<?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="420"/><title>Determining Immediate Constituents of Compounds in GermaNet</title><pubinfo>Proceedings of Recent Advances in Natural Language Processing,pages 420—426, Hissar, Bulgaria, 12-14 September 2011.</pubinfo><author surname="Henrich" givenname="Verena"><org  name="University of Tubingen" country="Germany" city="Tubingen"/></author><author surname="Hinrichs" givenname="Erhard"><org  name="University of Tubingen" country="Germany" city="Tubingen"/></author></firstpageheader><frontmatter><p><b>Determining Immediate Constituents of Compounds in GermaNet</b></p><p><b>Verena Henrich Erhard Hinrichs</b></p><p>University of Tübingen University of Tübingen</p><p>verena.henrich@uni- erhard.hinrichs@uni-tuebingen.de tuebingen.de</p></frontmatter><abstract>In order to be able to systematically link compounds in GermaNet to their constit­uent parts, compound splitting needs to be applied recursively and has to identify the immediate constituents at each level of analysis. Existing tools for compound splitting for German only offer an analy­sis of all component parts of a compound at once without any grouping of subcon-stituents. Thus, existing tools for splitting compounds were adapted to overcome this issue. Algorithms combining three heterogeneous kinds of compound split­ters are developed to achieve better re­sults. The best overall result with an ac­curacy of 92.42% is achieved by a hybrid combined compound splitter that takes into account all knowledge provided by the individual compound splitters, and in addition some domain knowledge about German derivation morphology and compounding. </abstract></header><body><section number="1" title="Introduction"><p>The present paper presents a compound splitter for German that is tailored to the needs of sys­tematically enriching the set of lexical relations of <i>GermaNet </i>(Kunze and Lemnitzer, 2002; Hen-rich and Hinrichs, 2010), the German version of the Princeton WordNet for English (Fellbaum, 1998). Compounding is a highly productive word formation process resulting in complex words with two or more constituent parts. Baroni et al. (2002) report that almost half (47%) of the word types in the APA German news corpus are com­pounds.</p><p>For GermaNet, the numbers are comparable: The morphological analyzer <i>SMOR </i>(Schmid et al., 2004) for German classifies 46.89% of all lexical units contained in release 6.0 of Germa-</p><p>Net as compounds. Among those, nominal com­pounds make up 95% and are thus by far the largest class of compounds. It is for this reason that we concentrate exclusively on the treatment of nominal compounds in the present study.</p><p>Given the prevalence of compounds in Germa-Net and its current coverage of 84586 lexical units, a systematic treatment of compounds is badly needed in order to enhance the usability of GermaNet for a wide variety of NLP applica­tions, including machine translation, natural lan­guage generation, information extraction, etc. The size of GermaNet and the high frequency of compounds clearly prohibit a purely manual so­lution and mandate an automatic treatment. The treatment of compounds for GermaNet needs to be systematic along at least three dimensions: (i) it should cover all combinations of word classes present in GermaNet which can enter into noun compounding, (ii) it should apply to all lexical units already entered into GermaNet, and (iii) it should be extendable to all compounds which are candidates for inclusion in GermaNet in future data releases.</p></section><section number="2" title="Nominal Compounds in German"><p>Peter Eisenberg (Eisenberg, 2006) defines four major subclasses for compounds, where the rightmost head constituent is a noun.</p><p>1. Noun + Noun: <i>Apfelbaum </i>'apple tree'.</p><p>2. Adjective + Noun: <i>Weißbrot </i>'white bread'.</p><p>3. Verb + Noun: <i>Esstisch </i>'eating table'.</p><p>4. Preposition + Noun: <i>Oberarm </i>'upper arm'.</p><p>In addition to these four major classes, there is a small class of bound morphemes (i.e., mor­phemes that cannot appear as an independent word), such as <i>Him-<footnote anchor="1"/>, </i>that can also serve as the initial constituent of a nominal compound:</p><footnote label="1">In the German linguistics literature such bound mor­phemes are referred to as unikale Elemente.</footnote><page local="2" global="421"/><p>5.   Bound Morpheme + Noun: <i>Himbeere </i>'rasp­berry'.