in the sentence
with a given karaka relation with the verb,
their vibhakti is given by the the karaka chart.
The basic TAM label chosen for Hindi is {taa\_hei} and
roughly corresponds to present indefinite
tense. Any other TAM label could have been chosen as basic
without any problem.
(The TAM labels are purely syntactic in nature and can be
determined by looking at the verb form and the associated
auxiliary verbs, etc.)
For TAM labels other than basic, there are karaka chart
transformation rules.
Thus, for a given verb with some TAM label in a sentence,
appropriate
karaka-vibhakti mapping can be obtained using its default
karaka chart and the transformation rule depending on its
TAM label.
The default karaka chart
ex{default karaka chart, see under karaka chart}
for three of the karakas is given in Figure \ref{def-kc}.
{tabular}{lcc} %{p{20mm}p{30mm}p{30mm}}
line
{ Karaka} & { Vibhakti} & { Presence}\\
line
Karta & $hi{}$ & mandatory\\
Karma & ko or $hi{}$ & mandatory \\
Karana & se or dvaaraa & optional\\
This explains the vibhaktis in sentences A.1 to A.2 as
described below.
Take for example, the following modifier-modified structure:
piiTa (beat) [present]
k1 k2
Raama Mohana
Using the karaka chart, we get the following vibhaktis:
raama [$\phi$], mohana [ko], piiTa [taa\_hei]
which yield the following sentences:
A.1 raama mohana ko piiTataa hei.
Ram Mohan -ko beat is
(Ram beats Mohan.)
A.2 mohana ko raama piiTataa hei.
Mohan -ko Ram beat is
(Ram beats Mohan.)
Note that no order is specified by the grammar among the nouns.
It is important to reemphasize that the transformation depends on TAM
label which is purely syntactic, and not on tense, aspect
and modality which are semantic. The TAM label can be determined for
Hindi and other Indian languages by syntactic forms of the verb
and its auxiliaries without the need to refer to any semantic
aspects. The specification for obtaining TAM labels is given
by a finite state machine as described in
Bharati et al. (1995; Chap. 4).
The Paninian theory outlined above (i.e., karaka charts and karaka chart
transformations) can be used to generate (or analyze) the sentences
given above as we have seen. However, there are additional constraints
that would disallow the following sentences to be generated, for
example\footnote{A `*' before a sentence indicates that it is not
a good sentence.}:
B.1 *ladake ne raama ne laDakii ko kitaaba dii.
boy -ne Ram -ne girl -ko book gave
(*The Ram the boy gave a book to the girl.)
B.2 *laDake ne laDakii ko kitaaba phoola dii.
boy -ne girl -ko book flower gave
(*The boy gave a book a flower to the girl.)
The constraints are given below:
-
Each mandatory karaka\ in the karaka chart for each
verb group, is expressed {\em exactly once}. (In other words,
a given mandatory karaka generates only one noun group
with the specified vibhakti in its karaka chart unlike in
B.1.)
- Each optional karaka in the karaka chart for each verb group,
is expressed {\em at most once}.
- Each source word group satisfies some karaka relation with some
verb (or some other relation). In other words, there should
no unconnected source word group in a sentence, otherwise,
the sentence becomes bad as in B.2.
Karaka charts are based on the idea of aakaankshaa and yogyataa.
A karaka chart for a verb expresses its
aakaankshaas or demands, and specifies the vibhaktis that must be used
(i.e., yogyataa) with word groups that satisfy the demands.
The same ideas can be used to handle noun-adjectives,
noun-noun relations, verb-verb relations etc, each of which
can be viewed as a demand-satsifaction pair.
|
{Complex Sentences}
(TOP)
Let us now consider the generation of sentences with more than one
verb group. We begin with an example. Suppose the speaker (or writer)
wants to express the fact that
raama ne mohana ko phala khaakara bulaayaa.
Ram erg. Mohan dat fruit having-eaten called
(Ram called Mohan having eaten the fruit.)
This can be expressed by the structure at the karaka level
shown in Figure
bulaa (call) [past]
karta precede
karma
raama mohana khaa (eat)
.
karta . karma
.
raama phala (fruit)
Modifier-modified relations for a complex sentence
(shared karakas shown by dotted lines)}
One way of realizing it in a sentence is to choose TAM label
for
{khaa} (eat). This label specifies the temporal
precedence relation with its parent
node {bulaa} (call).
