Chapter. 1 Introduction
Chunking POS Tagged Corpus Samples
Yüklə 1,52 Mb.
|
- Bu səhifədəki naviqasiya:
- Kashmiri Sentence 43
- Chunks [[
- Chunking Issues
- Statisticsl Results
- Chunking Guidelines
Chunking POS Tagged Corpus Samples As aforementioned, chunking is labeling a cluster of POS annotated words (with an obvious head) or a solitary POS annotated word (which itself acts as a head), with a higher level tag. During chunking, words have been clustered together and assigned a particular chunk tag, keeping in view their POS tags, adjacency and dependency relations between them that make them perceptually closed entities. It has been done in such a way that each chunk has a definite internal structure, i.e. words constituting a chunk are asymmetrically related to each other, with one word as a head and the remaining words as its dependents or in case the word is solitary, it is itself head with no dependents. However, there are certain cases where the word which has been given chunk status is neither a head nor a dependent as per semantic dependency is concerned, e.g. AUXP, as discussed above. The chunking process has been carried out using the same interface which was used to carry out POS tagging. The chunk layer has been built on the POS layer as illustrated below in three steps for the sentences 43 and sentence 42 (taken from the corpus), given in the Table.2 along with their English translation and the chunk information. The POS annotated file in SSF format can be opened in the Sanchay SA Interface (GUI), as shown in the Fig.1.a in order to carry out manual chunking.
Table.2. Showing Example Sentence 
Figure.1.a. SA Interface Showing POS Tagged Sentence Step-1: In this step, the contiguous words which form a chunk have been selected by holding control key and clicking on the nodes so that all the contiguous nodes are selected, simultaneously, as shown in the Fig.1.a. Although, the first three chunks (NP, VGF and CCP) consist of solitary words, they have been also chunked following the same steps as shown for the forth chunk (highlighted one in Fig.1.b), i.e. by selecting the nodes, adding a layer and changing the name of the layer (chunk name) for the selected nodes. 
Figure.1.b. SA Interface Showing Step-1 of Chunking Step-2: In step two, one can right click on the selected chunk so that the drop down list of actions opens in which ‘Add Layer’ option can be selected and new chunk layer can be added in the format as shown in the Fig.1.b.
Figure.1.c. SA Interface Showing Step-2 in Chunking Step-3: In this step, the newly added layer would have some default tag (NP) which can be easily changed by clicking on the chunk tag itself and using keyboard by pressing the first letter key of the desired chunk tag. One can keep pressing the letter key unless the desired chunk tag is assigned to the newly added chunk layer, as shown in the Fig.1.d.
Figure.1.d. SA Interface Showing Step-3 in Chunking As shown above, sentence-43 has 26 token which have been grouped into 18 chunks and sentence-42 has 29 tokens which have been grouped into 16 chunks. The ratio between the tokens/words and chunks is not very large (approx 1.6) which is indicative of the fact that there is high frequency of the solitary words that have been given status of chunk. The 18 chunks of the sentence-43, as viewed in the tree viewer of the interface are shown in Fig.2.a and Fig.2.b.
Figure.2.a. SA Interface Showing Chunks in Sentence 43 
Figure.2.b. SA Interface Showing Chunks in Sentence 43
Table.3. Showing Chunked Sentences in SSF
As discussed above, chunking covers half of the dependency relations, though they are not explicitly marked as relational labels which is a general practice in dependency treebanks and can be clearly seen in the works of Nivre (2009) who has translated Tesnière’s (1959) seminal work on dependency grammar. However, applying the framework which has been designed for treebanking in ILs (Bharati et. al. 1995, 2006) on Kashmiri, entirely new issues come to fore. Such issues have partly stemmed from the underlying theory and partly from the peculiar morphosyntactic or syntactic properties of Kashmiri that distinguish it from rest of the ILs and bring it closer to Germanic Languages like German and Yiddish. The main issues that have been encountered during the manual chunking of Kashmiri corpus are briefly given below.
