<?xml version="1.0"?><!DOCTYPE article SYSTEM "/project/take/software/searchbench_offline_processing/paperxml_generator/aclextractor/src/python/../resource/dtd/paperxml.dtd"><article><header><firstpageheader><page local="1" global="1243"/><title>Automatic Transformation of the Thai Categorial Grammar Treebank to Dependency Trees</title><pubinfo>Proceedings of the 5th International Joint Conference on Natural Language Processing,pages 1243-1250, Chiang Mai, Thailand, November 8 - 13, 2011. ©2011 AFNLP</pubinfo><author surname="Rishøj" givenname="Christian"><org  name="University of Copenhagen" country="Denmark" city="Copenhagen"/></author><author surname="Ruangrajitpakorn" givenname="Taneth"><org  name="University of Copenhagen" country="Denmark" city="Copenhagen"/></author><author surname="Boonkwan" givenname="Prachya"><org  name="University of Copenhagen" country="Denmark" city="Copenhagen"/></author><author surname="Supnithi" givenname="Thepchai"><org  name="University of Copenhagen" country="Denmark" city="Copenhagen"/></author></firstpageheader><frontmatter><p><b>Automatic Transformation of the Thai Categorial Grammar Treebank to Dependency Trees</b></p><p><b>Christian Rish0j</b></p><p>CST</p><p>University of Copenhagen crjensen@hum.ku.dk</p><p><b>Taneth Ruangrajitpakorn Prachya Boonkwan</b></p><p>HLT Lab NECTEC prachya.boo@nectec.or.th</p><p>HLT Lab</p><p>NECTEC</p><p>taneth.rua@nectec.or.th</p><p><b>Thepchai Supnithi</b></p><p>HLT Lab NECTEC</p><p>thepchai.sup@nectec.or.th</p></frontmatter><abstract>A method for deriving an approximately labeled dependency treebank from the Thai Categorial Grammar Treebank has been implemented. The method involves a lexical dictionary for assigning depen­dency directions to the CG types associ­ated with the grammatical entities in the CG bank, falling back on a generic map­ping of CG types in case of unknown words. Currently, all but a handful of the trees in the Thai CG bank can unambiguously be transformed into directed dependency trees. Dependency labels can optionally be assigned with a learned classifier, which in a preliminary evaluation with a very small training set achieves 76.5% label accuracy. In the process, a number of annotation er­rors in the CG bank were identified and corrected. Although rather limited in its coverage, excluding e.g. long-distance de­pendencies, topicalisations and longer sen­tences, the resulting treebank is believed to be sound in terms of structural annota-tional consistency and a valuable comple­ment to the scarce Thai language resources in existence. </abstract></header><body><section number="1" title="Introduction"><p>Syntactic resources play an essential role for the majority of NLP applications, but for the Thai lan­guage, openly available syntactic resources are few in number: So far, the only reported resources are the CG treebank [Ruangrajitpakorn et al., 2009] and the NAiST dependency bank [Wacharaman-otham et al., 2007, Sudprasert, 2008]. Others are either unpublished or minuscule in size. Rather than relying exclusively on labor-intensive manual annotation for further expanding the resources, it would be economically sound to leverage existing efforts and transform an existing treebank in one formalism into another.</p><subsection number="1.1" title="Categorial grammar"><p>Categorial grammar (CG) is a lexicalised theory in natural language syntax motivated by the principle of constitutionality and organised according to the syntactic elements [Steedman, 2000, Ajdukiewicz, 1935], and forms the theoretical basis for the Thai CG treebank. The resource building effort has been very fruitful, but there remains phenomena of Thai language, including long-distance dependen­cies and topicalisation [Warotamasikkhadit, 1997], which are unhandled by the instantiation of CG currently in use.</p><p>Additionally, although the Thai language belongs to a fixed word order typology, Thai spoken language exhibits some flexibility in word order, due to the occasional preference of Thai language users for correspondence in rhyme. As an example, consider the following sentence<footnote anchor="1"/>:</p><p>Ü£Hmu€a Ill's ImiitFi'un</p><p><b>(Lit: </b>England/NE invent/V "-ly"/ADVPFX heavy/ADJ vaccine/N protect/V virus/N avian_flu/NP) "The British are strenuously developing an Avian Flu vaccine."</p><p>The adverbial compound formed by ajj<b>"N </b>VlTn ("strenuously") conventionally occurs after the di­rect object, but is in this sentence, it is realised in the pre-direct object position<footnote anchor="2"/> in order for the last syllable [ <i>nakL] </i>of ajj<b>"N </b>VlTn to rhyme with the first syllable [ <i>wakH] </i>of Tfl§T ("vaccine"). To some degree, this phenomenon from spoken lan­guage shines through in written language, espe­cially in the domains of news and recent politics, and is causing a challenge for the employment of CG grammar.