08-2001_0	A preparatory study on the capitalization of Portuguese BN has been performed by ( )	DT JJ NN IN DT NN IN NP NP VHZ VBN VVN IN ( )	BackGround	GRelated	Neutral
08-2001_0	Results from previous experiment are still worse than results achieved by other work on the area ( ) (about 94% precision and 88% recall), specially in terms of recall	NNS IN JJ NN VBP RB JJR IN NNS VVN IN JJ NN IN DT NN ( ) NN CD NN CC CD NN , RB IN NNS IN NN	Compare	Compare	Positive
08-2001_1	The modeling approach here described is discriminative, and is based on maximum entropy (ME) models, firstly applied to natural language problems in ( )	DT NN NN RB VVN VBZ JJ , CC VBZ VVN IN JJ NN NN NNS , RB VVN TO JJ NN NNS IN ( )	BackGround	SRelated	Neutral
08-2001_2	The capitalization problem can be seen as a sequence tagging problem ( ), where each lower-case word is associated to a tag that describes its capitalization form	DT NN NN MD VB VVN IN DT NN VVG NN ( ) , WRB DT JJ NN VBZ VVN TO DT NN WDT VVZ PP$ NN NN	BackGround	GRelated	Neutral
08-2001_3	Concerning this subject, ( ) shows that, as the time gap between training and test data increases, the performance of a named tagger based on co-training ( ) decreases	VVG DT NN , ( ) VVZ IN/that , IN DT NN NN IN NN CC NN NNS NNS , DT NN IN DT VVN NN VVN IN NN ( ) NNS	BackGround	SRelated	Neutral
08-2001_5	Evaluation results may be influenced when taking such words into account ( )	NN NNS MD VB VVN WRB VVG JJ NNS IN NN ( )	BackGround	SRelated	Neutral
08-2001_6	The capitalization task, also known as truecasing ( ), consists of rewriting each word of an input text with its proper case information	DT NN NN , RB VVN IN NN ( ) , VVZ IN VVG DT NN IN DT NN NN IN PP$ JJ NN NN	BackGround	GRelated	Neutral
08-2001_6	( ) builds a trigram language model (LM) with pairs (word, tag), estimated from a corpus with case information, and then uses dynamic programming to disambiguate over all possible tag assignments on a sentence	( ) VVZ DT NN NN NN NN IN NNS NN , NN , VVN IN DT NN IN NN NN , CC RB VVZ JJ NN TO VV IN DT JJ NN NNS IN DT NN	BackGround	GRelated	Neutral
08-2001_7	The evaluation is performed using the metrics: Precision, Recall and SER (Slot Error Rate) ( )	DT NN VBZ VVN VVG DT JJ NN , NP CC NP NP NP NP ( )	Fundamental	Basis	Neutral
08-2001_8	In fact, subtitling of BN has led us into using a baseline vocabulary of 100K words combined with a daily modification of the vocabulary ( ) and a re-estimation of the language model	IN NN , VVG IN NP VHZ VVN PP IN VVG DT JJ NN IN NN NNS VVN IN DT JJ NN IN DT NN ( ) CC DT NN IN DT NN NN	Fundamental	Basis	Neutral
08-2001_10	Other related work includes a bilingual capitalization model for capitalizing machine translation (MT) outputs, using conditional random fields (CRFs) reported by ( )	JJ JJ NN VVZ DT JJ NN NN IN VVG NN NN NN NNS , VVG JJ JJ NNS NN VVN IN ( )	BackGround	GRelated	Neutral
08-2002_0	Whereas the more commonly applied Akaike Information Criterion ( ) requires the number of estimated parameters to be determined exactly, the DIC facilitates the evaluation of mixed-effects models by relaxing this requirement	IN DT RBR RB VVN NP NP NP ( ) VVZ DT NN IN JJ NNS TO VB VVN RB , DT NP VVZ DT NN IN JJ NNS IN VVG DT NN	BackGround	SRelated	Negative
08-2002_1	A growing body of work in cognitive science characterizes human readers as some kind of probabilistic parser ( )	DT VVG NN IN NN IN JJ NN VVZ JJ NNS IN DT NN IN JJ NN ( )	BackGround	GRelated	Neutral
08-2002_1	This observation is consistent with Brants and Crocker's ( ) observation that accuracy can be maintained even when restricted to 1% of the memory required for exhaustive parsing	DT NN VBZ JJ IN NP CC NP ( ) NN IN/that NN MD VB VVN RB WRB VVN TO CD IN DT NN VVN IN JJ VVG	Compare	Compare	Positive
08-2002_4	This basic notion has proved remarkably applicable across sentence types and languages ( )	DT JJ NN VHZ VVN RB JJ IN NN NNS CC NNS ( )	BackGround	GRelated	Positive