It has been found that the notion of verb group that was proposed for ILs, do not stand for Kashmiri corpus because of a unique syntactic feature of Kashmiri language known as V2 Phenomenon. V2 phenomenon occurs in all tensed clauses be it matrix clause or embedded clause, both in active and passive configurations. It is due to this phenomenon that the tense auxiliary and the main verb cease to exist contiguously. The tense auxiliary (VAUX) occur at the second (V2) position and the main verb (VM) occur at the final position of the sentence but if there is also modal auxiliary in the sentence, it occupies the third (V3) position. For example: farooq/NNP chu/VAUX batI/NN khevaan/VM Farooq is eating rice. farooq/NNP chu/VAUX aasaan/VAUX batI/NN khevaan/VM Farooq keeps eating rice. However, in interrogative sentence the tense auxiliary can also occur at third (V3) position in the sentence if the auxiliary carrying aspectual information is also present, it occurs at forth (V4) position. For example:
What Farooq is giving to Riyaz? farooq/NNP kya/WH chu/VAUX aasaan/VAUX reyaazas/NNP divaan/VM What Farooq keeps giving to Riyaz? The problem with the finite clauses in Kashmiri is that they can’t be easily chunked like in other IA languages, e.g. Hindi, Urdu or Punjabi, etc, due to the presence of V2 phenomenon. The tensed verb stands discontinuous from its main verb as shown in the above examples. Usually, a group or cluster of words is assigned a chunk label if the words are adjacent or contiguous to each other and also have an asymmetric relation of dependence with each other or simply share unequal category status, so that the one with higher status can be projected as the head but in this case the VAUXs and the VMs are neither adjacent to each other nor the relationship they hold with each other is dependency relation in the real sense. Dependencies are essentially modifier-modified relations and between the discontinuous VAUX and VM in finite clauses, there is hardly any modifier-modified kind of relation but definitely a part-whole kind of relationship. Therefore, it impossible to posit verb as a chunk as noun and adjective chunks have been posited. Some ad hoc decisions need to be taken to tackle the V2 problem as the language data is far from being ideal to fit for our perceived notion of chunk.
Adverbs are considered as the most floating or movable elements in a sentence. They frequently occur, discontinuously, away from their heads (VMs), at beginning, at the final position or at elsewhere in the sentence. However, sometimes they occur adjacent to the VM, they modify and thus, become parts of VGF, VGNF or VGNN. When adverbs (RB) occur discontinuously, they have no governing or influencing head adjacent to them and are authority in themselves. Under such circumstances, RBs can be considered as heads, though they would be pseudo-heads and would be still having a clear cut dependency relation with their far away head, which is also the ultimate head, the root. It is evident that the dependency relation of lower level (chunk level) has been promoted to the dependency relation of higher level (argument structure level) to handle discontinuous verb chunk.
Since there is well known notion of functional heads in both the constituency based as well as in dependency based frameworks, at least for exocentric constructions, a distinction has been maintained for when a cluster of words (N-PP) has noun its head and when it has post/preposition its head. In other words, when N-PP cluster is a noun phrase/chunk and when it is a post/prepositional phrase/chunk, has been properly distinguished. However, no such distinction has been drawn on the functional basis, i.e. on the basis of functioning as an argument or an adjunct. In all the clusters of words containing noun, nouns have been treated as their head but never the pre/postposition, irrespective of the fact that some of them perform core functions (subject or objects) and many of them perform mere subsidiary functions (adverbial) in the sentence. This uniformity has been maintained at this level because the underlying notion of the head in PCG (Bharati, 1995) is essentially a semantic notion, with few exceptions76. The function words are devoid of semantics or content and can’t be treated as heads based on the underlying theory. Therefore, there is no possibility of existence of pre/postpositional phrases or chunks unlike what been originally posited in Bloomfieldean and Post-Bloomfieldean literature for exocentric constructions, as already given chapter two, which latter appeared in PSG (Chomsky, 1956) and DG (Tesniere, 1959). It is worth to mention that in these works, NPs are generally arguments and PPs are adjuncts but this distinction has been avoided here for the sake of theory and has been encoded at the next level of annotation.
