"... In PAGE 4: ... In order to perform learning in the meta level, we need inductive algorithms that can handle features, of that nature. The full set of morphology characteristics is presented in Table1 . Although the set is limited, and morphology characteristics that de- scribe class di erences are missing, we get interesting and interpretable results.... In PAGE 6: ... The distance d between two datasets D; D0 with respective morpholo- gies lt; m1; : : : ; mn gt; and lt; m0 1; : : : ; m0 n gt; is d = n X i=1 d2 ii where dii = mi ? m0 i if mi; m0 i 2 R = 1 if one of mi; m0 i is non-appl = 0 if both of mi; m0 i are non-appl 4.2 Combining the Inductive Models with NOEMON-SEL Whenever a new dataset is input to the system, NOEMON-SEL extracts its morphological character- istics,shown in Table1 , in a preprocessing step and feeds them to the inductive models in the KB. Each model proposes one of the two algorithms (associated with the meta-learning problem from which it was pro- duced) or indicates a tie.... In PAGE 6: ... Although the accuracy achieved is rather low, we still get an im- provement over the default accuracy of 40% for the IBL-NB problem, 36% for the ID3-NB, and 33% for the IBL-ID3, for the models induced from the full set of data sets characteristics, given in Table 1. In an e ort to improve the accuracy we ran various suites of experiments where the sets of morphology characteristics of the meta-learning datasets were sub- sets of the initial set presented in Table1 . The results of these experiments are also given in Table 3.... ..."

"... In PAGE 7: ... As with the EM trained models, combining lin- ear and morphological contexts is always beneficial. To put these numbers in context, Table3 lists current state-of-the art results for the same task. CE+spl is the Contrastive-Estimation CRF method of SE.... ..."

"... In PAGE 5: ... 4.3 Experiments and results Table2 summarizes the relative performance of the var- ious methods (cf. section 3) in four different learning set- tings (cf.... ..."

"... In PAGE 4: ... The other disadvantage is the shrinkage of LM span which has probably re- sulted in an increase in the normalized perplexity. The word error rates (with/without adaptation) presented in Table5 clearly indicate the improvement with word splitting. The best results are obtained using the data driven word splitting us- ing counts.... ..."

"... In PAGE 10: ... We want to suggest a classi cation of all possible applications of ILP learning to word morphology based on two factors, the extent to which linguistic concepts are used as background concepts, and the amount of annotation in data. The lower part of Table2 (see page 18) gives a graphical representation of that clas- si cation. The horizontal axis represents a partial ordering between data sets|if two data sets containing the same words are displayed in di erent columns, the one on the left-hand side contains all annotation from the rst data set plus some additional one.... In PAGE 11: ...available in the data. The application of any preference bias for unsupervised learning from the upper part of Table2 would result in a right-to-left move in parallel to the horizontal axis. Indeed, each of these biases would add extra annotation to the data yet their contribution to the concept language would usually be limited to the creation of new theory constants, typically segments of words.... ..."

"... In PAGE 2: ..., 1990). The tag- ging and bracketing tasks utilized approximately 2 million words in each language, with the sample sizes for morphology induc- tion given in Table3 . All word alignments utilized strictly raw- word-based model variants for English/French/Spanish/Czech and character-based model variants for Chinese, with no use of mor- phological analysis or stemming, POS-tagging, bracketing or dic- tionary resources.... In PAGE 5: ...a86 a81 a78a80a79a25a92 a90a93a52a94a95a87a88a85a89a90 a91 croyaient croissant believe believing believed French Inflections BELIEVE English Bridge Lemmas thought think THINK croitreV French Roots croire a91 a90a93a85a94 a78a80a79a96a81a25a82 a83a52a83a85a84a96a86 a92 Figure 8: Multi-bridge French inflection/root alignment lemmatization can be potentially utilized for all other English lem- mas (such as THINK) with which croyaient and croire also asso- ciate, offering greater potential coverage and robustness via multi- ple bridges. Formally, these multiple transitive linkages can be modeled as shown below, by summing over all English lemmas (a61a98a97a100a99a102a101a104a103 ) with which either a candidate foreign inflection (a105 infl) or its root (a105 root) exhibit an alignment in the parallel corpus: a22a54a101a67a106a107a18a85a105a54a108a43a109a43a109a111a110a43a19 a105 a49a41a112a21a113 a97 a20a67a66a72a68 a49 a22 a16 a18a85a105a114a108a50a109a43a109a111a110a111a19 a61 a97a100a99a102a101 a103a111a20a70a22 a16 a18a85a61 a97a100a99a102a101 a103a48a19 a105 a49a58a112a115a113 a97 a20 For example: a22a114a101a67a106a74a18 croirea19 croyaienta20a71a66 a22a54a16a74a18 croirea19 BELIEVEa20a73a22a54a16a70a18 BELIEVEa19 croyaienta20a50a75 a22 a16 a18 croirea19 THINKa20a70a22 a16 a18 THINKa19 croyaienta20a116a75a77a56a59a56a59a56 This projection/bridge-basedsimilarity measure a22 mpa18a85a105 roota19 a105 infla20 can be quite effective on its own, as shown in the MProj only entries in Table3 (for multiple parallel corpora in 3 different languages), especially when restricted to the highest-confidence subset of the vocabulary (5.2% to 77.... In PAGE 6: ...ranslation of a common source (e.g. the Latin Vulgate Bible). As shown in Table3 , using the previously analyzed French Bible as a... ..."

"... In PAGE 5: ...1. Table2 shows the runtime for a few industrial instances. We can see that the induction-based learning can be very powerful, espe- cially for hard UNSAT cases.... ..."