\footnote{However, it comes packaged, so to
say, with the karaka sharing constraint
that karta of {khaa} must be the same as that of its parent. Since
this constraint is satisfied, the choice of TAM as {kara} is
acceptable. More on this later.}
The vibhaktis for the nouns
can now be generated using the karaka chart (i.e., the
default karaka chart together with karaka chart transformation rule
for {kara} shown in Figure \ref{tam-transform}). In this example,
the transformed karaka chart shows that karta of {khaa}
is not expressed. The vibhaktis for the nouns in Figure \ref{m-m-g1}
are as follows:
raama [$\phi$], mohana [ko], phala [$\phi$], khaa [kara], bulaa [taa\_hei]
The generated sentence reads as:
raama mohana ko phala khaakara bulaataa hei.
The karaka chart transformation rules are
given in Figure tam-transform}.
-
Karta must not be expressed.\\
-
Karma is optional.\\
-
naa & Karta and karma are optional.\\
-
taa\_huaa & Karta and karma are optional.\\
\caption{Transformation rules for complex sentences}
The karaka sharing rule is given in Rule S1.
Karta of intermediate verb with TAM label {kara} is the same as the
karta of the verb modified by the intermediate verb.
|
{Constraint Based Parsing}
(TOP)
The Paninian theory outlined above can be used for building
a parser.
First stage of the parser takes care of morphology.
For each word
in the input sentence, a dictionary or a lexicon is looked up,
and the associated grammatical information is retrieved.
In the next stage, local word grouping takes place, in which based on
local information certain words are grouped together yielding
noun groups and verb groups.
These are the word groups at the vibhakti
level (i.e., typically each word group is a noun or verb with
its vibhakti, TAM label, etc.).
These involve grouping post-positional
markers with nouns, auxiliary verbs with main verbs etc.
Rules for local word grouping are given by finite state machines.
Finally, the karaka relations among the elements are
identified in the last stage called the {\em core parser}
(Bharati et al., 1995; Chap. 6).
The task of the core parser to identify karaka relations.
It requires karaka charts and transformation rules.
For a given sentence after the word groups have been formed,
each of the noun groups is tested against each row (called
{\em karaka restriction}) in each
karaka chart for each of the verb groups
(provided the noun group is to the left of the verb group
whose karaka chart is being tested).
When testing a noun group
against a karaka restriction of a verb group, vibhakti information
is checked, and if found satisfactory, the noun
group becomes a candidate for the karaka of the verb group.
The above can be shown in the form of a constraint graph.
Nodes of the
graph are the word groups and there is an arc labeled by a
karaka from a verb group to a
noun group, if the noun group satisfies
the karaka restriction in the karaka chart of the verb group. (There
is an arc from one verb group to another, if the karaka chart
of the former shows that it takes a sentential or verbal karaka.)
The verb groups are called demand groups as they make demands about
their karakas, and the noun groups are called source groups because
they satisfy demands.
As an example, consider a sentence containing the verb khaa (eat):
E.1 baccaa haatha se kelaa khaataa hei.
child hand -se banana eats
(The child eats the banana with his hand.)
Its word groups are marked and
khaa (eat) has the same karaka chart as in Figure \ref{def-kc}.
Its constraint graph\index{constraint graph} is shown in Figure
\ref{tam-transform}.
karma karta
baccaa haatha se kelaa khaataa hei
karana karma
karta
{Constraint graph for sentence E.1}
A parse is a sub-graph of the constraint graph\index{constraint graph} containing all the nodes
of the constraint graph and satisfying the following conditions:
-
For each of the mandatory karakas in a karaka chart for each demand
group, there should be {\em exactly one}
outgoing edge labelled by the karaka
from the demand group.
-
For each of the desirable or optional
karakas in a karaka chart for each demand
group, there should be {\em at most one} outgoing edge labelled by the karaka
from the demand group.
-
There should be {\em exactly one} incoming arc into each source group.
Efficient methods based on bipartite graph matching are known
for finding solution graphs.
If several sub-graphs of a constraint graph satisfy
the above conditions, it means that there are multiple parses and
the sentence is ambiguous. If no sub-graph satisfies the above
constraints, the sentence does not have a parse, and is probably
ill-formed.
With the constraints (C1, C2 and C3 ) specified above, the parsing
problem reduces to bipartite graph matching\index{bipartite graph matching}\footnote{We
are indebted to Somnath Biswas for suggesting the reduction.}
and assignment\index{assignment}
problems (Bharati et al. (1995; Chap. 6)). These have efficient solutions even in the worst case.
{Paninian Framework Applied to English}
>
(TOP)
|
We have already discussed in the last section that PG gives
primacy to information and its coding. Thus, the karaka-vibhakti
mapping for a language specifies how the karaka information
can be coded in the vibhakti in that language. The vibhakti
for a free word order language, say an Indian language,
consists of case and post-positional markers. It is natural,
therefore, to include position information as part of the vibhakti in a
positional language such as English.