</p><p>What makes compound splitting for German a challenging task is the fact that compounding is not always simple string concatenation, but often involves the presence of intervening linking ele­ments or the elision of word-final characters in the non-head constituent of a compound<footnote anchor="2"/>. Word­final <i>e, </i>for example, is absent in compounds such as <i>Hüftschwung </i>'hip swing', whose non-head constituent is <i>Hüfte </i>'hip'. While such elision cases are relatively rare, the presence of linking morphemes in nominal compounds is a much more frequent phenomenon. Eisenberg (2006) distinguishes between the following linking ele­ments: <i>n (Blumenvase: Blume + n + Vase; </i>'flower vase'), <i>s (Zweifelsfall: Zweifel + s + Fall; </i>'case of doubt'), <i>ns (Glaubensfrage: Glaube + ns + Frage; </i>'question of believe'), <i>e (Pferdewagen: Pferd + e + Wagen; </i>'horse car­riage'), <i>er (Kindergarten: Kind + er + Garten), en (Heldenmut: Held + en + Mut; </i>'hero's cour­age'), <i>es (Siegeswille: Sieg + es + Wille; </i>'will to win'), and <i>ens (Schmerzensschrei: Schmerz + ens + Schrei; </i>'scream of pain').</p></section><section number="3" title="Modeling Compounds in GermaNet"><p><i>GermaNet </i>is a lexical semantic network that is modeled after the Princeton WordNet for Eng­lish. It partitions the lexical space into a set of semantic concepts (modeled by <i>synsets) </i>that are interlinked by semantic relations. A synset is a set of words (called <i>lexical units) </i>where all the words are taken to have the same meaning. There are two types of semantic relations in GermaNet. <i>Conceptual relations </i>hold between two synsets, including hypernymy, part-whole relations, en-tailment, or causation. <i>Lexical relations </i>hold be­tween two individual lexical units.</p><p>To the best of our knowledge, a systematic treatment of compounds is largely absent from monolingual wordnets presently available. The only programmatic approach for how to treat compounds is documented in the final report of the <i>EuroWordNet </i>project (Vossen, 2002) from which the following illustrative example is taken:</p><p>guitar player</p><p>HAS_HYPERONYM player CO_AGENT_INSTRUMENT guitar</p><footnote label="2">Langer (1998) presents a frequency table for German link­ing morphemes and elisions, according to which approxi­mately half of the compounds he investigated contain some kind of linking morpheme or elision.</footnote><p>In this EuroWordNet proposal, compounds such as <i>guitar player </i>are linked via conceptual relations to their component parts. The com­pound as a whole is related via the hypernymy relation to its head constituent <i>(player) </i>and via the bidirectional COROLE relation to its modi­fier constituent <i>(guitar). </i>This CO_ROLE relation is then further specified by the particular themat­ic role realized by the modifier constituent. In short, the EuroWordNet treatment focuses on the semantics of compounds.</p><p>The current proposal of how to treat com­pounds in GermaNet is to some extent more modest in that it focuses on the morphosyntactic structure of compounds and leaves a semantic treatment to future work. A strong requirement for a compounding analysis for GermaNet is that it has to reflect the recursive nature of com­pounding in the case of compounds that have more than two constituent parts such as <i>Kraftfahrzeugsteuer </i>'motor vehicle tax'. The immediate constituents of this compound are <i>Kraftfahrzeug </i>and <i>steuer, </i>with the first constitu­ent then splitting further into <i>Kraft </i>and <i>fahrzeug, </i>etc. (see Figure 1). In order to be able to system­atically link compounds in GermaNet to their constituent parts, compound splitting needs to be applied recursively and has to identify only the immediate constituents at each level of analysis.</p><p>Kraftfahrzeugsteuer</p><p><b>c_modifier N^cjiead</b></p><p>Kraftfahrzeug Steuer</p><doubt alpha="63.2" length="19" tooSmall="False" monospace="0.0">(^modifier/ \^_head</doubt><p>Kraft Fahrzeug</p><doubt alpha="59.1" length="22" tooSmall="False" monospace="0.0">c modifie^/    \c head</doubt><p>fahren Zeug Figure 1. Compounds in GermaNet.</p></section><section number="4" title="Related Work on Compound Splitting"><p>For German, there are a number of morphologi­cal tools available that include compound split­ting, such as GERTWOL (Haapalainen and Ma­jorin, 1994), SMOR (Schmid et al., 2004), ASV Toolbox (Witschel and Biemann, 2005), Bana-naSplit<footnote anchor="3"/> , and Morfessor (Creutz and Lagus, 2005). After an initial evaluation of all publicly available tools, SMOR and ASV Toolbox are used as baseline tools for the present project.</p><footnote label="3">See http://niels.drni.de/s9y/pages/bananasplit.html</footnote><page local="3" global="422"/><p><i>SMOR </i>is a morphological analyzer for Ger­man inflection and productive word formation including composition, which has been devel­oped at the University of Stuttgart. It provides analyses consisting of sequences of morphemes enriched with morphological information, how­ever without grouping them into immediate con­stituents. Furthermore, although SMOR disam-biguates its results to a certain extend, for many compounds there are still several distinct se­quences of morphemes provided.