</p><p>xThe Thai adverbialising prefix Bjj"N can be likened to the English "-ly" suffix, which produces an adverbial form from an adjective. Artificial word boundaries are inserted for clarity.</p><footnote label="2">When language users exploit this flexibility in word order to produce aesthetically pleasing sound patterns, it results in a marked form, but the phenomenon is nonetheless productive, and encountered frequently enough to necessitate handling in NLP applications.</footnote><page local="2" global="1244"/><figure caption="Figure 1: Example of ambiguity arising only in dependency representation [Daum et al., 2004]"></figure></subsection><subsection number="1.2" title="Dependency representation"><p>In recent years, dependency grammar has seen a dramatic increase in interest, likely due to a num­ber of appealing properties of the representation. In comparison to phrase structure grammar, d°-pendency structures provide a relatively direct en­coding of predicate-argument structure, which is relevant to sub°equent analyses [Nivre, 2005]. De­pendency representation is arguably better suited for languages with flexible "word order. Addition­ally, having no non-terminal nodes, dependency structures are often petceived as leaving room for less ambiguity as well as being more computation­ally m anageable.</p><p>Certainly, dependency represegtation hesdnaw-backs of its own in terms of gmbigu-ty, some of which aire specificto depyndency repre°entytion. In particular, Figu-o 1 sdaws ft consteoctkm with an unambiguous œnstituœtstrurtnre, wMdi in dependency space isdmbiguyuewfthrecpyct tothoattachment of the adverb [Daum <i>e</i>t nf., ny04].</p><p>Furthermore, dependency stracPuoa p-lowcfota number of wayg to represent coorflnatep phrases-fome having thd <i>tosrdinat-ng </i>tsftj'oycpfona- <i>heoO </i>(CCI<footnote anchor="0"/>!) oe the coordinnte structure, an- f sptdal dependency label t^«^^0 oCat doec not; describe -he grammatical functioaoCohesenjun-os iFiy^v^^^2a). Another option is hoving tlee <i>cotrOCnaCtng </i>-tajoyti <i>tion as dependent </i>jCCD)of oneof -he conjunets, thus allowing one oonjuncl te ocsurwfca e do^ pendency label denoting i-f j;rat^ai-iti-ot hrnc-km (Figure 2b). McDonsld toait Nevre <i>{200t] oSeg </i>a thorough review ot -Soest an-l ottar candodate -oal-yses in use. UnforS-^n^t^^ifi-ne^i^^ot Shi conweivable representations are unyroblseoatic fooma lingmsSic perspective [Daumofah, 2004], or- offer- ths same transparency as the coordinationruks ooCCG do [Boonkwan, 2009].</p><p>Nonetheless, givonChe avayability ofgenotally applicable, trainabit Oependsncf po-feys, and re­ports of beneficial apphcatiogsefdependencyanal-ysis in tasks such ay word-ahgement [Ma eC r-., 2008] and reordering [Chang et al., 2009] for statis­tical machine translation, a dependency treebank of good quality is a highly desirable resource.</p><doubt alpha="100.0" length="2" tooSmall="False" monospace="0.0">PU</doubt><doubt alpha="100.0" length="4" tooSmall="False" monospace="0.0">PRED</doubt><doubt alpha="100.0" length="3" tooSmall="False" monospace="0.0">OBJ</doubt><doubt alpha="44.4" length="9" tooSmall="False" monospace="0.0">/ TV/ Wg\</doubt><p>ROOT They operate ships and banks .</p><p>ROOTTheyoperateshipsand banks .</p><doubt alpha="57.1" length="7" tooSmall="False" monospace="0.0">(b) CCD</doubt><figure caption="Figure 2: Two possible analyses ofacoordination [Kübler et al., 2009]"></figure></subsection><subsection number="1.3" title="Outline"><p>The rest of the paper is organised as follows. In Section 2 wd briefly review other works dealing; waith similarotran.sformt ations,ebeforetpresentingdthe.approach taken in this work in Section 3. Section 4 describes the experimental setting and results. A discussion follows in Section 5, and conclusions in Section 6 .</p></subsection></section><section number="2" title="Related work"><p>In preparation for the CoNLL-X shared task on dependency parsing [Buchholz and Marsi, 2006], a number dependency trees were derived from a number of constituency-based phrase structure treebanks, most of which have grammatical func­tion (e.g. "subject" and "object") as part of the annotation. The convtrsion process for such tree-banks would involve a <i>head table </i>with rules of the form</p><p>• "trie rread child of <b><i>a </i></b>VP/clause is the drild with the IID/predicator/hd/Head function" and</p><p>• "[the dependency label] for a token is the func­tion of the biggest constituent of which this token is the lexical head".</p><p>The case of the Thai CG bank is different, as it does not directly contain any grammatical func­tions. On the other hand, identifying head tokens is relatively straight-forward when augmenting the CG annotation with dependency directions.