We introduce the notion of generalized vibhakti. The
generalized vibhakti includes relative position of a constituent
besides case endings, post-positions and prepositions.
As was just mentioned, in a
(relatively) fixed word order language like English or French,
the relative position plays an important role, whereas in a
(relatively) free word order language such as an Indian language
or Japanese, it is the other markers that are more important.
Thus, a language may choose to have any of these devices or a
combination thereof for encoding semantic information.
There are two important vibhaktis for English based on word
order:
- [Subject:] This is the pre-verbal position, that is the position
to the immediate left of a verb group.
- [Object:] This is the post-verbal position, that is, the
position to the immediate right of a verb group.
These do not have any other implications regarding tree
structure. Thus, they should not be confused with subject
and object
terms in the generative enterprise or some other formalism in
linguistics where these terms are defined with respect to the trees.
Take for example the modifier-modified structure in Figure \ref{mm-eat}.
mmeat
eat (past)
k1 k2
Ram fruits
{An example modifier-modified structure}
The karaka chart for eat is given in Figure \ref{kc-eat}
where `acc' means that the word (which is karma) must be in
accusative case. For karana no position is specified in the
vibhakti, so that it can occur anywhere as long as vibhakti
of other karakas are not affected. (If we had wanted to
place a constraint that it must occur to the right of object, we
could have written its vibhakti as ` > obj'.)
kc-eat
{Karaka} & {Vibhakti} & {Presence}\\
Karta (k1) & subj & mandatory\\
Karma (k2) & obj + acc(accusative) & mandatory \\
Karana (k3) & prep.: with & optional\\
{Default karaka chart for English verb {eat}}
The following vibhakti level representation is produced, when we
try to generate it for the modifier-modifier structure using the
karaka chart for eat.
Ram [subj] fruits [obj] ate
No order is explicitly shown in the above. Finally, when we
generate the sentence we get:
Ram ate fruits.
where Ram and fruits have taken their (positional) vibhakti.
(Determiners and adjectives are handled at the level of local
word grouping. They are ignored here for simplicity, and it will
be assumed that they can be produced when needed by local word
grouping.)
Let us introduce a more pictorial notation for karaka charts
containing generalized vibhakti. In the new notation, the karaka
chart for {eat} given earlier would appear as:
*
______ eat ______[acc] with_______
k1 k2 Opt k3
An underline indicates that a constituent is to appear
there. The placement of the line shows its relative position with
respect to its neighbour unless a `*' is marked above it which
indicates that there is no constraint on word order for that
constituent. The
prepositions and case are shown along with the underline.
The default case is
nominative, if no case is shown. The relation that the constituent
will have with the verb {eat} is written below the line.
Optional karaka relations are marked by {opt} and this indicates
that it is not mandatory that the constituent with the given
karaka relation and vibhakti be present.
For the modifier-modified structure in Figure \ref{m-m-ex2},
{m-m-ex2}
eat [past]
karta karma karana
Ram fruits his hands
{Another example of modifier-modified structure}
the following sentences can get generated using the above karaka
chart:
Ram ate the fruits with his hands.\\
With his hands, Ram ate the fruits.
Note that {with his hands} cannot come between {Ram} and {ate} or
between {ate} and {fruit} because then the adjacency relation
between karta and the verb and karma and the verb (specified by
subj. and obj. vibhakti, respectively) would be violated.
{kct-passv}
{TAM label} & {Transformation}\\
Passive & k2 subj.;\\
& k1 opt, prep: by; *\\
{Karaka chart transformation for passive TAM}
|
For passives, we have the following karaka chart:
* *
______ eat[passive] by_____[acc] with_____
k2 opt k1 opt k3
The karaka chart transformation rule given in Figure \ref{kct-passv}
generates the
above chart from the basic or default chart.
Pictorially the transformation rule can also be shown as:
passivization *
____ eat ______[acc] ===========>> _____ eat[passive] by_____[acc]
k1 k2 k2 Opt k1
The passive voice sentence gets generated when the verb in the
modifier-modified tree is marked as passive semantically (meaning
thereby that the role of the karta is de-emphasized).
As an example, we have a modifier-modified structure and the
generated sentence as shown in Figure \ref{m-m-passv}.
{m-m-passv}
bring [past,+passive]
k1 k2
she books
The books were brought by her.