</p><p><i>ASV Toolbox </i>has been developed at the Uni­versity of Leipzig. It comprises several tools for linguistic classification and clustering, amongst them compound splitting, which is included in the tool described as <i>ASV Toolbox Baseforms<footnote anchor="4"/>. </i>The result of the compound analysis identifies all constituent parts of the compound without inter­nal bracketing. It reduces inflected word forms of constituents to their base forms.</p></section><section number="5" title="Compound Splitting Algorithms"><p>Three individual compound splitters are used in the present project: a compound splitter incorpo­rating GermaNet (GN-CS) developed by the au­thors of this paper, a modified version of SMOR (SMOR-CS), and a modified version of the ASV Toolbox compound splitter (ASV-CS).</p><subsection number="5.1" title="Compound     Splitter Incorporating GermaNet (GN-CS)"><p>This compound splitter is especially tailored for determining compounds in GermaNet and their immediate constituents. It uses pattern matching for gathering all potential modifiers and heads of a compound, considering intervening linking morphemes and the elision of word-final charac­ters (as described in section 2). In case the pat­tern matching yields more than one potential modifier-head composition, the correct constitu­ents are verified incorporating the semantic re­source GermaNet and its graph structure. For example, compositions having both constituents in GermaNet are preferred over compositions where only one constituent is an existing entry in GermaNet. Further, more probability is assigned to compositions of simple string concatenation than to compositions showing a linking mor­pheme or the elision of word-final characters.</p><p>The availability of semantic relations, such as part-whole relations, direct or indirect hyper-nymy, or synonymy, is employed as well. Thus, a modifier or head that is semantically related to the compound determines the correct splitting of compounds into its immediate constituents with high probability. The following example illus­trates this. For the compound <i>Flughafengelände </i>'airport area', all relevant parts of the two candi­date parses <i>Flug + Hafengelände </i>and <i>Flughafen + Gelände </i>are existing entries in GermaNet, i.e., existing words. Further, both potential analyses show neither linking morphemes nor the elision of word-final characters. In this case, the usage of GermaNet's semantic relations determines, that <i>Flughafen </i>is a holonym of the compound <i>Flughafengelände, </i>and thus clearly and correctly determines the modifier, resulting in the correct parse <i>Flughafen + Gelände.</i></p><footnote label="4">See http://wortschatz.uni-leipzig.de/~cbiemann/software/ toolbox/Baseforms%20Tool.htm</footnote><p>If there are two different modifier-head com­binations having both their heads as hypernyms of the compound, GN-CS disambiguates the cor­rect splitting by taking into account the hyper-nym's distances<footnote anchor="5"/>. The splitting belonging to the head with the larger hypernym distance is pre-ferred.<footnote anchor="6"/> For example, <i>Nachttischlampe </i>'bedside lamp' has both hypernyms <i>Tischlampe </i>'table lamp' (hypernym distance is 1, i.e., direct hyper-nym) and <i>Lampe </i>'lamp' (hypernym distance is 2, i.e., indirect hypernym). Thus, <i>Nachttischlampe </i>is correctly split into <i>Nachttisch + Lampe.</i></p></subsection><subsection number="5.2" title="Modified SMOR Compound Splitter"><doubt alpha="66.7" length="9" tooSmall="False" monospace="0.0">(SMOR-CS)</doubt><p>To achieve better results in the specific task of determining compounds in GermaNet and their immediate constituents, SMOR's output has been adapted. Some steps, such as the denominaliza-tion of the head constituents or the splitting of all affixes, need to be reverted. Other results, such as the splitting into more than two constituents or the indication of more than one splitting possibil­ity, require further processing. For example, SMOR splits <i>Änderungsanforderung </i>'change request' into <i>ändern + ung + an + fordern + ung. </i>After reverting the denominalization and the separation of prefixes and suffixes, SMOR-CS returns: <i>Änderung </i>+ <i>Anforderung.</i></p><p>For those compounds with several distinct re­sults, it is not trivial to disambiguate the correct splitting. Furthermore, the splitting of com­pounds having more than two constituents, such as <i>Brennstofflagerungsbehälter </i>('fuel storage container'), which is split into <i>brennen + Stoff + lagern + Behälter </i>cannot be used in this form for determining immediate constituents, since the constituents are not grouped.<page local="4" global="423"/></p><footnote label="5">Here, hypernym distance describes the path length be­tween the compound and a direct or indirect hypernym, i.e., a direct hypernym has a hypernym distance of one.</footnote><footnote label="6">Preference of longer hypernym distance may seem coun­terintuitive, but surprisingly turns out to be the correct heu­ristic.