12.07.2015 Views

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

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

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

SHOW MORE
SHOW LESS

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

System IN O1 O2[Malouf, 2000] 91.5 65.6 71.6webc(a 1 ,a 2 ) vs. c(a 2 ,a 1 ) 87.1 83.7 86.0SVM with N-GM features 90.0 85.8 88.5SVM with LEX features 93.0 70.0 73.9SVM with N-GM + LEX 93.7 83.6 85.4Table 5.3: Adjective ordering accuracy (%). SVM and [Malouf, 2000] trained on BNC,tested on BNC (IN), Gutenberg (O1), and Medline (O2).Finally, we also have features <strong>for</strong> all suffixes of length 1-to-4 letters, as these encodeuseful in<strong>for</strong>mation about adjective class [Malouf, 2000]. Like the adjective features, thesuffix features receive a value of +1 <strong>for</strong> adjectives in the first position and −1 <strong>for</strong> those inthe second.N-GM features[Lapata and Keller, 2005] propose a web-based approach to adjective ordering: take themost-frequent order of the words on the web, c(a 1 ,a 2 ) vs. c(a 2 ,a 1 ). We adopt this as ourunsupervised approach. We merge the counts <strong>for</strong> the adjectives occurring contiguously andseparated by a comma.These are the most useful N-GM features; we include them but also other, tag-basedcounts from Google V2. Raw counts include cases where one of the adjectives is notused as a modifier: “the special present was” vs. “the present special issue.” We includelog-counts <strong>for</strong> the following, more-targeted patterns: 5 c(a 1 a 2 N.*), c(a 2 a 1 N.*),c(DT a 1 a 2 N.*), c(DT a 2 a 1 N.*). We also include features <strong>for</strong> the log-counts of eachadjective preceded or followed by a word matching an adjective-tag: c(a 1 J.*), c(J.* a 1 ),c(a 2 J.*), c(J.* a 2 ). These assess the positional preferences of each adjective. Finally, weinclude the log-frequency of each adjective. The more frequent adjective occurs first in57% of pairs.As in all tasks, the counts are features in a classifier, so the importance of the differentpatterns is weighted discriminatively during training.5.3.2 Adjective Ordering ResultsIn-domain, with both feature classes, we set a strong new standard on this data: 93.7%accuracy <strong>for</strong> the N-GM+LEX system (Table 5.3). We trained and tested [Malouf, 2000]’sprogram on our data; our LEX classifier, which also uses no auxiliary corpus, makes 18%fewer errors than Malouf’s system. Our web-based N-GM model is also superior to thedirect evidence web-based approach of [Lapata and Keller, 2005], scoring 90.0% vs. 87.1%accuracy. These results show the benefit of both our new lexicalized and our new web-basedfeatures.Figure 5.1 gives the in-domain learning curve. With fewer training examples, the systemswith N-GM features strongly outper<strong>for</strong>m the LEX-only system. Note that with tens ofthousands of test examples, all differences are highly significant.Out-of-domain, LEX’s accuracy drops a shocking 23% on Gutenberg and 19% on Medline(Table 5.3). [Malouf, 2000]’s system fares even worse. The overlap between training5 In this notation, capital letters (and reg-exps) are matched against tags while a 1 and a 2 match words.73

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!