{Modifier-modified structure for a passive}
TAMs in English are the counterpart of those in Indian languages:
They specify tense, aspect and modality (compositely) associated
with the verb, and are given by the verbal ending and the
auxiliary. Past-passive TAM, for example, means that the auxiliary
verb {be} is with the appropriate tense ({were}) followed by
the participle form of the main verb ({brought}). As defined
earlier, a TAM label of a
verb can also be called as the vibhakti of the verb.
{mm-lex-ctrl}
Promise
k1 k2 k2-v
Ram Mohan go
k1 k2
Ram(i) party
{Modifier-modified structure with shared karta in lexical control}
|
We will now look at some verbs that take another verb as
their argument, and share some karakas with it. In the following
sentence, for example:
Ram promised Mohan to go to the party.
{go} is the karma-vishaya (abbreviated as k2-v)
\footnote{Karma-vishaya
is a karma which gives the topic or subject related to
the main verb.}
of { promise}, and { Ram} is the karta of { promise}
as well as { go} in Figure {mm-lex-ctrl}.
Consider another sentence where the karta of the modifier verb is
the same as the karma of the modified verb:
Ram persuaded Mohan to go to the party.
This is generated from the structure shown in Figure {mm-persuade}.
{mm-persuade}
persuade[past]
k1 k2 k2-v
Ram Mohan go
k1 k2
Mohan party
{Modifier modified structure for { persuade}}
The karaka charts for { promise} and { persuade}
are given in Figure {kc-promise}.
{kc-promise}
_______ promise _______[acc] ___________[inf]
k1 k2 k2-v (k-control sharing)
_______ persuade _______[acc] ___________[inf]
k1 k2 k2-v (k-control sharing)
{Karaka charts for {promise} and {persuade}}
They are the similar to karaka charts shown for other verbs
except that they differ in the karaka control or k-control.
{ Promise} is said to be a { karta-control} verb whereas
{ persuade} is a karma-control verb. The former has its k-control
equal to karta, and the latter has k-control as karma.
The modified verb, shares its karta with a karaka of the
modified verb depending on its k-control. It is important to note
that the control is independent of the voice (or TAM) of the
verb. For example, { Mohan} is still the karta of { go} in the
following sentence:
Mohan was persuaded by Ram to go to the party.
In the standard linguistics analysis, one talks of {persuade}
as an object-control verb. But this is not quite right because
when passivized, it acts as a subject-control verb. Alternatively,
one will have to say that {persuade} is on object-control
verb when active and a subject-control verb when passive.
PG analysis is simpler and more elegant.
|
In this section, we look at complex sentences with infinitives
indicating purpose relations. The modifier verb is an infinitive (TAM
equal to infinitive) and specifies a purpose for the modified
verb. For example, {cut} is the purpose for the main verb in the
following sentence:
Mohan brought a knife to cut fruit.
There is a relation at the karaka level called the taadarthya or purpose.
Using it the karaka charts can indicate a purpose relation between two verbs (usually the main verb and another verb):
* * *
_____ bring ______[acc] with_____[acc] in_____[acc] to___[v,inf]
k1 k2 opt k3 opt k7 opt purpose
The karaka chart specifies that the
argument verb must take the infinitive TAM.
If we begin with the default karaka chart of {cut} verb with its TAM
equal to simple present:
______ cut ________[acc]
k1 k2
then, for the infinitive form of {cut}, the karaka chart can be
obtained by karaka chart transformation rule for the infinitive
TAM shown in Figure \ref{kct-inf}. This
kct-inf}
TAM & Transformation \\
infinitive & karta must not be expressed
(sharing rule T1),\\
& or karta has prep.: {for}, [acc] \\
{Karaka chart transformation for infinitive TAM} yields the karaka chart:
For______[acc] cut[inf] _______
opt k1 opt k2
It also places a constraint that if karta is not expressed, it
must be the same as the karta of the main verb.\footnote{In the
example sentence, rule T1 applies. With {purpose} relation,
however, karta could be shared with other karakas also. For
example in the following sentence:
Ram gave him a toy to play.
karta of {play} is possibly shared with sampradana of {give},
namely {him}. But
not so for the modifier verb {win} in the following sentence:
Ram gave him a toy to win his trust.
More generally this can be stated as the sharing rule T1.
Sharing rule T1: If the karta of a modifier verb is not
expressed, it must be the same as the karta of the modified verb.
With the above karaka charts we can generate a sentence for
the modifier-modified structure in Figure {mm-share}.
{mm-share}
bring (past)
k1 k2 purpose
Mohan knife cut
i
k1 k2
Mohan fruit
i
{Modifier-modified tree with a shared karta}
The generated vibhakti structure is:
Mohan [subj of bring] knife [obj (bring)] fruit [obj (cut)] brought cut[inf]
and finally the sentence:
Mohan brought a knife to cut fruit.
or
To cut fruit, Mohan brought a knife.