</footnote></subsection><subsection number="5.3" title="Modified   ASV   Toolbox Compound"><p><b>Splitter (ASV-CS)</b></p><p>The output of the ASV Toolbox compound split­ter is further processed in order to better fit the needs of the present project. To enhance the reli­ability of the determined constituents, the en­hanced compound splitter ASV-CS searches for entries in GermaNet. If a result consists of more than two constituents, the different bracketing alternatives need to be verified. This is done by incorporating GermaNet's graph structure in the same way as for GN-CS (see section 5.1).</p></subsection></section><section number="6" title="Combination of Compound Splitters"><p>It has been shown for various NLP tasks, such as part-of-speech tagging (van Halteren et al., 2001) or word sense disambiguation (Florian and Yarowsky, 2002), that multiple classifier systems outperform single decision systems. Further, the performance of such methods is usually better the more diverse the individual systems are (Polikar, 2006). Thus, having three classifiers<footnote anchor="7"/>(compound splitters) available that produce di­verse results, the application of a combined method seems reasonable. As the compound splitters in the present project each return exactly one decision, the range of applicable combina­tion algorithms is restricted. In the following subsection, the application of <i>majority voting </i>and <i>weighted majority voting </i>is described. Further, a combined algorithm, which is developed by the authors of this paper, is presented.</p><subsection number="6.1" title="Majority Voting (MV) and Weighted Majority Voting (WMV)"><p>In majority voting, equal weight is given to all compound splitters when voting for a result (i.e., a splitting of a compound into its immediate con­stituents). The votes from all compound splitters are summed up and the result with the highest number of votes is selected. In case a compound splitter does not return an analysis, it is disre­garded, while the other two compound splitters vote for the final result.<footnote anchor="8"/> In weighted majority voting, individual compound splitters are as­signed different weights in such a way that the combination of weights minimizes errors.<footnote anchor="9"/></p><footnote label="7">The task of compound splitting is, in a strict sense, not a classification task, because there is no predefined result set, such as a tagset for part-of-speech tagging. The results of the compound splitters are rather variable and, from a tech­nical point of view, describe arbitrary content (although describing the splitting of a compound into its immediate constituents).</footnote></subsection><subsection number="6.2" title="Combined Hybrid Compound Splitter"><doubt alpha="57.1" length="7" tooSmall="False" monospace="0.0">(CH-CS)</doubt><p>In order to further increase performance, we cre­ated a hybrid combined compound splitter that takes into account all knowledge provided by the individual compound splitters, but that also takes into account some domain knowledge about German derivation morphology and compound­ing. One of the frequent mistakes made is to treat words like <i>Gutherzigkeit<footnote anchor="10"/> </i>'kindheartedness' or <i>Teilhaberschaft</i><footnote anchor="11"/> 'partnership' as compounds, while in reality these are complex nouns formed by derivation morphology. The hybrid model therefore incorporates knowledge about deriva­tion morphology and filters out such erroneously marked compounds. As will be shown in the evaluation section, the hybrid model outperforms all individual compound splitters as well as the other combined compound splitters in all tasks described in section 7.</p></subsection></section><section number="7" title="Evaluation"><p>The automatic predictions of compounds and their immediate constituents are manually veri­fied. The order of the manual verification is in the order of the IDs of the lexical units, which is actually randomly concerning the nouns them­selves. For the purpose of evaluation, 68743<footnote anchor="12"/>nouns were chosen, of which 42191 (61.37%) are compounds and 26552 (38.63%) are not. The evaluation is fourfold: (i) section 7.1 evaluates how many compounds are correctly identified, (ii) section 7.2 evaluates how many predicted compounds are split at the correct position, (iii) how many compounds are correctly predicted</p><footnote label="8">In case of a tie, giving priority to SMOR-CS turned out to be the best strategy.</footnote><footnote label="9">Experimenting with several weighting combinations re­sulted in giving weight 2.0 to SMOR-CS, 0.9 to GN-CS,</footnote><p>and 0.8 to ASV-CS. This adjustment helps in cases where both GN-CS and ASV-CS agree on an erroneous analysis.</p><footnote label="10">SMOR-CS treats Gutherzigkeit erroneously as a com­pound, although it is derived from the adjective gutherzig with the derivation suffix -keit.</footnote><footnote label="11"> Teilhaberschaft is derived from the noun Teilhaber with the derivation suffix -schaft.