<page local="3" global="1245"/></p><p>Chanev et al. [2006] faced a similar situation in their transformation of the BulTreeBank to depen­dency representation. Heads were first identified from explicitly stated rules in a head table. Lack­ing explicit grammatical functions in the source treebank, they explored a heuristic rule-based ap­proach for the labeling with a <i>dependency table, </i>containing rules based on parent constituents. Al­though good results are achieved, they report of errors like mistaken subjects and objects.</p></section><section number="3" title="Methodology"><p>The situation with the Thai CG bank is a lit­tle different. Together with a set of combinatory grammar rules, the CG type tags and bracketing present in the treebank unambiguously specify the constituent structure of the treebank sentences. When the CG type tags are augmented with depen­dency directions, a dependency tree can be derived with relative ease from the CG-based constituents. Grammatical functions, however, are not immedi­ately evident from the CG trees.</p><p>We first describe a relatively straight-forward method for deriving the dependency trees, and next consider the more daunting task of assigning functional labels to the dependency arcs. Figure 3 shows a schematical overview of the proposed method.</p><subsection number="3.1" title="Terminology"><p>For any given CG type t, we use the <i>arity </i>(admit­tedly a bit sloppily) to denote the ordered list of arguments expected by a type. The arity of the type <b>s\np/ws/np </b>is thus <b>/np, /ws </b>and <b>\np </b>— that is,• a noun phrase from the right, followed by</p><p>• a subordinate clause beginning with the Thai word 11 ("that", subordinate clause marker), and</p><p>• another noun phrase from the left.</p><p>Complementary, we define the <i>yield </i>of t as the set of possible CG types which may result from func­tional application of a CG rule. The transitive verb type <b>s\np/np, </b>for example, yields after receiving a <b>np </b>to the right, and another <b>np </b>to the left, respectively. This is simply the basic combinatory CCG rules:</p><doubt alpha="66.7" length="9" tooSmall="False" monospace="0.0">• s\npand</doubt><p><b>Algorithm 1 </b>Pseudo-code for propagating depen­dency directions to non-terminal nodes.</p><p><b>def </b>propagateDirections (node ) : <b>for  </b>child  <b>in </b>node . children :</p><p>propagateDirections (child) <b>if </b>node . hasDependencyDirection : <b>return</b></p><p><b>for  </b>child  <b>in </b>node . children : <b>if </b>child . yields(node.type ) : node . depDirs = child . depDirs <b>break</b></p></subsection><subsection number="3.2" title="Dependency directions"><p>The first transformation step, which can be re­garded as a preprocessing step before the actual transformation, involves a lexical dictionary for as­signing dependency directions to the CG types as­sociated with the grammatical entities in the CG bank, falling back on a generic CG to CDG map­ping in case of unknown words.</p><p>Note that only terminal nodes will have as­signed dependency directions assigned by this pro­cedure. Dependency directions are propagated to non-terminal nodes in a bottom-up fashion by the procedure <b>propagateDirections </b>(Algorithm 1). In identifying which child to adopt dependency directions from, the parent node type is checked against the yield of each child node.</p></subsection><subsection number="3.3" title="Head finding"><p>Dependency arcs are assigned by a procedure that implements the CDG dependency derivation rules introduced by Boonkwan and Steedman [2011] (motivated by Collins, 1999). The idea is to trace the derivation implied by the CG rules, registering the dependency relations specified by the CDG-augmentation along the way.</p><p>The derivation rules handled are as follows. Let c : d signify a CDG type c and a dependency struc­ture d, and the notion h(di) — h(d2) represent a dependency arc between the head of d1 and the head of d2 (with h(d1) governing h(d2)) . Then the derivation rules are:</p><doubt alpha="47.7" length="44" tooSmall="False" monospace="0.0">X/&lt;Y:d1Y:d2X/&gt;Y:d1Y:d2Y:diX\&lt;Y:d2Y:d1X\&gt;Y:d2</doubt><doubt alpha="32.0" length="50" tooSmall="False" monospace="0.0">h(d1)&lt;- h(d2)h(d1)— h(d2)h(d1)&lt;- h(d2)h(d1)— h(d2)</doubt><doubt alpha="34.8" length="23" tooSmall="False" monospace="0.0">X/Y Y =&gt; X Y X\Y   =&gt; X</doubt><p>In a simple example, Mary<b>[np] </b><i>drinks</i><b><i>[s\ </i></b><b>&lt; np/ &gt; np] </b><i>fresh</i><b><i>[np/ </i></b><b>&lt; np] </b>rm'lfc<b>[np], </b><i>fresh </i>would combine with <i>milk, </i>yielding an <b>np </b>and a dependency arc specifying <i>milk </i>as head of <i>fresh.