<?xml version="1.0" standalone="yes"?>
<Paper uid="P06-1084">
  <Title>Sydney, July 2006. c(c)2006 Association for Computational Linguistics An Unsupervised Morpheme-Based HMM for Hebrew Morphological Disambiguation</Title>
  <Section position="4" start_page="665" end_page="666" type="intro">
    <SectionTitle>
2 Hebrew and Arabic Tagging -
</SectionTitle>
    <Paragraph position="0"/>
    <Section position="1" start_page="665" end_page="666" type="sub_section">
      <SectionTitle>
Previous Work
</SectionTitle>
      <Paragraph position="0"> Several works have dealt with Hebrew tagging in the past decade. In Hebrew, morphological analysis requires complex processing according to the rules of Hebrew word formation. The task of a morphological analyzer is to produce all possible analyses for a given word. Recent analyzers provide good performance and documentation of this process (Yona and Wintner, 2005; Segal, 2000).</Paragraph>
      <Paragraph position="1"> Morphological analyzers rely on a dictionary, and their performance is, therefore, impacted by the occurrence of unknown words. The task of a morphological disambiguation system is to pick the most likely analysis produced by an analyzer in the context of a full sentence.</Paragraph>
      <Paragraph position="2"> Levinger et al. (1995) developed a context-free method in order to acquire the morpho-lexical probabilities, from an untagged corpus. Their method handles the data sparseness problem by using a set of similar words for each word, built according to a set of rules. The rules produce variations of the morphological properties of the word analyses. Their tests indicate an accuracy of about 88% for context-free analysis selection based on the approximated analysis distribution. In tests we reproduced on a larger data set (30K tagged words), the accuracy is only 78.2%. In order to improve the results, the authors recommend merging their method together with other morphological disambiguation methods - which is the approach we pursue in this work.</Paragraph>
      <Paragraph position="3"> Levinger's morphological disambiguation system (Levinger, 1992) combines the above approximated probabilities with an expert system, based on a manual set of 16 syntactic constraints . In the first phase, the expert system is applied, dis24-86). null ambiguating 35% of the ambiguous words with an accuracy of 99.6%. In order to increase the applicability of the disambiguation, approximated probabilities are used for words that were not disambiguated in the first stage. Finally, the expert system is used again over the new probabilities that were set in the previous stage. Levinger reports an accuracy of about 94% for disambiguation of 85% of the words in the text (overall 80% disambiguation). The system was also applied to prune out the least likely analyses in a corpus but without, necessarily, selecting a single analysis for each word. For this task, an accuracy of 94% was reported while reducing 92% of the ambiguous analyses. null Carmel and Maarek (1999) use the fact that on average 45% of the Hebrew words are unambiguous, to rank analyses, based on the number of disambiguated occurrences in the text, normalized by the total number of occurrences for each word. Their application - indexing for an information retrieval system - does not require all of the morphological attributes but only the lemma and the PoS of each word. As a result, for this case, 75% of the words remain with one analysis with 95% accuracy, 20% with two analyses and 5% with three analyses.</Paragraph>
      <Paragraph position="4"> Segal (2000) built a transformation-based tagger in the spirit of Brill (1995). In the first phase, the analyses of each word are ranked according to the frequencies of the possible lemmas and tags in a training corpus of about 5,000 words. Selection of the highest ranked analysis for each word gives an accuracy of 83% of the test text - which consists of about 1,000 words. In the second stage, a transformation learning algorithm is applied (in contrast to Brill, the observed transformations are not applied, but used for re-estimation of the word couples probabilities). After this stage, the accuracy is about 93%. The last stage uses a bottom-up parser over a hand-crafted grammar with 150 rules, in order to select the analysis which causes the parsing to be more accurate. Segal reports an accuracy of 95%. Testing his system over a larger test corpus, gives poorer results: Lembersky (2001) reports an accuracy of about 85%.</Paragraph>
      <Paragraph position="5"> Bar-Haim et al. (2005) developed a word segmenter and PoS tagger for Hebrew. In their architecture, words are first segmented into morphemes, and then, as a second stage, these morphemes are tagged with PoS. The method proceeds in two sequential steps: segmentation into morphemes, then tagging over morphemes. The segmentation is based on an HMM and trained over a set of 30K annotated words. The segmentation step reaches an accuracy of 96.74%. PoS tagging, based on unsupervised estimation which combines a small annotated corpus with an untagged corpus of 340K  Word Segmentation Tag Translation bclm bclm PNN name of a human rights association (Betselem) bclm bclm VB while taking a picture bclm bcl-m cons-NNM-suf their onion bclm b-cl-m P1-NNM-suf under their shadow bclm b-clm P1-NNM in a photographer bclm b-clm P1-cons-NNM in a photographer bclm b-clm P1-h-NNM in the photographer hn'im h-n'im P1-VBR that are moving hn'im hn'im P1-h-JJM the lovely hn'im hn'im VBP made pleasant  words by using smoothing technique, gives an accuracy of 90.51%.</Paragraph>
      <Paragraph position="6"> As noted earlier, there is as yet no large scale Hebrew annotated corpus. We are in the process of developing such a corpus, and we have developed tagging guidelines (Elhadad et al., 2005) to define a comprehensive tag set, and assist human taggers achieve high agreement. The results discussed above should be taken as rough approximations of the real performance of the systems, until they can be re-evaluated on such a large scale corpus with a standard tag set.</Paragraph>
      <Paragraph position="7"> Arabic is a language with morphology quite similar to Hebrew. Theoretically, there might be 330,000 possible morphological tags, but in practice, Habash and Rambow (2005) extracted 2,200 different tags from their corpus, with an average number of 2 possible tags per word. As reported by Habash and Rambow, the first work on Arabic tagging which used a corpus for training and evaluation was the work of Diab et al. (2004). Habash and Rambow were the first to use a morphological analyzer as part of their tagger. They developed a supervised morphological disambiguator, based on training corpora of two sets of 120K words, which combines several classifiers of individual morphological features. The accuracy of their analyzer is 94.8% - 96.2% (depending on the test corpus).</Paragraph>
      <Paragraph position="8"> An unsupervised HMM model for dialectal Arabic (which is harder to be tagged than written Arabic), with accurracy of 69.83%, was presented by Duh and Kirchhoff (2005). Their supervised model, trained on a manually annotated corpus, reached an accuracy of 92.53%.</Paragraph>
      <Paragraph position="9"> Arabic morphology seems to be similar to Hebrew morphology, in term of complexity and data sparseness, but comparison of the performances of the baseline tagger used by Habash and Rambow - which selects the most frequent tag for a given word in the training corpus - for Hebrew and Arabic, shows some intriguing differences: 92.53% for Arabic and 71.85% for Hebrew. Furthermore, as mentioned above, even the use of a sophisticated context-free tagger, based on (Levinger et al., 1995), gives low accuracy of 78.2%. This might imply that, despite the similarities, morphological disambiguation in Hebrew might be harder than in Arabic. It could also mean that the tag set used for the Arabic corpora has not been adapted to the specific nature of Arabic morphology (a comment also made in (Habash and Rambow, 2005)).</Paragraph>
      <Paragraph position="10"> We propose an unsupervised morpheme-based HMM to address the data sparseness problem. In contrast to Bar-Haim et al., our model combines segmentation and morphological disambiguation, in parallel. The only resource we use in this work is a morphological analyzer. The analyzer itself can be generated from a word list and a morphological generation module, such as the HSpell wordlist (Har'el and Kenigsberg, 2004).</Paragraph>
    </Section>
  </Section>
class="xml-element"></Paper>