</footnote><footnote label="12">Altogether, there are 93407 nouns in GermaNet. Note that all foreign words and named entities are disregarded in this evaluation.</footnote><page local="5" global="424"/><p>regarding the word forms of their immediate constituents is evaluated in section 7.3, and, fi­nally, (iv) there is an error analysis in section 7.4.</p><subsection number="7.1" title="Identification of Compounds"><p>The first part of the evaluation concerns the pre­diction whether a noun in GermaNet is a com­pound or not. Table 1 lists all <i>true positives </i>(TP; correctly identified compounds), <i>false positives </i>(FP; erroneously identified as a compound), <i>true negatives </i>(TN; correctly identified as no com­pound), and <i>false negatives </i>(FN; erroneously not identified as a compound). The numbers are sep­arately calculated for the individual algorithms and for the combined algorithms.</p><p>The reason for MV and WMV performing alike in this task of identifying compounds is that in case a compound splitter does not return an analysis, it is disregarded. This means that, if at least one compound splitter returns a result, both MV and WMV decide that this noun is a com­pound regardless of any weighting.</p><p>There are remarkable improvements especially in the numbers of true positives and false nega­tives of the combined algorithms compared to the individual ones. The reason for these remark­able differences is obvious: the individual split­ting algorithms are very heterogeneous, which leads to an improved overall coverage. Table 2 shows the calculated percentages for accuracy, precision, and recall of the task of identifying compounds.</p><p>Highest accuracy and best recall are achieved by CH-CS, whereas ASV-CS and SMOR-CS yield highest precision. The values in this section (Tables 1 and 2) are gathered with the aim of identifying if a noun in GermaNet is a compound or not. The correctness of the splitting into two constituents is considered in the following sec­tions.</p></subsection><subsection number="7.2" title="Predicting Immediate Splitting Position"><p>This part of the evaluation regards the splitting position. It is evaluated for all 42191 compounds whether the predicted position at which the algo­rithms split the compounds into two constituents is correct. An obvious error is, e.g., the splitting of <i>Tiefkühltruhe </i>'deep-freezer' into <i>tief + Kühl­truhe </i>instead of <i>tiefkühlen + Truhe. </i>An example of an erroneous splitting that is not as obvious is the splitting of <i>Muskelshirt </i>'muscle shirt' into <i>Muskel + Hirt </i>instead of <i>Muskel + Shirt. </i>In con­trast, the predicted position of the splitting <i>Bund­faltenhose </i>'pleated pants' into <i>Bundfalten + Hose </i>(instead of <i>Bundfalte + Hose) </i>is correct, although this example reveals a wrong inflection of the modifier. The evaluation results are pre­sented in Table 3; where the accuracy specifies the number of correctly predicted splitting posi­tions divided by the total number of compounds.</p><p>For the task of predicting the immediate split­ting position again all combined algorithms out­perform the individual compound splitters.</p></subsection><subsection number="7.3" title="Prediction of Immediate Constituents"><p>This section evaluates the correctness of the en­tire prediction of two immediate constituents, including word class and inflection. The predict­ed constituents for all 42191 compounds are ana­lyzed and the results listed in Table 4. The evalu­ation takes into account, that for some com­pounds, there is more than one composition cor­rect. For <i>Nachtspeicherheizung </i>'night storage heater', e.g., two internal groupings are semanti-cally correct: <i>Nacht + Speicherheizung </i>and <i>Nachtspeicher + Heizung. </i>In other compounds, two word classes are possible for the modifier. For example, <i>Spielecke </i>'kid's corner' might be composed of <i>Spiel + Ecke </i>or <i>spielen + Ecke.</i></p><table caption="Table 1: Identification of Compounds" 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></tr><tr class="row"><td class="cell"></td><td class="cell"><p><i>Algorithm TP</i></p></td><td class="cell"><p><i>FP</i></p></td><td class="cell"><p><i>TN FN</i></p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>GN-CS 38489</p></td><td class="cell"><p>1559</p></td><td class="cell"><p>24993 3702</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>SMOR-CS 33765</p></td><td class="cell"><p>544</p></td><td class="cell"><p>26008 8426</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>ASV-CS 36356</p></td><td class="cell"><p>555</p></td><td class="cell"><p>25997 5835</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>MV &amp; WMV 39675</p></td><td class="cell"><p>1806</p></td><td class="cell"><p>24746 2516</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>CH-CS 41894</p></td><td class="cell"><p>1974</p></td><td class="cell"><p>24578 297</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></tr></table><table