</i></p><p>In addition to the standard combinatory rules for forward and backward functional application above, the Thai CG bank makes use of a CCG-style <i>serialisation </i>rule to handle e.g.  serial verb<page local="4" global="1246"/></p><doubt alpha="100.0" length="1" tooSmall="False" monospace="0.0">s</doubt><p>Lexical dictionary</p><p>CG to CDG</p><doubt alpha="100.0" length="3" tooSmall="False" monospace="0.0">map</doubt><p>Gold CDG &amp; labeled dependency tree pairs</p><p>Approximately labeled dep. trees</p><figure caption="Figure 3: Transformation overview"></figure><p>constructions, which are used in Thai (and Chi­nese) to express serial or consecutive events. As Boonkwan [2009], we take the notion of a <i>serial verb construction </i>to mean a series of verbs or verb phrases without explicit connectives marked with (or understood to have) the same grammatical cat­egories, and sharing at least one common argu­ment, typically a subject.</p><p>As an example, the verbs flll^ ("examine") and <b>W</b>U ("find") occur serially in the following sen­tence<footnote anchor="3"/> from the CG bank, indicating a resultative course of events [Thepkanjana and Uehara, 2009]:</p><p>u'nîïnnu em*! wu Iiis TflTiih</p><p>Til tfîJJJfl</p><p><b>(Lit: </b>scientist/N examine/V find/V virus/N corona/N in/PP civet/N) "The scientist examined the civet and found coronavirus."</p><p>We introduce a generalised derivation rule for se­rial constructions which simply designates the head of the first dependency structure as governing the head of the following dependency structure:</p><doubt alpha="30.8" length="26" tooSmall="False" monospace="0.0">X:d1X:d2==   h(d1) — h(d2)</doubt><p>The rule is generalised in the sense that it han­dles <i>serial noun constructions </i>as well as serial verb constructions.</p><p>Further CCG-style combinatory rules, such as functional composition and type raising, are not currently in use in the Thai CG bank, and therefore not handled by the transformation.</p><p>An outline of the head finding procedure is given as Algorithm 2. Intuitively, the algorithm proceeds by, for each node in turn, beginning at the termi­nal nodes, identifying sibling nodes which satisfy the arity of the of the node CG type. For each sibling node satisfying an argument, the depen­dency derivation rule is applied and the sibling is removed.</p><footnote label="3">Artificial word boundaries inserted for clarity.</footnote><p><b>Algorithm 2 </b>Pseudo-code for the head-finding procedure.</p><p><b>def </b>assignHeads (node ) :</p><p><b>for </b>c  <b>in </b>node . nonterminalChildren :</p><p>assignHeads(c) <b>for </b>c  <b>in </b>node . terminalChildren :</p><p>assignHeads(c) <b>for  </b>arg  <b>in </b>node . type . arguments : <b>if </b>arg . side == 'right':</p><p>sibl = node.right Siblings. first <b>else</b></p><p>sibl = node.leftSibli ngs.last <b>break </b>unless arg . matches( sibl ) <b>i f </b>arg . side == ' right ' :</p><doubt alpha="62.1" length="29" tooSmall="False" monospace="0.0">i farg. dependencyDir=='&gt; ' :</doubt><p>registerHead ( sibl , node) <b>else </b>: <i>#</i><i> </i><i>&lt;</i></p><p>registerHead (node , sibl) node . rightSiblings . shift <b>else </b>: <i># left</i></p><p>registerHead (node , sibl) <b>else </b>: <i># &lt;</i></p><p>registerHead ( sibl , node) node . leftS iblings . pop</p><p>It is worth noting that while both terminal and non-terminal nodes are involved in this process, we are only interested in assigning dependency arcs to terminal nodes, as non-terminals are absent in the all-terminal dependency structure. This is en­sured by an implementation detail of the proce­dure <b>registerHead </b>(omitted from Algorithm 2), in which non-terminal nodes act as proxies for their terminal heads.</p></subsection><subsection number="3.4" title="Dependency labeling"><p>Although the CDG-augmentation of the CG tree-bank implies a dependency structure for each sentence, there are no immediate clues avail­able about the specific grammatical functions of dependency arcs. Obviously, there are some clear-cut cases: When a token with CDG type <b>((s\ &lt; np)\ &gt; (s\ &lt; np))\ &lt; num </b>modifies a token with CDG type <b>num, </b>it must be an application of a <i>quantifier </i>(with dependency type "quan"), as ex­emplified in Figure 4.</p><p>Other cases are less obvious. Even when taking<page local="5" global="1247"/></p><doubt alpha="100.0" length="4" tooSmall="False" monospace="0.0">subj</doubt><doubt alpha="100.0" length="4" tooSmall="False" monospace="0.0">modr</doubt><doubt alpha="100.0" length="4" tooSmall="False" monospace="0.0">modp</doubt><doubt alpha="40.0" length="5" tooSmall="False" monospace="0.0">1 I I</doubt><p>-Root- <b><i>wb     </i></b><b><i>vh</i></b><b><i> </i></b><b>5       </b>jjinjna nana «ita fauna principal have   abundantly diverse species</p><doubt alpha="41.7" length="60" tooSmall="False" monospace="0.0">np     np s\&lt;np (s\&lt;np)\&gt;(s\&lt;np) num ((s\&lt;np)\&gt;(s\&lt;np))\&lt;num</doubt><p>Figure 4: An unmistakable labeling case: Given the CDG types of waifl ("diverse") and <b><i>"tili® </i></b>("species"), the only dependency label supported by the examined data is "quan" <i>(quantifier).</i></p><p>dependency direction and argument position into consideration, there are still cases with several pos­sible dependency types, as shown in Figure 5.