If the kartas for the two verbs were different, we could not drop
the karta of the {cut} verb. For example, for the
structure in Figure \ref{mm-noshare},
{mm-noshare}
bring (past)
k1 k2 purpose
Mohan knife cut
k1 k2
Ram fruit
{Modifier-modified structure without shared karta}
the karaka charts would permit the generation of the following
sentence:
Mohan brought a knife for Ram to cut fruit.
but not one in which karta of cut is not explicitly expressed:
* Mohan brought a knife to cut fruit.} (Bad sentence for the above tree.)
|
Relative clauses allow us to refer to any participant in an
action, and relate to another action (say, the main action) etc.
In English, it requires putting a wh-word for the participant to
be referred to and moving it to the beginning of the clause. For
example, in the sentence below, {the woman} participates in two
actions: {work} and {donate}. It occurs with
the main verb {works} and is shown to be a participant in the
action {donate} using a wh-word ({who}):
(S1) The woman [who donated the necklace to the relief fund] works in a quarry.
Similarly, here is another example where the {necklace} is the entity referred to by the relative clause involving the action {donate}:
(S2) The necklace [which the woman donated to the relief fund] was bought by her husband.
Note that {who} in the sentence S1
(or {which} in the sentence S2),
is at the beginning of the clause, rather than, its
normal position with respect to its verb {donate}.
The modifier-modified structure for the sentence S1 is
given in Figure \ref{rel-cl-donate}, where a dot marks
{rel-cl-donate}
.
work donate
k2 k1 k1 k2 k4
quarry woman necklace relief fund
{Modifier-modified tree for relative clause}
the main action ({work}). Equivalently, the above structure can
be drawn as a tree using $k1^{-1}$ relation
as shown in Figure \ref{rel-cl-donate2}
(where a dot is not
necessary as the tree has a single root node
denoting the main action).
{rel-cl-donate2}
work
k2 k1
quarry woman
-1
k1
donate
k2 k4
necklace relief fund
{An alternate modifier-modified structure for relative clause}
\end{figure}
When a verb takes a sentence as its argument in a
relative clause, the wh-element still moves to the front, but the
rest of the order is not disturbed in the clause. For example,
when {say} is introduced, we have:
(S3) The woman [who Ram said donated the necklace to the relief fund] works in a quarry.
The modifier-modified structure (M3) in Figure \ref{rel-cl-donate3} corresponds to the sentence S3.
{rel-cl-donate3}
say
. k2 k1
work
donate Ram
k2 k1 k1 k2 k4
quarry woman necklace relief fund
{Modifier-modified tree M3 for an embedded sentence in relative clause}
Note that {the woman} is still a participant in the two actions:
{work} and {donate}; except now, {donate} is itself an argument of
another action {say}. The positions of the arguments of {say}
remain undisturbed, in the sentence S3, only the positions of
arguments of donate are altered. In particular, {who},
the referred argument of {donate},
occurs in the first place in the clause.\footnote {If there are
several ancestor nodes of {donate}, they would all maintain
normal positions of their arguments, and the referred argument of
{donate} would be further away from its normal position.
Therefore, this is called long distance movement or long distance
dependency.}
We now present a theory in the Paninian spirit for generating
the above kind of sentences with relative clauses. We look at the
theory in two steps: first for clauses with only one verb (simple
clauses) and then for non-simple clauses.
Whenever we have a modifier-modified structure which has more
than one root (as in Figures \ref{rel-cl-donate} and \ref{rel-cl-donate3}),
we break the structure up into trees as
shown in Figure \ref{rel-cl-two-trees}.
{rel-cl-two-trees}
.
work donate [rel.clause]
(M4)
k2 k1 k1 k2 k4
quarry woman who necklace relief fund
(i) (i)
{Breaking into two modifier-modified trees for a relative clause}
For a shared node (i.e., a node with two parents), its link to a
parent which is not under a root node marked by a dot, is broken
and instead a new node consisting of wh-element is created and
attached to the broken link. The parent ({donate}) is marked to
be of relative clause type.