caption="Table 3: Predicting Immediate Splitting Position" 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><i>Algorithm</i></p></td><td class="cell"><p><i>Correct position</i></p></td><td class="cell"><p><i>Erroneous position</i></p></td><td class="cell"><p><i>Accuracy</i></p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>GN-CS</p></td><td class="cell"><p>37779</p></td><td class="cell"><p>4411</p></td><td class="cell"><p>89.54%</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>SMOR-CS</p></td><td class="cell"><p>32863</p></td><td class="cell"><p>9326</p></td><td class="cell"><p>77.89%</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>ASV-CS</p></td><td class="cell"><p>35407</p></td><td class="cell"><p>6783</p></td><td class="cell"><p>83.92%</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>MV</p></td><td class="cell"><p>38548</p></td><td class="cell"><p>3636</p></td><td class="cell"><p>91.38%</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>WMV</p></td><td class="cell"><p>38688</p></td><td class="cell"><p>3496</p></td><td class="cell"><p>91.71%</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>CH-CS</p></td><td class="cell"><p>40010</p></td><td class="cell"><p>2181</p></td><td class="cell"><p>94.83%</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: Accuracy, Precision, and Recall of Identifying Compounds" 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><i>Algorithm</i></p></td><td class="cell"><p><i>Accuracy</i></p></td><td class="cell"><p><i>Precision</i></p></td><td class="cell"><p><i>Recall</i></p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>GN-CS</p></td><td class="cell"><p>92.34%</p></td><td class="cell"><p>96.11%</p></td><td class="cell"><p>91.23%</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>SMOR-CS</p></td><td class="cell"><p>86.95%</p></td><td class="cell"><p>98.41%</p></td><td class="cell"><p>80.03%</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>ASV-CS</p></td><td class="cell"><p>90.70%</p></td><td class="cell"><p>98.50%</p></td><td class="cell"><p>86.17%</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>MV &amp; WMV</p></td><td class="cell"><p>93.71%</p></td><td class="cell"><p>95.65%</p></td><td class="cell"><p>94.04%</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>CH-CS</p></td><td class="cell"><p>96.70%</p></td><td class="cell"><p>95.50%</p></td><td class="cell"><p>99.30%</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><page local="6" global="425"/><p>Table 4 reveals that all combined compound splitters outperform the individual compound splitters in the main task of the present project, i.e., in determining immediate constituents of compounds in GermaNet. The best overall result with an accuracy of 92.42% is achieved by the hybrid combined compound splitter CH-CS.</p></subsection><subsection number="7.4" title="Error Analysis"><p>To distinguish different cases that cause errone­ous predictions of the immediate constituents, the following error types were identified. The occurrences of these error types - presented in Table 5 - are gathered for the combined algo­rithm CH-CS only as this error classification is done in a manual verification step.</p><p><i>• Position: </i>The proposed splitting position is wrong, e.g., <i>Eislaufbahn </i>'ice rink' is split into <i>Eis + Laufbahn </i>instead of <i>Eislauf + Bahn.</i></p><p><i>• Not parsed: </i>Some compounds are recognized but not parsed. For example, a compound such as <i>Kreuzschlitzschraubenzieher </i>,Philips screwdriver', consisting of four parts, is rec­ognized as a compound, but not grouped into its immediate constituents.</p><p><i>• Wrong lemma: </i>For some predictions, the lemmatization of the modifier is erroneous. For example, the immediate lemmatized con­stituents of <i>Hühnerleiter </i>'chicken ladder' are <i>Huhn </i>and <i>Leiter, </i>but CH-CS splits the com­pound into <i>Hühner + Leiter </i>without lemma-tizing the modifier.</p><p><i>• Word class: </i>The modifier has been assigned a wrong word class. Two different subcases are distinguished:</p><p>1. The proposed word does not exist. For ex­ample, <i>Mischanlage </i>'mixing plant' is er­roneously split into <i>Misch + Anlage, </i>but the modifier needs to be the verb <i>mischen, </i>because a noun like <i>Misch </i>does not exist.</p><p>2. The proposed word (class) has a wrong reading, e.g., the splitting of <i>Allesschnei­der </i>'slicing maschine' into <i>All + Schnei­der </i>instead of <i>alles + Schneider </i>reveals a wrong reading of the modifier.</p><p><i>•  False negatives: </i>Those compounds that are erroneously not identified as a compound.</p><p>Two (obvious) causes of errors are identified in Table 6: bound morphemes and missing en­tries in GermaNet. Bound morphemes such as <i>Him- </i>in <i>Himbeere </i>'raspberry' (cf. section 2 above) are a common source of error because the algorithm cannot reliably identify such words. Second, if either the modifier or the head is not in GermaNet, the algorithm may propose a wrong splitting. For example, the correct split­ting of <i>Feincordhose </i>'narrow wale corduroy pants' is <i>Feincord + Hose, </i>but as <i>Feincord </i>is not in GermaNet, the algorithm erroneously proposes <i>fein + Cordhose </i>as those two constituents are entries in GermaNet.