</p><p>While for many practical purposes an unlabeled dependency structure is sufficient, having proper dependency labels is nonetheless desirable. In lack of an exact transformation, the approach explored in this work relies instead on training a classifier to predict the correct dependency label given local features of the tokens involved, as they occur in the dependency structure derived from the CDG tree.</p><p>From a related work (manuscript in prepara­tion), we have obtained a number of CDG trees from labeled dependency trees. Such pairs of <i>la­beled </i>dependency trees and CDG trees can serve as training material for the label classifier.</p><p>We evaluate different feature sets for classifica­tion task. The basic feature set contains:</p><p>• CDG type of token and its head</p><p>• # of left siblings</p><p>• # of right siblings</p><p>• Dependency direction (L/R)</p><p>• Concatenations of token/head CDG and # of left siblings</p><p>Other feature sets extend the basic set including additional information:</p><p><b>+forms </b>Token and head word form.</p><p><b>+POS(L) </b>Part-of-speech tag for the token, as as­signed by a learned part-of-speech tagger<footnote anchor="4"/>.</p><p><b>+POS(G) </b>"Gold" part-of-speech tag as found in the labeled dependency tree.</p></subsection></section><section number="4" title="Experiments"><p>The Thai CG bank made available for this re­search contains 1,428 phrases with corresponding CG trees. Trees are mostly medium to low in token count, with 49 single-token trees, and the longest phrase having 9 tokens. Median phrase length is 3</p><doubt alpha="100.0" length="2" tooSmall="False" monospace="0.0">ir</doubt><footnote label="4">SVMTool [Giménez and Marquez, 2004], trained l the Orchid corpus [Sornlertlamvanich et al., 1997].</footnote><p>vifwnnuu</p><p>afterwards</p><doubt alpha="40.0" length="10" tooSmall="False" monospace="0.0">s/&gt;(s\&lt;np)</doubt><p><b>"i I </b><i>r^n.</i></p><p>fall effect s\&lt;np/&gt;np np less s\&lt;np</p><doubt alpha="66.7" length="18" tooSmall="False" monospace="0.0">(a) Subject (subj)</doubt><doubt alpha="100.0" length="4" tooSmall="False" monospace="0.0">UBII</doubt><doubt alpha="100.0" length="7" tooSmall="False" monospace="0.0">descend</doubt><doubt alpha="37.5" length="16" tooSmall="False" monospace="0.0">(s\&lt;np)\&gt;(s\&lt;np)</doubt><doubt alpha="66.7" length="6" tooSmall="False" monospace="0.0">-Root-</doubt><doubt alpha="100.0" length="3" tooSmall="False" monospace="0.0">hoe</doubt><doubt alpha="100.0" length="4" tooSmall="False" monospace="0.0">dirt</doubt><p>nan bury uai then</p><doubt alpha="100.0" length="5" tooSmall="False" monospace="0.0">water</doubt><doubt alpha="41.0" length="39" tooSmall="False" monospace="0.0">s\&lt;np/&gt;np    np    s\&lt;np    s\&lt;np s\&lt;np</doubt><p>(b) Direct object (dobj)</p><doubt alpha="33.3" length="9" tooSmall="False" monospace="0.0">r55-! I r</doubt><p>-Root- iJ§n plant</p><doubt alpha="56.2" length="16" tooSmall="False" monospace="0.0">s\&lt;np/&gt;pp pp/&gt;np</doubt><doubt alpha="60.0" length="15" tooSmall="False" monospace="0.0">np s\&lt;np/&gt;np np</doubt><p>(c) adverbialiser (advm)</p><p>Figure 5: Different labeling of left-pointing depen­dency arcs with <b>s\ &lt; np/ &gt; np </b>as head CDG type and <b>np </b>as dependent CDG type. See Figure 6 for depency type abbreviations.</p><p>tokens. With a total token count of 5,143 tokens, the arithmetic mean length is 3.60.</p><p>As there is no explicit sentence boundary marker in Thai [Satayamas and Kawtrakul, 2004], it is of­ten unclear what constitutes a sentence. Thus, rather than ordinary sentences, the treebank con­tains phrases of different types, reflecting the par­titioning of token sequences made by the treebank annotators for the purpose of treebanking:</p><p>• 539 verb phrases or subject-omitted phrases</p><doubt alpha="42.9" length="7" tooSmall="False" monospace="0.0">(s\np),</doubt><doubt alpha="52.6" length="19" tooSmall="False" monospace="0.0">• 363 sentences(s),</doubt><doubt alpha="64.0" length="25" tooSmall="False" monospace="0.0">• 372 noun phrases(np)and</doubt><p>• 4 prepositional phrases <b>(pp).</b></p><p>The lexical dictionary used contains possible CDG types for 38.250 word forms, with an average of 2 types listed per word form. For six of the word forms, the dictionary lists several possible depen­dency directions for a single CG type. These con-fusable CDG types and the dictionary entries they occur for are listed in Table 1.</p><p>An example sentence from the treebank affected by the ambiguous mapping of the adverb flatiü</p><doubt alpha="100.0" length="4" tooSmall="False" monospace="0.0">comp</doubt><p>seed be norm<page local="6" global="1248"/></p><doubt alpha="57.9" length="19" tooSmall="False" monospace="0.0">("at present") is5:</doubt><p>eiaiid iea <b>m</b>a"^ <b><i>ùa </i></b>wnn</p><p><b>(Lit: </b>this-moment/ADV she/PRON "-ing"/AUX busy/ADJ very/ADV) "At this moment she is being very busy."