The karaka chart for a verb marked to be a relative clause
type has a vibhakti called TOPIC for one of its karaka. This
vibhakti has the constraint that it occurs in the first position
of the clause obtained after substitution operation. Thus, the
following is the karaka chart for {donate} [rel. clause]:
[TOPIC] *
[+wh]____ donate[rel.clause] _____[acc] to________
k1 k2 k4
M4 trees generate the following representations at vibhakti
level:
(a) woman(i)[subj(works)] works quarry[prep:in]\\
(b) who(i)[TOPIC] donate[past] the necklace[obj (donate)] relief fund[prep:to]
Strings with ordered elements can be produced out of the above
representations. Finally, the two strings are merged together by
looking at coindexing across them and performing the join
operation. In the join operation, the coindexed item in the dependent
string immediately follows the other coindexed item and the
dependent string is inserted there. In the past example, we first
get the following strings:
(a) The woman(i) works in the quarry.\\
(b) who(i) donated the necklace to the relief fund
After the join operation we have:
The woman who donated the necklace to the relief fund
works in a quarry.
In a similar way if we generate a sentence for the structure M3,
then the representations at tree level, vibhakti level and
strings are as given in Figure \ref{repr-m4}.
{repr-m4}
work say [past]
k2 k1 k2 k1
quarry woman(i) donate [rel.- Ram
clause]
k1 k2 k4
+wh(i) necklace relief fund
(i) Trees
(a) woman(i)[subj] works quarry [obj+prep:in]\\
(b) Ram[subj(say)] say[past] who(i)[TOPIC] donate[past]
necklace[obj(donate)] relief fund[prep:to]\\
(ii) Vibhakti representation
(a) the woman(i) works in the quarry\\
(b) who(i) Ram said donated the necklace to the relief fund\\
(iii) Strings
{Three level representation for M4}
Note that the TOPIC vibhakti forces {who} to appear as the first
element in the string. Joining operation produces the correct
sentence.
|
Wh-questions are handled by a similar mechanism. An example
Karaka chart for a wh-question regarding the karta is given as
follows:
[TOPIC] *
[+wh]_____ donate[wh-Q] ________[acc] to______
k1 k2 k4
For the modifier-modified structure in Figure \ref{rel-cl-donate3},
this generates (ignoring the rules for placement of {did}):
Who did Ram say donated the necklace to the relief fund?
|
{Sentential Subjects}
(TOP)
Many verbs take a sentence as an argument. For example,
{tell, say, think, expect} take a sentence as their
karma. In active voice, the vibhakti assigned to karma is object
position. For example:
Ram {said/expected} that the boy will come to the party.
When such verbs are passivized, there is a tendency in English not to make these sentential arguments as subjects.]
\footnote {This might flow from basic principles relating to sentence processing by human mind which prefer that the head of the main sentence (i.e., the verb) not be postponed indefinitely.}
Passivization for such verbs produces a karaka chart in which the subject position is filled by a dummy or pleonastic element
It is said that the boy will come to the party.\\
It is expected by Ram that the boy will come to the party.
The karaka chart transformation rule for passivization of such verbs is shown in Figure \ref{kct-pass}.
{kct-pass}
Passiv.
____ say that______ ====> It say[passv] by_______ that____
k1 k2 k1 opt k2
{Karaka chart transformation for passives of {say} verbs}
The usual passivization rule shown in Figure \ref{kct-pass-usual}
is less preferred (or not acceptable).
{kct-pass-usual}
Passiv.
___ say that_____ ====> that_____ say[passv] by ______
k1 k2 k2 k1 opt
{Karaka chart transformation same as for usual passives}
and would produce a bed sentence such as:
*That the boy will come to the party is said.\\
?That the boy will come to the party is expected.
|
Raising is a peculiar phenomenon in English which occurs with
a small number of verbs that take sentential arguments.
(Such verbs are called raising verbs or exceptional case-marking
verbs.) We illustrate it by means of an example sentence involving
{expect}, and the modifier-modified structure given in Figure \ref{m-m-raising}.
{m-m-raising}
expect[passv]
k2
go[+infin]
k1 k2
Boy home
The boy is expected to go home.
{Modifier-modified structure involving {expect}}
Note that an argument of the embedded verb {come} has moved
to the subject position of {expect}. In Paninian terms, a karaka
of an embedded verb is given a vibhakti (position) that seems to
be defined with respect to the main verb.
The karaka charts for {expect} and {come} that handle
this phenomenon are shown in Figure \ref{kc-raising}.
{kc-raising}
v[inf,TOPIC]
expect[passv.] by_____ ___________
k1 opt k2
[TOPIC]
________ come[inf] to_______
k1 k2
{Karaka charts for {expect}, a raising verb}
Note that for raising to take place, the embedded verb must be in
the infinitive form.
{kc-no-raise}
it expect[passv.] for ______v[infin]
k2
_____ come[infin] to______
k1 k2
{Karaka charts for {expect} without raising}
The infinitive form can also occur without raising as shown in Figure
\ref{kc-no-raise}
to generate the marginally acceptable sentence:
It is expected for the boy to come to the party.
