<?xml version="1.0" standalone="yes"?>
<Paper uid="P98-2123">
  <Title>A Freely Available Morphological Analyzer, Disambiguator and Context Sensitive Lemmatizer for German</Title>
  <Section position="2" start_page="0" end_page="744" type="metho">
    <SectionTitle>
1 The Morphology System
</SectionTitle>
    <Paragraph position="0"> Since German is a highly inflectional language, the morphological algorithms used in Morphy are rather complex and can not be described here in detail (see Lezius, 1996). In essence, Morphy is a computer implementation of the morphological system described in the Duden grammar (Drosdowsky, 1984).</Paragraph>
    <Paragraph position="1"> An overview on other German morphology systems, namely GERTWOL, LA-Morph, Morph, Morphix, Morphy, MPRO, PC-Kimmo  and Plain, is given in the documentation for the Morpholympics (Hausser, 1996). The Morpholympics were an attempt to compare and evaluate morphology systems in a standardized competition. Since then, many of the systems have been further developed. The version of Morphy as described here is a new release. Improvements over the old version include an integrated part-of-speech tagger, a context-sensitive lemmatizer, a 2.5 times larger lexicon and more user-friendliness through an interactive Windows-environment. null The following subsections describe the three submodules of the morphological analyzer.</Paragraph>
    <Paragraph position="2"> These are the lexical system, the generation module and the analysis module.</Paragraph>
    <Section position="1" start_page="743" end_page="743" type="sub_section">
      <SectionTitle>
1.1 Lexical System
</SectionTitle>
      <Paragraph position="0"> The lexicon of Morphy is very compact as it only stores the base form for each word together with its inflection class. Therefore, the complete morphological information for 324,000 word forms takes less than 2 Megabytes of disk space.</Paragraph>
      <Paragraph position="1"> In comparison, the text representation of the same lexicon, which can be generated via Morphy's export facility, requires 125 MB when full morphological descriptions are given.</Paragraph>
      <Paragraph position="2"> Since the lexical system has been specifically designed to allow a user-friendly extension of the lexicon, new words can be added easily. To our knowledge, Morphy is the only morphology system for German whose lexicon can be expanded by users who have no specialist knowledge. When entering a new word, the user is asked the minimal number of questions necessary to infer the grammatical features of the new word and which any native speaker of German should be able to answer.</Paragraph>
    </Section>
    <Section position="2" start_page="743" end_page="743" type="sub_section">
      <SectionTitle>
1.2 Generation
</SectionTitle>
      <Paragraph position="0"> Starting from the root form of a word and its inflection type as stored in the lexicon, the generation system produces all inflected forms.</Paragraph>
      <Paragraph position="1"> Morphy's generation algorithms were designed with the aim of producing 100% correct output.</Paragraph>
      <Paragraph position="2"> Among other morphological characteristics, the algorithms consider vowel mutation (Haus -H/iuser), shift between B and ss (FaB - Fasser), eomission (segeln - segle), infixation of infinitive markers (weggehen - wegzugehen), as well as pre- and infixation of markers of participles (gehen - g.egangen; weggehen - wegg.egangen).</Paragraph>
    </Section>
    <Section position="3" start_page="743" end_page="744" type="sub_section">
      <SectionTitle>
1.3 Analysis
</SectionTitle>
      <Paragraph position="0"> For each word form of a text, the analysis system determines its root, part of speech, and - if appropriate - its gender, case, number, person, tense, and comparative degree. It also segments compound nouns using a longest-matching rule which works from right to left and takes linking letters into account. To compound German nouns is not trivial: it can involve base forms and/or inflected forms (e.g. Haus-meister but H~iuser-meer); in some cases the compounding is morphologically ambiguous (e.g. Stau-becken means water reservoir, but Staub-ecken means dust corners); and the linking letters e and s are not always determined phonologically, but in some cases simply occur by convention (e.g.</Paragraph>
      <Paragraph position="1"> Schwein-e-bauch but Schwein-s-blase and Schwein-kram).</Paragraph>
      <Paragraph position="2"> Since the analysis system treats each word separately, ambiguities can not be resolved at this stage. For ambiguous word forms, all possible lemmata and their morphological descriptions are given (see Table 1 for the example Winde). If a word form can not be recognized, its part of speech is predicted by a guesser which makes use of statistical data derived from German suffix frequencies (Rapp, 1996).</Paragraph>
      <Paragraph position="3"> morphological description lemma  Morphy's algorithm for analysis is motivated by linguistic considerations. When analyzing a word form, Morphy first builds up a list of possible roots by cutting off all possible prefixes and suffixes and reverses the process of vowel mutation if umlauts are found (shifts between B  and ss are treated analogously). Each root is looked up in the lexicon, and - if found - all possible inflected forms are generated. Only those roots which lead to an inflected form identical to the original word form are selected (Lezius, 1996).</Paragraph>