08-2002_5	The length of time that a reader's eyes spend fixated on a particular word in a sentence is known to be affected by a variety of word-level factors such as length in characters, n-gram frequency and empirical predictability ( )	DT NN IN NN IN/that DT NNS NNS VVP VVN IN DT JJ NN IN DT NN VBZ VVN TO VB VVN IN DT NN IN JJ NNS JJ IN NN IN NNS , NN NN CC JJ NN ( )	BackGround	SRelated	Neutral
08-2002_6	From a cognitive perspective, the utility of small k parsers for modeling comprehension difficulty lends credence to the view that the human processor is a single-path analyzer ( )	IN DT JJ NN , DT NN IN JJ NN NNS IN VVG NN NN VVZ NN TO DT NN IN/that DT JJ NN VBZ DT NN NN ( )	BackGround	GRelated	Neutral
08-2002_7	Hale ( ) suggests this quantity as an index of psycholinguistic difficulty	NP ( ) VVZ DT NN IN DT NN IN JJ NN	BackGround	GRelated	Neutral
08-2002_8	The parser's outputs define a relation on word pairs ( )	DT NNS NNS VV DT NN IN NN NNS ( )	BackGround	SRelated	Neutral
08-2002_11	Our methodology imposes this requirement by fitting a kind of regression known as a linear mixed-effects model to the total reading times associated with each sentence-medial word in the Potsdam Sentence Corpus (PSC) ( )	PP$ NN VVZ DT NN IN NN DT NN IN NN VVN IN DT JJ NN NN TO DT JJ NN NNS VVN IN DT JJ NN IN DT NP NP NP NN ( )	Fundamental	Basis	Neutral
08-2002_12	When more than one transition is applicable, the parser decides between them by consulting a probability model derived from the Negra and Tiger newspaper corpora ( )	WRB JJR IN CD NN VBZ JJ , DT NN VVZ IN PP IN VVG DT NN NN VVN IN DT NP CC NP NN NNS ( )	BackGround	SRelated	Neutral
08-2002_14	From the theoretical side, we calculate word-by-word surprisal predictions from a family of incremental dependency parsers for German based on Nivre ( ); these parsers differ only in the size k of the beam used in the search for analyses of longer and longer sentence-initial substrings	IN DT JJ NN , PP VVP NN NN NNS IN DT NN IN JJ NN NNS IN JJ VVN IN NP ( NN DT NNS VVP RB IN DT NN NN IN DT NN VVN IN DT NN IN NNS IN JJR CC RBR JJ NNS	Fundamental	Basis	Neutral
08-2002_16	Perhaps human parsing is boundedly rational in the sense of the bound imposed by Stack3 ( )	RB JJ VVG VBZ RB JJ IN DT NN IN DT VVN VVN IN NP ( )	BackGround	MRelated	Neutral
08-2002_18	We evaluated the change in relative 7 quality of fit due to surprisal with the Deviance Information Criterion (DIC) discussed in Spiegelhalter et al.( )	PP VVD DT NN IN JJ CD NN IN NN JJ TO NN IN DT NP NP NP NN VVN IN NP NP NP ( )	Fundamental	Basis	Neutral
08-2003_0	Because of this, ( ) defines the tree kernel algorithm whose computational complexity does not depend on m	IN IN DT , ( ) VVZ DT NN NN NN WP$ JJ NN VVZ RB VV IN NN	BackGround	SRelated	Neutral
08-2003_1	Recently, the graph-based method (LexRank) is applied successfully to generic, multi-document summarization ( )	RB , DT JJ NN NN VBZ VVN RB TO JJ , NN NN ( )	BackGround	GRelated	Positive
08-2003_1	In ( ), the concept of graph-based centrality is used to rank a set of sentences, in producing generic multi-document summaries	IN ( ) , DT NN IN JJ NN VBZ VVN TO VV DT NN IN NNS , IN VVG JJ NN NNS	BackGround	SRelated	Neutral
08-2003_2	Initiatives such as PropBank (PB) ( ) have made possible the design of accurate automatic Semantic Role Labeling (SRL) systems like ASSERT ( )	NNS JJ IN NP NP ( ) VHP VVN JJ DT NN IN JJ JJ NP NP NP NN NNS IN NP ( )	BackGround	GRelated	Positive
08-2003_4	We carried out automatic evaluation of our summaries using ROUGE ( ) toolkit, which has been widely adopted by DUC for automatic summarization evaluation	PP VVD RP JJ NN IN PP$ NNS VVG NP ( ) NN , WDT VHZ VBN RB VVN IN NP IN JJ NN NN	Fundamental	Basis	Positive
08-2003_6	A topic-sensitive LexRank is proposed in ( )	DT JJ NN VBZ VVN IN ( )	BackGround	GRelated	Neutral