The TOPIC constraints block the generation of the following bad sentences:
*The boy it is expected for to come to the party.\\
*Is expected the boy to come to the party.
|
There are certain adjectives which display a phenomenon
similar to raising. For example, the following sentences:
It is tough to play this tune on the sitar.\\
This tune is tough to play on the sitar.
are generated for the modifier-modified structure
in Figure \ref{mm-tough}.
{mm-tough}
is-tough
k2
play
k1 k7
tune sitar
{Modifier-modified structure involving {tough}}
The karaka charts shown in Figure \ref{kc-tough} are similar to those used for handling raising. By simply marking certain karakas as having the vibhakti TOPIC in the karaka chart for the embedded verb {play}, the tough-movement is handled.
{kc-tough}
[TOPIC]
for______ play[inf] ________ on________
k1 opt k2 k7
[TOPIC]
for______ play[inf] ________ on________
k1 opt k2 k7
v[+TOPIC]
be tough _______
k2
{Karaka chart for {tough}}
If the tough-adjective is viewed as a passive form of {find tough} as in:
Ram finds it tough to play this tune on the sitar.\\
It is tough for Ram to play this tune on the sitar.
then in a possible analysis the usual rules for sharing of karta
between the main verb ({find tough}
or {is tough}) and the infinitive verb ({to play}) apply
as discussed in an earlier section.
The tough movement can occur over sentences embedded several
levels deep:
It is tough to believe Mohan can play this tune on the sitar.\\
This tune is tough to believe Mohan can play on the sitar.
However, exactly the same mechanism as outlined earlier handles
it. Figure \ref{kc-tough2} shows
the karaka charts used as before, except one of
the vibhaktis is marked as TOPIC.
{kc-tough2}
______ to believe ______ is-tough
k1 opt k2
k2
believe
[TOPIC] *
_____ can play ______ on_____ k2
k1 k2 k7
play
k1 k2 k7
Mohan tune sitar
{Karaka charts for {tough} with embedded sentences}
Note that by simply making the vibhakti of k2 as TOPIC in an
embedded sentence, everything else falls into place.
The following sentence which has a combination of shared karta
between the main verb and the infinitive and tough movement is
handled correctly as the reader can verify:
This tune is tough for Ram to believe Mohan can play on the sitar.
|
subsection{Pied Piping}
(TOP)
In a normal relative clause, only the wh-element ``moves" from
its position to the TOPIC position. For example,
The party to which it was a privilege to be admitted
was held in Lucknow.
In pied piping, a number of items move with the wh-element (or
follow the pied piper) as shown below:
The party to be admitted to which was a privilege was held in Lucknow.
Modifier-modified structure in Figure \ref{pied-pipe} corresponds to the above sentences.
{pied-pipe}
.
hold [passv.] is-privilege
k7 k2 k
admit [passv.]
k2
Lucknow party
{Modifier modified structure to illustrate pied-piping}
It is interesting to bring a contrast with Hindi.
For the above modifier-modified structure, the following sentences could be
generated in Hindi which respectively correspond to the two
English sentences just given:
? bhavya paartii jisameM aadara kii baata thii pravesha diyaa jaanaa
grand party in-which respect's thing was to-be-admitted
lakhanauu meM hotii thii.
Lucknow -in held was
(The grand party to which it was an honour to be admitted was held in Lucknow.)
bhavya paartii jisameM pravesha diyaa jaanaa aadara kii baata thii
lakhanauu meM hotii thii.
(The grand party to be admitted to which was an honour was held in Lucknow.)
Note that what is called pied piping is the normal preferable
sentence in Hindi\footnote{This possibly has to do with the free word
order and verb final nature of Hindi together with ease of
information extraction.}.
The English sentence with pied-piping can be derived by
relaxing the wh-movement (karaka chart wh-transformation) out of
{admit} and performing normal passive karaka chart transformation
for {is-privilege}. Again the normal rules for karaka sharing
apply. Figure \ref{kc-pied} has
the relevant transformed karaka charts with some
example strings that can be generated by them.
{kc-pied}
_______ admit[passv,inf] ______ by______
k4 opt. k2 k1 opt
(Ex. Ram to be admitted to the party by Krishna)
______[v,inf] was a privilege for_________
k2 k1 opt.
(Ex. To attend the party was a privilege for Dasarath)
{Karaka charts for pied piping}
Note that normal karaka sharing rules between main verb and
infinitives apply if the optional k1 (karta) does not occur with
the infinitive verb.