</p><p>A third error type is identified for false posi­tives - actually for 35.3% of all false positives (696 of 1974): Words like <i>Bausparen </i>'building society savings' or <i>Zusammenprall </i>'collision' are frequently treated as compounds, while these are nouns derived from compound verbs.</p></subsection></section><section number="8" title="Conclusion and Future Work"><p>Existing tools for splitting compounds were adapted to overcome issues with determining immediate constituents of compounds. Combina-tory algorithms using three heterogeneous kinds of compound splitters are developed to achieve better results. As the combined compound splitting algorithms all outperform the individual compound splitters, the overall combined result should improve further, including even more in­dividual compound splitters.<page local="7" global="426"/> The best overall result with an accuracy of 92.42% is achieved by a hybrid combined compound splitter that takes into account all knowledge provided by the indi­vidual compound splitters, and in addition some domain knowledge about German derivation morphology and compounding.</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></tr><tr class="row"><td class="cell"></td><td class="cell"><p><i>Algorithm</i></p></td><td class="cell"><p><i>Correct constituents</i></p></td><td class="cell"><p><i>Erroneous constituents</i></p></td><td class="cell"><p><i>Accuracy</i></p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>GN-CS</p></td><td class="cell"><p>32738</p></td><td class="cell"><p>9449</p></td><td class="cell"><p>77.60%</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>SMOR-CS</p></td><td class="cell"><p>31757</p></td><td class="cell"><p>10432</p></td><td class="cell"><p>75.27%</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>ASV-CS</p></td><td class="cell"><p>31621</p></td><td class="cell"><p>10568</p></td><td class="cell"><p>74.95%</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>MV</p></td><td class="cell"><p>33349</p></td><td class="cell"><p>8832</p></td><td class="cell"><p>79.06%</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>WMV</p></td><td class="cell"><p>33176</p></td><td class="cell"><p>9005</p></td><td class="cell"><p>78.65%</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>CH-CS</p></td><td class="cell"><p>38994</p></td><td class="cell"><p>3197</p></td><td class="cell"><p>92.42%</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>Table 4: Prediction of Immediate Constituents</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 5: Occurrences of Different Error Types" 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><i>Error type</i></p></td><td class="cell"><p><i>CH-CS</i></p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>Position</p></td><td class="cell"><p>384 (12.01%)</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>Not parsed</p></td><td class="cell"><p>1490 (46.60%)</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>Wrong lemma</p></td><td class="cell"><p>207 (6.47%)</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>Word class 1</p></td><td class="cell"><p>325 (10.17%)</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>Word class 2</p></td><td class="cell"><p>311 (9.73%)</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>False negatives</p></td><td class="cell"><p>297 (9.29%)</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>Other</p></td><td class="cell"><p>183 (5.72%)</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>Total errors</p></td><td class="cell"><p>3197</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><table caption="Table 6: Causes of Errors" 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><i>Error type</i></p></td><td class="cell"><p><i>Total</i></p></td><td class="cell"><p><i>Bound morpheme</i></p></td><td class="cell"><p><i>No entry in GermaNet</i></p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>Position</p></td><td class="cell"><p>384</p></td><td class="cell"><p>18 (4.7%)</p></td><td class="cell"><p>280 (72.9%)</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>Not parsed</p></td><td class="cell"><p>1490</p></td><td class="cell"><p>98 (6.6%)</p></td><td class="cell"><p>1061 (71.2%)</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>Wrong lemma</p></td><td class="cell"><p>207</p></td><td class="cell"><p>7 (3.4%)</p></td><td class="cell"><p>150 (72.5%)</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>Word class 1</p></td><td class="cell"><p>325</p></td><td class="cell"><p>112 (34.5%)</p></td><td class="cell"><p>226 (69.5%)</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>Word class 2</p></td><td