alogue system domain, we propose a stack-basedsemantic representation at the phrase level, whichis expressive enough to generate natural utterancesfrom unseen inputs, yet simple enough for data tobe collected from 42 untrained annotators with aminimal normalisation step. A human evaluationover 202 d[r]
ata, 2007)) is orthogonal to the use of co-trainingin our work, since it only enhances each MT sys-tem in our ensemble by exploiting the multilingualdata. In future work, we plan to incorporate trian-gulation into our active learning approach.7 ConclusionThis paper introduced the novel[r]
Corpus-based supervised learning is now a stan-dard approach to achieve high-performance in nat-ural language processing. However, the weaknessof supervised learning approach is to need an anno-tated corpus, the size of which is reasonably large.Even if we have a good supervised-lea[r]
is selected using algorithms developed here, and areannotated by human beings and are then added to trainingdata to rebuild the model. The procedure is iterated untilthe model reaches a certain accuracy level.Our efforts are devoted to two aspects: first, we be-lieve that the selected samples should[r]
To our knowledge, it is the first work on consider-ing the three criteria all together for active learning. Furthermore, such measures and strategies can be easily adapted to other active learning tasks as well. 2 Multi-criteria for NER Active Learning Supp[r]
data. The results are summarized in the right panelof Fig. 1, where we plot one point per dataset. Pointsabove the diagonal-line demonstrate the superiorityof ACL over CW Margin. ACL required fewer la-bels than CW margin twice as often as the oppositeoccurred (8 vs 4). Note that CW Margin used morel[r]
currently the methodological backbone for lots ofNLP activities. Despite their success they create acostly follow-up problem, viz. the need for humanannotators to supply large amounts of “golden”annotation data on which ML systems can betrained. In most annotation campaigns, the lan-guage material c[r]
dently of the other. In the general language UPennannotation efforts for the WSJ sections of the PennTreebank (Marcus et al., 1993), sentences are anno-tated with POS tags, parse trees, as well as discourseannotation from the Penn Discourse Treebank (Milt-sakaki et al., 2008), while verbs and verb a[r]
tion noise in the context of supervised classification.Section 3 describes the experimental setup of oursimulation study and presents results. In Section 4we present our filtering approach and show its im-pact on AL performance. Section 5 concludes andoutlines future work.2 Related WorkWe are interest[r]
sentence, there can be more than one manuallyaligned target word. The restricted training willallow only those paths, which are consistent with366the manual alignments. Therefore, the restrictionof the alignment paths reduces to restricting thesummation in EM.4 Query Strategies for Link SelectionWe[r]
Proceedings of the ACL-IJCNLP 2009 Conference Short Papers, pages 61–64,Suntec, Singapore, 4 August 2009.c2009 ACL and AFNLPA Combination of Active Learning and Semi-supervised Learning Starting with Positive and Unlabeled Examples for Word Sense Disambiguation: An Empirical St[r]
Previous work tries to do as much as possibleusing only s. Typically s is assumed to contain atleast 2 words and often many more. Pantel et al.(2009) discusses the issue of seed set size in detail,concluding that 5-20 seed words are often requiredfor good performance.There are several problems with[r]
Table 1: The average number of senses in BC andWSJ, average MFS accuracy, average number of BCtraining, and WSJ adaptation examples per noun.data, and the rest of the WSJ examples are desig-nated as in-domain adaptation data. The row 21nouns in Table 1 shows some information aboutthese 21 nouns. For[r]
and added into the training data.In experiments about chunk wise selection 4000pairs of bunsetsus, which are roughly equal to theaveraged number of bunsetsus in 500 sentences,are selected at each iteration of active learning.6.6 Dependency AccuracyWe use dependency accuracy as a perfor[r]
utze and Keh-Yih Su, editors, Proceedings of the 2000 Joint SIG-DAT Conference on Empirical Methods in NaturalLanguage Processing, pages 45–53. Association forComputational Linguistics, Somerset, New Jersey.Ashish Kapoor, Eric Horvitz, and Sumit Basu. 2007.Selective supervision: Guiding supervised l[r]
logical issues relevant to this decision?Family History Project 225Mark R. Warren, Harvard UniversityWith this exercise, you have the opportunity to explore in depth your own family history. Youwill conduct interviews with six family members, going back as many generations as possi-ble. As you and y[r]
Proceedings of ACL-08: HLT, Short Papers (Companion Volume), pages 65–68,Columbus, Ohio, USA, June 2008.c2008 Association for Computational LinguisticsAssessing the Costs of Sampling Methods in Active Learning for AnnotationRobbie Haertel, Eric Ringger, Kevin Seppi, James Carroll, Pet[r]
(AL) is compelling. AL techniques for IMT se-lectively ask an oracle (e.g. a human transla-tor) to supervise a small portion of the incomingsentences. Sentences are selected so that SMTmodels estimated from them translate new sen-tences as accurately as possible. There are threechallenges when apply[r]
port the number of words along the x-axis. Wehave verified that our system does not substantiallyalter the average number of letters per word in thetraining set for any of these languages. Hence, thenumber of words reported here is representative ofthe true annotation effort.4Subtree raising is an ex[r]
Bayesian Methods for Machine LearningZoubin Ghahramani* Tutorial Notes Now Available Here *TopicMany topics in Machine Learning (e.g. kernel methods, clustering, semi-supervised learning, feature selection, active learning, reinforcement learning)can be addressed w[r]