The above would not only correctly generate the sentence with
pied piping given above, but would also generate correctly the
following sentences from their respective modifier-modified structures:
For Ram to be admitted to the party
was a privilege for Dasarath.\\
The party for Ram to be admitted to which was
a privilege was held in Delhi.\\
The party to be admitted to which was
a privilege for Ram was held in Delhi.\\
?For Krishna to admit Ram to the party was a
privilege for Vasudeva.\\
The party to admit Ram to which was a privilege for Krishna
was held in Vrindavan.
In this paper, we have presented how the computational Paninian Grammar
formalism can
account for English. The approach is information theoretic and
uses the notion of vibhakti and karaka. Since vibhakti is a marker
at the surface level, and English is a positional language,
therefore, the position of a word in a sentence is also a part
of its generalized vibhakti. To account for the long distance
dependency, a special vibhakti called TOPIC is introduced which
stands for the first position in a clause even though the sentence
might be embedded several levels deep in the clause.
With the above generalization of vibhakti, the karaka charts
and karaka chart transformations work same as before (but naturally,
a transformation of vibhakti appears as movement). A new
operation called join is introduced, besides the normal substitute
operation, for connecting the relative clause modifier with
a noun. The whole framework is based on information theoretic
ideas and the extensions mentioned above are in the Paninian spirit.
Interestingly, the join operation is the same as the
adjoin operation in Tree Adjoining Grammar (Joshi, 1985)
provided the auxiliary trees have their foot node
at either the extreme left or the extreme right leaf node.
There are a number of issues which are being looked at as part
of the ongoing research. One of the issues for example, is how
to handle double movement, e.g., tough and wh-movement in the
same clause:
The instrument which this raaga is tough to play on was designed in
Banaras.
This requires two TOPIC positions, which seem to be required
in a definite order. Similarly, there are issues in why `that' blocks
movement. It seems to be related to the nature of modifier-modified
structure and the vibhakti. In the paper, we have also not
discussed subject-verb agreement. While no difficulty is
anticipated, it needs to be spelt out.
At a more general level, we would also like to compare the
adjoining operation in TAG with no restrictions on auxiliary
trees with a suitable operation in PG.
Finally, we have not discussed parsing issues in this paper.
Paninian parser for Indian languages has already been described
earlier in Bharati et al. (1995). Parser for the new generalized
vibhakti and the join operation needs to be worked out in detail.
It is expected to be similar in spirit to before, as only the
constraints have changed. However, a more efficient parsing algorithm
is expected to be designed after the study of the particular nature
of the constraints.
|
{Acknowledgements}
>
(TOP)
We would like to acknowledge the earlier work done by
Rajesh Bhatt on the same problem of applying PG to
English (Bhatt, 1993) where pre-cursors to some of the ideas
given here were explored. The solution using TOPIC position
for wh-movement and other phenomena related to long-distance
dependency outlined in this paper,
differs from the treatment outlined in Bhatia (1995), and
was worked out after Medhavi Bhatia moved out of Kanpur. He should not
be held responsible for errors, if any, in this part of the theory.
The Bibliography
-
{Bharati:93c}
{Bharati, Akshar}, Vineet Chaitanya, and Rajeev Sangal,
\newblock {Anusaraka or Language Accessor: A Short Introduction},
\newblock In { Automatic Translation},
Thiruvananthpuram, Int. school of Dravidian Linguistics, 1994.
-
{Bharati:95}
{Bharati, Akshar}, Vineet Chaitanya, and Rajeev Sangal,
{ Natural Language Processing: A Paninian Perspective},
Prentice-Hall, New Delhi, 1995.
-
{Bhatia:95}
{Bhatia, Medhavi}.
\newblock {\em Paninian Theory Applied to English}.
\newblock Dept. of CSE, IIT Kanpur, 1995.
\newblock Bachelor's Thesis.
-
{Bhatt:93}
{Bhatt, Rajesh}.
\newblock {\em Paninian Theory for English}.
\newblock Dept. of CSE, IIT Kanpur, 1993.
\newblock Bachelor's Thesis.
-
Joshi, Aravind K.,
Tree Adjoining Grammar: How much Context-Sensitivity is Required
to provide reasonable Structural Description. In D. Dowty,
L. Karttunen, and A. Zwicky (eds.), {\em Natural Language Parsing},
Cambridge University Press, Cambridge, UK, 1985.
-
Narayana, V.N., {\em Anusarak: A Device to Overcome
the Language Barrier}, Ph.D. thesis, Dept. of CSE, IIT Kanpur,
January 1994.
Anusaaraka Home Page
|
| | |
|