class="cell"><p>311</p></td><td class="cell"><p>14 (4.5%)</p></td><td class="cell"><p>87 (28.0%)</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>FN</p></td><td class="cell"><p>297</p></td><td class="cell"><p>15 (5.2%)</p></td><td class="cell"><p>153 (51.5%)</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>Other</p></td><td class="cell"><p>183</p></td><td class="cell"><p>2 (1.1%)</p></td><td class="cell"><p>23 (12.6%)</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>Total errors</p></td><td class="cell"><p>3197</p></td><td class="cell"><p>266 (8.3%)</p></td><td class="cell"><p>1980 (61.9%)</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><p>There are two obvious problems with the used individual compound splitters. First, lemmatized forms are never generated by GN-CS. Extending GN-CS with a lemmatizer to determine base forms can enhance this drawback. Second, the immediate constituents of compounds consisting of more than two or three constituents are not determined by SMOR-CS and ASV-CS, respec­tively. This issue can be improved through bracketing those compounds by ASV-CS and</p><p>SMOR-CS.</p><p>In future work, we plan to automatically pre­dict compound-internal relations between the now determined immediate constituents by using GermaNet's relations. This would also mean that the immediate compound constituents would have to be automatically disambiguated. Further, an automatic extension of GermaNet with com­pounds by using statistical information of exist­ing compounds in GermaNet is envisioned.</p></section><section title="Acknowledgments"><p>The research reported in this paper was jointly funded by the SFB 833 grant of the DFG and by the CLARIN-D grant of the BMBF.</p><p>We would like to thank our research assistant Sarah Schulz for her help with the evaluation reported in Section 7. Special thanks go to our GermaNet colleague Reinhild Barkey for exten­sive discussions on the syntax and semantics of compounds and on their modeling in GermaNet.</p></section><references><p>Marco Baroni, Johannes Matiasek, and Harald Trost. 2002. Predicting the Components of German No­minal Compounds. <i>Frank van Harmelen (eds.), Proceedings of the 15th European Conference on Artificial Intelligence (ECAI), </i>Amsterdam: IOS Press, 470-474.</p><p>Mathias Creutz and Krista Lagus. 2005. Unsupervised Morpheme Segmentation and Morphology Induc­tion from Text Corpora Using Morfessor 1.0. <i>Pub­lications in Computer and Information Science,</i></p><p>Report A81. Helsinki University of Technology Helsinki, Finland.</p><p>Peter Eisenberg. 2006. <i>Das Wort - Grundriss der deutschen Grammatik. </i>3rd edition, Verlag J. B. Melzer, Stuttgart/Weimar, Germany.</p><p>Christiane Fellbaum (eds.). 1998. <i>WordNet - An Elec­tronic Lexical Database. </i>The MIT Press.</p><p>Radu Florian and David Yarowsky. 2002. Modeling consensus: classifier combination for word sense disambiguation. In <i>Proceedings of the ACL-02 conference on Empirical methods in natural lan­guage processing (EMNLP '02), </i>Vol. 10. Associa­tion for Computational Linguistics, Stroudsburg,</p><p>PA, USA, 25-32.</p><p>Mariikka Haapalainen and Ari Majorin. 1994. <i>GERTWOL: Ein System zur automatischen Wort­formerkennung Deutscher Wörter. </i>Technical re­port, Lingsoft Inc. https://files.ifi.uzh.ch/cl/volk/ LexMorphVorl/Lexikon04.Gertwol.html</p><p>Verena Henrich and Erhard Hinrichs. 2010. GernEdiT - The GermaNet Editing Tool. In <i>Proceedings of the Seventh Conference on International Language Resources and Evaluation (LREC 2010), </i>main con­ference. Valletta, Malta.</p><p>Claudia Kunze and Lothar Lemnitzer. 2002. Ger-maNet - representation, visualization, application. In <i>Proceedings of LREC 2002, </i>main conference, Vol V. pp. 1485-1491.</p><p>Stefan Langer. 1998. Zur Morphologie und Semantik von Nominalkomposita. <i>Tagungsband der 4. Kon­ferenz zur Verarbeitung natürlicher Sprache, </i>KONVENS, pp. 83-97.</p><p>Robi Polikar. 2006. Ensemble based systems in deci­sion making. In <i>IEEE Circuits and Systems Maga­zine, </i>16(3):21-45.</p><p>Helmut Schmid, Arne Fitschen, and Ulrich Heid.</p><p>2004. SMOR: A German Computational Morphol­ogy Covering Derivation, Composition, and Inflec­tion. In <i>Proceedings of the IVth International Con­ference on Language Resources and Evaluation (LREC 2004), </i>p. 1263-1266, Lisbon, Portugal.</p><p>Hans van Halteren, Walter Daelemans and Jakub Za-vrel. 2001. Improving accuracy in word class tag­ging through the combination of machine learning systems. In <i>Compuational Linguistics, </i>27, 2, 199­229.</p><p>Piek Vossen. 2002. EuroWordNet General Document. EuroWordNet Project LE2-4003 &amp; LE4-8328 re­port, Version 3, Final, University of Amsterdam.</p><p>Hans F. Witschel and Chris Biemann. 2005. Rigorous dimensionality reduction through linguistically mo­tivated feature selection for text categorisation. In <i>Proceedings of NODALIDA 2005, </i>Joensuu, Fin­land.</p></references></body></article>