</p><p>Examples of the latter two ambiguity classes of the lexical dictionary (Table 1) were not encoun­tered in the treebank — i.e. the affected word forms do not occur with an ambiguous CG type. In dealing with these ambiguous entries in the lex­ical dictionary, we simply (and naively) choose the first mapping option.</p><p>The generic CG to CDG mapping, used in addi­tion to the lexical dictionary as fallback for word forms not found in the lexical dictionary, also ex­hibits some degree of ambiguity. The seven CG types with multiple possible CDG equivalents are listed in Table 2.</p><p>In evaluating the classification-based approach to assigning dependency labels, a sample of 678 labeled dependency edges from the NAiST depen­dency treebank [Wacharamanotham et al., 2007] was used, along with corresponding CDG trees. The feature sets suggested in section 3.4 on page 4 were evaluated with four different classifiers<footnote anchor="6"/> (see</p><p>Table 3).</p></section><section number="5" title="Discussion"><p>Transforming CDG-augmented CG trees to unla-beled dependency trees was successful. In this work, the issue of ambiguous CDG types affects only a very small number of trees, but remains an issue to be aware of.</p><footnote label="5">The Thai auxillary verb ma A indicates the present participle, meaning "in the act of", similar to the English suffix, "-ing". Articifial word boundaries are inserted for clarity.</footnote><footnote label="6">Experiments with the learners were done using leave-one-out cross-validation, with the exception of LibSVM, which was run using the standard K-fold cross-validation of the easy.py script [Chang and Lin,</footnote><p>2001].</p><p>For dependency labeling, only a small amount of training data (in the form of sentences with both CDG and labeled dependency analyses) was avail­able for this prelimenary experiment. Using this, dependency labels were not reliably recoverable (76.5% label accuracy). However, it seems hope­ful that better recovery of dependency labels can be obtained with this approach once more training material become available.</p><p>Head and dependent word forms, as well as part-of-speech tags, were beneficial as added features for the label classifier, resulting in a substantial reduction in error rate — from 0.357 to 0.235 (« 34%).</p><p>On the other hand, rather than the approx­imated classfier-based approach to labeling, one could consider settling for an exact but partial la­beling by only assigning those dependency labels which unambigously arise from tuples of head and dependent CDG types. However, the dependency labels obtainable with absolute certainty in this way are often of the less interesting kind — e.g. "conj" for a <b>s\ &lt; np </b>governed by a <b>s/ &gt; (s\ &lt; np) </b>— while more useful labels remain ambiguous.</p><table caption="Table 1: Entries in the lexical dictionary which are ambiguous with respect to dependency direction" class="main" frame="box" rules="all" border="1" regular="False"><tr class="row"><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>Word form(s)</p></td><td class="cell"><p>CG type</p></td><td class="cell"><p>Possible CDG equivalents</p></td><td class="cell"><p>Interpretation</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p></p></td><td class="cell"><p><b>(s\npV(s\np)</b></p></td><td class="cell"><p><b>(s\ &lt; np)/ &gt; (s\ &lt; np) (s\ &lt; np)/ &lt; (s\ &lt; np)</b></p></td><td class="cell"><p>Adverb "please" Auxiallary verb "should"</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p><b><i>WiiiiViii </i></b>("at that time") flaiiS ("at present") ^M^iTii ("at that moment") <b><i>^V&amp;ZV, </i></b>("this moment")</p></td><td class="cell"><p><b>s/s</b></p></td><td class="cell"><p><b>s/ &gt; s s/ &lt; s</b></p></td><td class="cell"><p>Conjunction Adverb of time</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>nil</p></td><td class="cell"><p><b>np\np</b>/<b>np</b></p></td><td class="cell"><p><b>np\ &gt; np/ &gt; np np\ &lt; np/ &gt; np</b></p></td><td class="cell"><p>Preposition "with" Conjunction "and"</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td></tr></table><table caption="Table 2: Cases of CD to CDG mappings which are ambigous with respect to dependency direction" class="main" frame="box" rules="all" border="1" regular="False"><tr class="row"><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>CG type</p></td><td class="cell"><p>Possible CDG equivalents</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p><b>np\np</b>/<b>np</b></p></td><td class="cell"><p><b>np\ &gt; np/ &gt; np</b></p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p></p></td><td class="cell"><p><b>np\ &lt; np/ &gt; np</b></p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p><b>s/s</b></p></td><td class="cell"><p><b>s/ &gt; s</b></p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p></p></td><td class="cell"><p><b>s/ &lt; s</b></p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p><b>s\s</b></p></td><td class="cell"><p><b>s\ &gt; s</b></p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p></p></td><td class="cell"><p><b>s\ &lt; s</b></p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>(<b>s\np</b>)/(<b>s\np</b>)</p></td><td class="cell"><p><b>(s\ &lt; np)/ &gt; (s\ &lt; np)</b></p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p></p></td><td class="cell"><p><b>(s\ &lt; np)/ &lt; (s\ &lt; np)</b></p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p><b>s</b>/(<b>s\np</b>)</p></td><td class="cell"><p><b>s/ &gt; (s\ &lt; np)</b></p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p></p></td><td class="cell"><p><b>s/ &lt; (s\ &lt; np)</b></p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p><b>s\</b>(<b>s\np</b>)</p></td><td class="cell"><p><b>s\ &lt; (s\ &lt; np)</b></p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p></p></td><td class="cell"><p><b>s\ &gt; (s\ &lt; np)</b></p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p><b>s</b>/(<b>s\np</b>)/<b>np</b></p></td><td class="cell"><p><b>s</b>/ <b>&gt; </b>(<b>s\np</b>)/ <b>&gt; np</b></p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p></p></td><td class="cell"><p><b>s</b>/ <b>&lt; </b>(<b>s\np</b>)/ <b>&gt; np</b></p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td></tr></table><page local="7" global="1249"/><p>Table 3: Accuracy of different classifiers and feature sets in recovering the correct dependency labels</p><p><b>Complements </b><i>subject </i>(subj) • <i>clausal subject </i>(csubj) • <i>direct object </i>(dobj) • <i>indirect object </i>(iobj) • <i>prepositional object </i>(pobj) • <i>preposi­tional complement </i>(pcomp) • <i>subject or ob­ject predicative </i>(pred) • <i>clausal predicative </i>(cpred) • <i>conjunction </i>(conj) • <i>subordinating conjunction </i>(sconj) • <i>nominaliser </i>(nom) • <i>adverbialiser </i>(advm)</p><p><b>Adjuncts </b><i>parenthetical modifier </i>(modp) • <i>re­strictive modifier </i>(modr) • <i>tense modifier </i>(modt) • <i>mood modifier </i>(modm) • <i>aspect modifier </i>(moda)  •  <i>locative modifier </i>(modl)<i>• parenthetical apposition </i>(appa) • <i>restric­tive apposition </i>(appr) • <i>relative clause modi­fication </i>(rel) • <i>determiner </i>(det) • <i>quantifier </i>(quan) • <i>classifier </i>(cl) • <i>coordination </i>(co­ord) • <i>negation </i>(neg) • <i>punctuation </i>(punc)</p><p><i>• double preposition </i>(dprep) • <i>parallel serial verb </i>(svp) • <i>sequence serial verb </i>(svs)</p><p>Figure 6: Dependency types from the annotation guidelines [Sudprasert, 2008] for the NAiST depen­dency treebank.</p></section><section number="6" title="Conclusion and future work"><p>In the process, a considerable amount of syntacti­cal and annotational errors in the Thai CG bank were identified and corrected. The authors are of the belief that this work has not only provided a means for continual expansion of resources for Thai natural language processing, but also helped improve the quality the existing CG resource.</p><p>As a related work (manuscript in preparation) progresses, more CDG trees for labeled NAiST de­pendency trees will become available. With this extra training material, the learned classifier used for labeling in this work should become more reli­able. Further improvement might also be achiev­able from an expanded feature set, including for ex­ample the argument position (in addition to side) and one or two generations of grandparent nodes.</p><p>Further development of the Thai CDG formal­ism is also expected, in particular for the analysis of sentence-like noun phrases. This will likely need special handling in the dependency representation as well. Currently, the NAiST annotation guidelines do not specify a label for this phenomenon. <b>Acknowledgements</b></p><p>The authors wish to thank the NAiST unit of Kasetsart University and the HLT Lab of NECTEC for making their treebanks available for experiments.</p><table class="main" frame="box" rules="all" border="1" regular="False"><tr class="row"><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>Classifier</p></td><td class="cell"><p>Basic</p></td><td class="cell"><p>+forms</p></td><td class="cell"><p>+POS(L)</p></td><td class="cell"><p>+POS(G)</p></td><td class="cell"><p>+POS(L) &amp; +forms</p></td><td class="cell"><p>+POS(G) &amp; +forms</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>Random Forest</p></td><td class="cell"><p>61.9%</p></td><td class="cell"><p>69.6%</p></td><td class="cell"><p>66.1%</p></td><td class="cell"><p>68.9%</p></td><td class="cell"><p>72.9%</p></td><td class="cell"><p>71.8%</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>LibSVM</p></td><td class="cell"><p>64.3%</p></td><td class="cell"><p>73.9%</p></td><td class="cell"><p>66.1%</p></td><td class="cell"><p>70.2%</p></td><td class="cell"><p>74.5%</p></td><td class="cell"><p><b>76.5%</b></p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>Nearest Neighbors</p></td><td class="cell"><p>60.5%</p></td><td class="cell"><p>64.8%</p></td><td class="cell"><p>62.4%</p></td><td class="cell"><p>64.5%</p></td><td class="cell"><p>70.5%</p></td><td class="cell"><p>70.4%</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>Naive Bayes</p></td><td class="cell"><p>60.6%</p></td><td class="cell"><p>63.7%</p></td><td class="cell"><p>61.7%</p></td><td class="cell"><p>65.8%</p></td><td class="cell"><p>63.7%</p></td><td class="cell"><p>66.2%</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td></tr></table></section><references><p>K. Ajdukiewicz. Die syntaktische konnexität. <i>Stu-dia Philosophica, </i>1(1):27, 1935.