08-2003_6	To apply LexRank to query-focused context, a topic-sensitive version of LexRank is proposed in ( )	TO VV NN TO JJ NN , DT JJ NN IN NP VBZ VVN IN ( )	BackGround	SRelated	Neutral
08-2004_2	So far, linguistic cues have played an important role in research of subjectivity recognition (e.g.( )), sentiment analysis (e.g.( )), and emotion studies (e.g.( ))	RB RB , JJ NNS VHP VVN DT JJ NN IN NN IN NN NN NP ( NP , NN NN NN ( NN , CC NN NNS JJ ( NN	BackGround	GRelated	Positive
08-2004_4	Wiebe ( ) further adapted this definition of subjectivity to be "the linguistic expression of private states ( )"	NP ( ) RBR VVN DT NN IN NN TO VB JJ JJ NN IN JJ NNS ( NN	BackGround	GRelated	Neutral
08-2004_5	They have shown that subjectivity annotations can be helpful for word sense disambiguation when a word has distinct subjective senses and objective senses ( )	PP VHP VVN IN/that NN NNS MD VB JJ IN NN NN NN WRB DT NN VHZ JJ JJ NNS CC JJ NNS ( )	BackGround	SRelated	Positive
08-2004_6	First, non-objectivity cannot be clearly identified without knowledge about its source ( )	RB , NN MD VB RB VVN IN NN IN PP$ NN ( )	BackGround	SRelated	Neutral
08-2004_7	Second, non-objectivity always lies in a context, which cannot be ignored ( )	RB , NN RB VVZ IN DT NN , WDT MD VB VVN ( )	BackGround	SRelated	Neutral
08-2005_0	We have examined sentence extraction agreement between experts using the prevalence-adjusted bias-adjusted (PABA) kappa to account for prevalence of judgments and conflicting biases amongst experts ( )	PP VHP VVN NN NN NN IN NNS VVG DT JJ JJ NN NN TO VV IN NN IN NNS CC JJ NNS IN NNS ( )	Fundamental	Basis	Neutral
08-2005_3	The gazetteer feature checks named entities from each sentence against the Alexandria Digital Library (ADL) Gazetteer ( )	DT NN NN NNS VVD NNS IN DT NN IN DT NP NP NP NP NP ( )	BackGround	SRelated	Neutral
08-2005_4	Knowledge maps consist of nodes containing rich concept descriptions interconnected using a limited set of relationship types ( )	NN NNS VVP IN NNS VVG JJ NN NNS VVN VVG DT JJ NN IN NN NNS ( )	BackGround	GRelated	Neutral
08-2005_5	We use ROUGE ( ) to assess summary quality using common n-gram counts and longest common subsequence (LCS) measures	PP VVP NP ( ) TO VV NN NN VVG JJ NN NNS CC JJS JJ NN NN NNS	Fundamental	Basis	Neutral
08-2005_6	Lin and Pantel ( ) discover concepts using clustering by committee to group terms into conceptually related clusters	NP CC NP ( ) VV NNS VVG VVG IN NN TO NN NNS IN RB VVN NNS	BackGround	GRelated	Neutral
08-2005_7	On-toLearn extracts candidate domain terms from texts using a syntactic parse and updates an existing ontology with the identified concepts and relationships ( )	NP VVZ NN NN NNS IN NNS VVG DT JJ VVP CC NNS DT JJ NN IN DT VVN NNS CC NNS ( )	BackGround	GRelated	Neutral
08-2005_8	Learning research indicates that knowledge maps may be useful for learners to understand the macro-level structure of an information space ( )	VVG NN VVZ IN/that NN NNS MD VB JJ IN NNS TO VV DT JJ NN IN DT NN NN ( )	BackGround	GRelated	Positive
08-2005_10	Finally, MEAD is a widely used MDS and evaluation platform ( )	RB , NP VBZ DT RB VVN NNS CC NN NN ( )	BackGround	GRelated	Neutral
08-2005_10	We have implemented an extractive summarizer for educational science content, COGENT, based on MEAD version 3.11 ( )	PP VHP VVN DT JJ NN IN JJ NN NN , NP , VVN IN JJ NN CD ( )	Fundamental	Basis	Neutral
08-2005_11	Knowledge Puzzle focuses on n-gram identification to produce a list of candidate terms pruned using information extraction techniques to derive the ontology ( )	NN NN VVZ IN NN NN TO VV DT NN IN NN NNS VVN VVG NN NN NNS TO VV DT NN ( )	BackGround	GRelated	Neutral
08-2007_0	For some restricted combinatorial spaces of alignmentsthose that arise in ITG-based phrase models ( ) or local distortion models ( )inference can be accomplished using polynomial time dynamic programs	IN DT JJ JJ NNS IN NN WDT VVP IN JJ NN NNS ( ) CC JJ NN NNS ( NN MD VB VVN VVG JJ NN JJ NNS	BackGround	GRelated	Neutral