      <Paragraph position="4"> Naturally, this procedure is much slower than a simple algorithm for the lookup of word forms in a full form lexicon. It results in an analysis speed of about 300 word forms per second on a fast PC, compared to many thousands using a full form lexicon. However, there are also advantages: First, as mentioned above, the lexicon can be kept very small, which is an important consideration for a PC-based system intended for Internet-distribution. More importantly, the processing of German compound nouns and the implementation of derivation rules - although only partially completed at this stage - fits better into this concept. For the processing of very large corpora under UNIX, we have implemented a lookup algorithm which operates on the Morphy-generated full form lexicon.</Paragraph>
      <Paragraph position="5"> The coverage of the current version of Morphy was evaluated with the same test corpus that had been used at the Morpholympics. This corpus comprises about 7.800 word forms in total and consists of two political speeches, a fragment of the LIMAS-corpus, and a list of special word forms. The present version of Morphy recognized 94.3%, 98.4%, 96.2%, and 88.9% of the word forms respectively. The corresponding values for the old version of Morphy, with a 2.5 times smaller lexicon, had been 89.2%, 95.9%, 86.9%, and 75.8%.</Paragraph>
    </Section>
  </Section>
  <Section position="3" start_page="744" end_page="745" type="metho">
    <SectionTitle>
2 The Disambiguator
</SectionTitle>
    <Paragraph position="0"> Since the morphology system only looks at isolated word forms, words with more than one reading can not be disambiguated. This is done by the disambiguator or tagger, which takes context into account by considering the conditional probabilities of tag sequences. For example, in the sentence &amp;quot;he opens the can&amp;quot; the verbreading of can may be ruled out because a verb can not follow an article.</Paragraph>
    <Paragraph position="1"> After the success of statistical part-of-speech taggers for English, there have been quite a few attempts to apply the same methods to German.</Paragraph>
    <Paragraph position="2"> Lezius, Rapp &amp; Wettler (1996) give an overview on some German tagging projects. Although we considered a number of algorithms, we decided to use the trigram algorithm described by Church (1988) for tagging. It is simple, fast, robust, and - among the statistical taggers - still more or less unsurpassed in terms of accuracy.</Paragraph>
    <Paragraph position="3"> Conceptually, the Church-algorithm works as follows: For each sentence of a text, it generates all possible assignments of part-of-speech tags to words. It then selects that assignment which optimizes the product of the lexical and contextual probabilities. The lexical probability for word N is the probability of observing part of speech X given the (possibly ambiguous) word N. The contextual probability for tag Z is the probability of observing part of speech Z given the preceding two parts of speech X and Y. It is estimated by dividing the trigram frequency XYZ by the bigram frequency XY. In practice, computational limitations do not allow the enumeration of all possible assignments for long sentences, and smoothing is required for infrequent events. This is described in more detail in the original publication (Church, 1988).</Paragraph>
    <Paragraph position="4"> Although more sophisticated algorithms for unsupervised learning - which can be trained on plain text instead on manually tagged corpora are well established (see e.g. Merialdo, 1994), we decided not to use them. The main reason is that with large tag sets, the sparse-data-problem can become so severe that unsupervised training easily ends up in local minima, which can lead to poor results without any indication to the user.</Paragraph>
    <Paragraph position="5"> More recently, in contrast to the statistical taggers, rule-based tagging algorithms have been suggested which were shown to reduce the error rate significantly (Samuelsson &amp; Voutilainen, 1997). We consider this a promising approach and have started to develop such a system for German with the intention of later inclusion into Morphy.</Paragraph>
    <Paragraph position="6"> The tag set of Morphy's tagger is based on the feature system of the morphological analyzer.</Paragraph>
    <Paragraph position="7"> However, some features were discarded for tagging. For example, the tense of verbs is not considered. This results in a set of about 1000 different tags. A fragment of 20,000 words from the Frankfurter Rundschau Corpus, which we have been collecting since 1992, was tagged with this tag set by manually selecting the correct choice from the set of possibilities generated  by the morphological analyzer. In the following we refer to this corpus as the training corpus. Of all possible tags, only 456 actually occurred in the training corpus. The average ambiguity rate was 5.4 tags per word form.</Paragraph>
    <Paragraph position="8"> The performance of our tagger was evaluated by running it on a 5000-word test sample of the Frankfurter Rundschau-Corpus which was distinct from the training text. We also tagged the test sample manually and compared the results.</Paragraph>