</p><p>P. Boonkwan. A memory-based approach to the treatment of serial verb construction in combi-natory categorial grammar. In <i>Proceedings of the 12th Conference of the European Chapter of the Association for Computational Linguistics: Student Research Workshop, </i>page 10-18, 2009.</p><p>P. Boonkwan and M. Steedman. Grammar in­duction from text using small syntactic proto­types. In <i>Proceedings of the 5th International Joint Conference on Natural Language Process­ing (to appear), </i>Chiang Mai, 2011.</p><p>S. Buchholz and E. Marsi. CoNLL-X shared task on multilingual dependency parsing. In <i>Tenth Conference on Computational Natural Language Learning, </i>page 149, 2006.</p><p>A. Chanev, K. Simov, P. Osenova, and S. Mari-nov. Dependency conversion and parsing of the BulTreeBank. In <i>Proc. of the LREC-Workshop Merging and Layering Linguistic Information,</i> 2006.</p><p>C. Chang and C. Lin.   LIBSVM: a library for support vector machines, 2001. URL http://www.csie.ntu.edu.tw/cjlin/libsvm<b>.</b></p><p>P. C Chang, H. Tseng, D. Jurafsky, and C. D Manning. Discriminative reordering with chinese grammatical relations features. In <i>Proceedings of the Third Workshop on Syntax and Structure in Statistical Translation, </i>page 51-59, 2009.</p><p>M. Collins. <i>Head-driven statistical models for nat­ural language parsing. </i>PhD thesis, University of Pennsylvania, 1999.</p><page local="8" global="1250"/><p>M. Daum, K. Foth, and W. Menzel. Automatic transformation of phrase treebanks to depen­dency trees. In <i>Proceedings of LREC, </i>2004.</p><p>J. Giménez and L. Marquez. SVMTool: a general POS tagger generator based on support vector machines. In <i>In Proceedings of the 4th Inter­national Conference on Language Resources and Evaluation, </i>2004.</p><p>S. Kübler, R. McDonald, and J. Nivre. Depen­dency parsing. <i>Synthesis Lectures on Human Language Technologies, </i>2(1):1-127, 2009.</p><p>Y. Ma, S. Ozdowska, Y. Sun, and A. Way. Improv­ing word alignment using syntactic dependen­cies. In <i>Proceedings of the ACL-08: HLT Second Workshop on Syntax and Structure in Statistical</i></p><p><i>Translation (SSST-2), </i>page 69-77, 2008.</p><p>R. McDonald and J. Nivre. Characterizing the er­rors of data-driven dependency parsing models. In <i>Proc. of the Joint Conf. on Empirical Meth­ods in Natural Language Processing and Compu­tational Natural Language Learning (EMNLP-</i></p><p><i>CoNLL), </i>2007.</p><p>J. Nivre. Dependency grammar and dependency parsing. <i>MSI report, </i>5133, 2005.</p><p>T. Ruangrajitpakorn, K. Trakultaweekoon, and T. Supnithi. A syntactic resource for thai: CG treebank. In <i>Proceedings of the 7th Workshop on Asian Language Resources, </i>page 96-101, 2009.</p><p>V. Satayamas and A. Kawtrakul. Wide-Coverage grammar extraction from thai treebank. In <i>Pro­ceedings of Papillon 2004 Workshops on Multi­lingual Lexical Databases, </i>2004.</p><p>V. Sornlertlamvanich, T. Charoenporn, and H. Isa-hara. ORCHID: thai part-of-speech tagged cor­pus. Technical Report TR-NECTEC-1997-001, 1997.</p><p>M. Steedman. <i>The syntactic process, </i>volume 131.</p><p>MIT Press, 2000.</p><p>S.     Sudprasert. Dependency annota-</p><doubt alpha="56.9" length="51" tooSmall="False" monospace="0.0">tion   guideline   for   thai,    version   1.4 (in</doubt><doubt alpha="56.4" length="39" tooSmall="False" monospace="0.0">thai).     Technical report,  2008. URL</doubt><p>http://naist.cpe.ku.ac.th/tred/<b>.</b></p><p>K. Thepkanjana and S. Uehara. Resulta-tive constructions with "implied-result" and "entailed-result" verbs in thai and en­glish:     a   contrastive   study. <i>Linguistics,</i></p><doubt alpha="17.5" length="40" tooSmall="False" monospace="0.0">47(3):589-618, 2009. ISSN 0024-3949. URL</doubt><p>http://www.thefreelibrary.com/-a0204693751<b>.</b></p><p>C. Wacharamanotham, M. Suktarachan, and A. Kawtrakul. The development of web-based   annotation   system   for   thai tree-</p><doubt alpha="50.0" length="38" tooSmall="False" monospace="0.0">bank.    page 305, Thailand, 2007. URL</doubt><p>U. Warotamasikkhadit. Fronting and backing top-icalization in thai.    <i>Mon-Khmer Studies, </i>27:</p><p>303-6, 1997.</p></references></body></article>