Large-Scale Semi-Supervised Learning for Natural Language ...

Foreign Language F Words f ∈ F Cognates E f+ False FriendsE f−Japanese (Rômaji) napukin napkin nanking, pumpkin, snacking, sneakingFrench abondamment abundantly abandonment, abatement, ... wondermentGerman prozyklische procyclical polished, prophylactic, prophylaxisSpanish viudos widows avoids, idiots, video, videos, virtuousTable 7.1: Foreign-English cognates and false friend training examples.2001] define a word pair (e,f) to be cognate if they are a translation pair (same meaning)and their edit distance is less than three (same form). We adopt an improved definition(suggested by [Melamed, 1999] for the French-English Canadian Hansards) that doesnot over-propose shorter word pairs: (e,f) are cognate if they are translations and theirLCSR ≥ 0.58. Note that this cutoff is somewhat conservative: the English/German cognateslight/Licht (LCSR=0.8) are included, but not the cognates eight/acht (LCSR=0.4).If two words must have LCSR ≥ 0.58 to be cognate, then for a given word f ∈ F , weneed only consider as possible cognates the subset of words in E having an LCSR with flarger than 0.58, a set we callE f . The portion ofE f with the same meaning asf,E f+ , arecognates, while the part with different meanings, E f− , are not cognates. The words E f−with similar spelling but different meaning are sometimes called false friends. The cognateidentification task is, for every word f ∈ F , and a list of similarly spelled words E f , todistinguish the cognate subset E f+ from the false friend setE f− .To create training data for our learning approaches, and to generate a high-quality labeledtest set, we need to annotate some of the (f,e f ∈ E f ) word pairs for whether ornot the words share a common meaning. In Section 7.5, we explain our two high-precisionautomatic annotation methods: checking if each pair of words (a) were aligned in a wordalignedbitext, or (b) were listed as translation pairs in a bilingual dictionary.Table 7.1 provides some labeled examples with non-empty cognate and false friendlists. Note that despite what it may appear from these examples, this is not a ranking task:even in highly related languages, most words in F have empty E f+ lists, and many haveempty E f− as well. Thus one natural formulation for cognate identification is a pairwise(and symmetric) cognation classification that looks at each pair (f,e f ) separately and individuallymakes a decision:+(napukin,napkin)– (napukin,nanking)– (napukin,pumpkin)In this formulation, the benefits of a discriminative approach are clear: it must find substringsthat distinguish cognate pairs from word pairs with otherwise similar form. [Klementievand Roth, 2006], although using a discriminative approach, do not provide theirinfinite-attribute perceptron with competitive counter-examples. They instead use transliterationsas positives and randomly-paired English and Russian words as negative examples.In the following section, we also improve on [Klementiev and Roth, 2006] by using acharacter-based string alignment to focus the features for discrimination.7.4 Features for Discriminative String SimilarityAs Chapter 2 explained, discriminative training learns a classifier from a set of labeledtraining examples, each represented as a set of features. In the previous section we showed99