The notion of dependency does not always provide unambiguous solutions when it comes to exocentric constructions. The dependency representation is at a loss when it comes to representing the notorious paratactic linguistic phenomena such as coordination, whose nature is symmetric (two or more conjuncts play the same role), as opposed to the head-modifier asymmetry of dependencies (Popel et al., 2013). In other words, coordination is a pending problem of natural languages and both PSG as well as DG struggle with it (Hudson, 1988, Covington, 1980). Conjuncts are also exocentric constructions but they have not been treated as endocentric constructions like the pre/postpositional phrases/chunks have been, given their crucial role in the structural organization of sentences.
Kashmiri has negative elements in free-form like na and nI (no and not) as well as in bound-form like -nI (not) in khe-yi-nI (will not eat) and sometimes there is also double negation, e.g. khe-yi-nI kehn (will not eat no) and na saa na (no +honrific not). The negative markers do not belong to this level and certain negative particles in double negation constructions (see the above example) which do have obvious heads, are of no concern here and can’t be projected as chunks. However, some negative particles, either solitary or in clusters (RP-RP), do not have any obvious heads and they themselves have the potential of being head.
Discourse elements are the particles that have been tagged as particle default (RPD) at POS level. They conjoin sentences at semantic or discourse level to bring cohesion in the text. Since, they were extraneous to the existing set of chunks and like conjunctions, in spite of being functional words don’t seem to be dependents of any existing semantic head. It must be noted that discourse elements have been also treated as heads (connectives) in discourse treebanks.
As aforementioned, at chunk level one needs to handle two kinds of grammatical relations, one lower level dependencies, e.g. between JJ and N or RB and VM and a kind of part-whole relations, e.g. between N and PP, N and RP, VAUX and VM. It would be more result oriented if one focuses on the one type of relation at a time. Therefore, ones need to keep track of the kind of the relations one has to handle without confusing between the dependencies and part-whole relations.
The quantitative results are given in terms of chunk statistics and qualitative results are given in terms of a miniature guideline. The four datasets that had been used for POS tagging have been reduced to only three datasets by merging second and third ones. These three datasets have been utilized in chunking which has been carried out by using SA Interface of Sanchay as aforementioned and chunk frequency of each dataset has been obtained with the help of the same interface. The frequency distribution table so obtained has been latter used to calculate the cumulative frequency and the percentage of the chunks. The same data is represented through the bar chats given in the Fig.3.a and 3.b. Figure.3.a. Showing Cumulative Frequency of Chunks
The NP being the most frequent and VGF being the second most frequent chunk is expected from the empirical facts about POS categories that have been as given in Chapter-IV. The statistical chunk results have shown an important empirical fact that 27.630% clauses show V2 phenomenon and 72.369 % clauses are devoid of V2 phenomenon in which tense is condensed in the main verb itself. However, it must be noted that only tensed verbs have not been considered finite but the all verbs, which have not become de-verbal and possess aspectual or modal information, have been considered as finite and it is because of this reason that comparatively lesser percentage of finite clauses have been found with V2 phenomenon which otherwise could have been larger. The statistical results reveal another empirical fact that 78.438 % verbs in Kashmiri are finite and 21.565 % verbs are non-finite. In the non-finite forms, 46.913 % are gerunds and remaining 53.086 % are other non-finite forms. The bar diagram in Fig.3.b shows the data represented in Fig.3.a in terms of percentage. It was done just to reveal the striking quantitative similarities among the three datasets and to put forward a numerical generalization about the percentage of various chunks in the corpus so that one can claim reliably that NPs constitute more that 50% of chunks in Kashmiri.
Figure.3.b. Showing Relative Proportion of Chunks
Chunk guidelines include various decisions that have been taken to resolve various chunk issues raised during the chunking the data. These guidelines can be followed in order to achieve the consistency in future chunking tasks.
|
