<?xml version="1.0" standalone="yes"?>
<Paper uid="W06-3206">
  <Title>constraint satisfaction inference</Title>
  <Section position="7" start_page="47" end_page="48" type="concl">
    <SectionTitle>
5 Discussion
</SectionTitle>
    <Paragraph position="0"> We have presented constraint satisfaction inference as a global method to repair errors made by a local classifier. This classifier is a memory-based learner predicting overlapping trigrams, creating a space of possible output sequences in which the inference procedure finds the globally optimal one. This globally optimal sequence is the one that adheres best to the trigram, bigram, and unigram sub-sequence constraints present in the predictions of the local classifier, weighted by the confidences of the classifier, in a back-off order from trigrams to unigrams.</Paragraph>
    <Paragraph position="1"> The method is shown to significantly outperform a memory-based classifier predicting atomic classes and lacking any global post-processing, which has previously been shown to exhibit successful performance (Van den Bosch and Daelemans, 1993; Van den Bosch and Daelemans, 1999). (While this was the reason for using memory-based learning, we note that the constraint satisfaction inference and its underlying trigram-based classification method can be applied to any machine-learning classifier.) The large improvements (27% and 26% error reductions on the two English tasks, 18% and 22% on the two Dutch tasks) can arguably be taken as an indication that this method may be quite effective in general in morpho-phonological sequence processing tasks.</Paragraph>
    <Paragraph position="2"> Apparently, the constraint-satisfaction method is able to avoid more errors than to add them. At closer inspection, comparing cases in which the atomic classifier generates errors and constraint satisfaction inference does not, we find that the type of avoided error, when compared to the unigram classifier, differs per task. On the morphological analysis task, we identify repairs where (1) a correct segmentation is inserted, (2) a false segmentation is not placed, and (3) a tag is switched. As Table 6 shows in its upper four lines, in the case of English most repairs involve correctly inserted segmentations, but the other two categories are also quite frequent. In the case of Dutch the most common repair is a switch from an incorrect tag, placed at the right segmentation position, to the correct tag at that point. Given that there are over three thousand possible tags in our complicated Dutch morphological analysis task, this is indeed a likely area where there is room for improvement. null Morphological analysis repairs English Dutch  three categories of morphological analysis classifications (top) and letter-phoneme conversions (bottom) of the constraint satisfaction inference method as compared to the unigram classifier.</Paragraph>
    <Paragraph position="3"> The bottom four lines of Table 6 lists the counts of repaired errors in word phonemization in both languages, where we distinguish between (1) alignment repairs between phonemes and alignment symbols (where phonemes are corrected to phonemic nulls, or vice versa), (2) switches from incorrect non-null phonemes to correct vowels, and (3) switches from incorrect non-null phonemes to correct consonants.</Paragraph>
    <Paragraph position="4"> Contrary to expectation, it is not the second vowel category in which most repairs are made (many of the vowel errors in fact remain in the output), but the alignment category, in both languages. At points where the local unigram classifier sometimes incorrectly predicts a phoneme twice, where it should have predicted it along with a phonemic null, the constraint satisfaction inference method never generates a double phoneme. Hence, the method succeeds in generating sequences that are possible, and avoiding impossible sub-sequences. At the same time, a possible sequence is not necessarily the correct sequence, so this method can be expected to still make errors on the identity of labels in the output sequence. null In future work we plan to test a range of n-gram widths exceeding the current trigrams. Preliminary results suggest that the method retains a positive effect over the baseline with n &gt; 3, but it does not outperform the n = 3 case. We also intend to test the method with a range of different machine learning methods, since as we noted before the constraint-satisfaction inference method and its underlying n-gram output subsequence classification method can  be applied to any machine learning classification algorithm in principle, as is already supported by preliminary work in this direction.</Paragraph>
    <Paragraph position="5"> Also, we plan comparisons to the work of Stroppa and Yvon (2005) and Damper and Eastmond (1997) on sequence-global analogy-based models for morpho-phonological processing, since the main difference between this related work and ours is that both alternatives are based on working units of variable width, rather than our fixedwidth n-grams, and also their analogical reasoning is based on interestingly different principles than our k-nearest neighbor classification rule, such as the use of analogical proportions by Stroppa and Yvon (2005).</Paragraph>
  </Section>
class="xml-element"></Paper>