    <Paragraph position="9"> 84.7% of the tags were correctly tagged. Although this result may seem poor at first glance, it should be noted that the large tag sets have many fine distinctions which lead to a high error rate. If a tag set does not have these distinctions, the accuracy improves significantly. In order to show this, in another experiment we mapped our large tag set to a smaller set of 51 tags, which is comparable to the tag set used in the Brown Corpus (Greene &amp; Rubin, 1971). As a result, the average ambiguity rate per word decreased from 5.4 to 1.6, and the accuracy improved to 95.9%, which is similar to the accuracy rates reported for statistical taggers with small tag sets in various other languages. Table 2 shows a tagging example for the large and the small tag set.</Paragraph>
  </Section>
  <Section position="4" start_page="745" end_page="746" type="metho">
    <SectionTitle>
3 The Lemmatizer
</SectionTitle>
    <Paragraph position="0"> For lemmatization (the reduction to base form), the integrated design of Morphy turned out to be advantageous. In the first step, the morphologymodule delivers all possible lemmata for each word form. Secondly, the tagger determines the grammatical categories of the word forms. If, for any of the lemmata, the inflected form corresponding to the word form in the text does not agree with this grammatical category, the respective lemma is discarded. For example, in the sentence &amp;quot;ich meine meine Frau&amp;quot; (&amp;quot;I mean my wife&amp;quot;), the assignment of the two middle words to the verb meinen and the possessive pronoun mein is not clear to the morphology system.</Paragraph>
    <Paragraph position="1"> However, since the tagger assigns the tag sequence &amp;quot;pronoun verb pronoun noun&amp;quot; to this sentence, it can be concluded that the first occurrence of meine must refer to the verb meinen and the second to the pronoun mein.</Paragraph>
    <Paragraph position="2"> Unfortunately, this may not always work as well as in this example. One reason is that there may be semantic ambiguities which can not be resolved by syntactic considerations. Another is that the syntactic information delivered by the tagger may not be fine grained enough to resolve all syntactic ambiguities, l Do we need the fine grained distinctions of the large tag set to resolve ambiguities, or does the rough information from the small tag set suffice? To address these questions, we performed an evaluation using another test sample from the Frankfurter Rundschau-Corpus. null We found that - according to the Morphy lexicon- of all 9,893 word forms in the sample, 9,198 (93.0%) had an unambiguous lemma. Of the remaining 695 word forms, 667 had two possible lemmata and 28 were threefold ambiguous (Table 3 gives some examples). Using the large tag set, 616 out of the 695 ambiguous word forms were correctly lemmatized (88.6%). The corresponding figures for the small tag set were slightly better: 625 out of 695 ambiguities were resolved correctly (89.9%). When the error-rate is related to the total number of word forms in the text, the accuracy is 99.2% for the large and 99.3% for the small tag set.</Paragraph>
    <Paragraph position="3"> The better performance when using the small tag set is somewhat surprising since there are a few cases of ambiguities in the test corpus which can only be resolved by the large tag set but not by the small tag set. For example, since the small tag set does not consider a noun's case, gender, and number, it can not decide whether Filmen is derived from der Film (&amp;quot;the film&amp;quot;) or from das Filmen (&amp;quot;the filming&amp;quot;). On the other hand, as shown in the previous section, the tagging accuracy is much better for the small tag set, which is an advantage in lemmatization and obviously compensates for the lack of detail.</Paragraph>
    <Paragraph position="4"> I For example the verb fuhren can be either a subjunctive form offahren (&amp;quot;to drive&amp;quot;) or a regular form offiihren (&amp;quot;to lead&amp;quot;). Since neither the large nor the small tag set consider mood, this ambiguity can not be resolved.</Paragraph>
    <Paragraph position="5">  However, we believe that with future improvements in tagging accuracy lemmatization based on the large tag set will eventually be better.</Paragraph>
    <Paragraph position="6"> Nevertheless, the current implementation of the lemmatizer gives the user the choice of selecting between either tag set.</Paragraph>
    <Paragraph position="7">  In this paper, a freely available integrated tool for German morphological analysis, part-of-speech tagging and context sensitive lemmatization was introduced. The morphological analyzer is based on the standard Duden grammar and provides wide coverage due to a lexicon of 324,000 word forms and the ability to process compound nouns at runtime. It gives for each word form of a text all possible lemmata and morphological descriptions. The ambiguities of the morphological descriptions are resolved by the tagger, which provides about 85% accuracy for the large and 96% accuracy for the small tag set. The lemmatizer uses the output of the tagger to disambiguate word forms with more than one possible lemma. It achieves an overall accuracy of about 99.3%.</Paragraph>
  </Section>
class="xml-element"></Paper>