08-2007_1	However, for more permissive models such as Marcu and Wong ( ) and DeNero et al.( ), which operate over the full space of bijective phrase alignments (see below), no polynomial time algorithms for exact inference have been exhibited	RB , IN JJR JJ NNS JJ IN NP CC NP ( ) CC NP NP NP ( ) , WDT VVP IN DT JJ NN IN JJ NN NNS NN NN , DT JJ NN NNS IN JJ NN VHP VBN VVN	BackGround	GRelated	Negative
08-2007_1	Then, we can formally define the set of bijective phrase alignments: A = |J   dj = e ;   |J   f ki = f > Both the conditional model of DeNero et al.( ) and the joint model of Marcu and Wong ( ) operate in A, as does the phrase-based decoding framework of Koehn et al.( )	RB , PP MD RB VV DT NN IN JJ NN NP NP SYM NP NP SYM SYM : NN SYM NP SYM SYM SYM CC DT JJ NN IN NP NP NP ( ) CC DT JJ NN IN NP CC NP ( ) VV IN DT , RB VVZ DT JJ VVG NN IN NP NP NP ( )	BackGround	SRelated	Neutral
08-2007_1	DeNero et al.( ) instead proposes an exponential-time dynamic program to systematically explore A, which can in principle solve either O or ?NP NP NP ( ) RB VVZ DT JJ JJ NN TO RB VV DT , WDT MD IN NN VV DT NN CC $		BackGround	SRelated	Neutral
08-2007_3	Using an off-the-shelf ILP solver, 4 we were able to quickly and reliably find the globally optimal phrase alignment under  <fi(eij, fki) derived from the Moses pipeline ( )	VVG DT NN NN NN , LS PP VBD JJ TO RB CC RB VV DT RB JJ NN NN IN NP , NN VVN IN DT NN NN ( )	Fundamental	Basis	Neutral
08-2007_4	Forced decoding arises in online discriminative training, where model updates are made toward the most likely derivation of a gold translation ( )	JJ VVG VVZ IN JJ JJ NN , WRB NN NNS VBP VVN IN DT RBS JJ NN IN DT JJ NN ( )	BackGround	SRelated	Neutral
08-2007_5	The existence of a polynomial time algorithm for ?implies a polynomial time algorithm for S, because A = 3  Complexity of Inference in  A For the space A of bijective alignments, problems ?and O have long been suspected of being NP-hard, first asserted but not proven in Marcu and Wong ( )	DT NN IN DT JJ NN NN IN $ VVZ DT JJ NN NN IN NP , IN NP SYM CD NN IN NN IN DT IN DT NN DT IN JJ NNS , NNS $ CC NP VHP RB VBN VVN IN VBG NP , RB VVD CC RB VVN IN NP CC NP ( )	BackGround	SRelated	Neutral
08-2007_5	Marcu and Wong ( ) describes an approximation to O	NP CC NP ( ) VVZ DT NN TO NP	BackGround	SRelated	Neutral
08-2007_6	CPM is #P-complete ( ), so S (and hence ? is #P-hard	NP VBZ JJ ( ) , RB NP NN RB JJ VBZ JJ	BackGround	SRelated	Neutral
08-2008_0	We selected a subset of the verbs annotated in the OntoNotes project ( ) that had at least 50 instances	PP VVD DT NN IN DT NNS VVN IN DT NP NN ( ) WDT VHD IN JJS CD NNS	Fundamental	Basis	Neutral
08-2008_1	The approaches to obtaining this kind of knowledge can be based on extracting it from ele c-tronic dictionaries such as WordNet ( ), using Named Entity (NE) tags, or a combi-nation of both ( )	DT VVZ TO VVG DT NN IN NN MD VB VVN IN VVG PP IN NP NP NNS JJ IN NP ( ) , VVG NP NP NN NNS , CC DT NN IN DT ( )	BackGround	GRelated	Neutral
08-2008_3	An automatic VSD system usually has at its disposal a diverse set of features among which the semantic features play an important role: verb sense distinctions often depend on the distinctions in the semantics of the target verb's arguments ( )	DT JJ NN NN RB VHZ IN PP$ NN DT JJ NN IN NNS IN WDT DT JJ NNS VVP DT JJ NN NN NN NNS RB VVP IN DT NNS IN DT NNS IN DT NN NN NNS ( )	BackGround	GRelated	Positive
08-2008_4	Hindle ( ) grouped nouns into thesaurus-like lists based on the similarity of their syntactic contexts	NP ( ) VVN NNS IN JJ NNS VVN IN DT NN IN PP$ JJ NNS	BackGround	GRelated	Neutral
08-2008_5	To collect this data, we utilized two resources: (1) MaltParser ( ) - a high-efficiency dependency parser; (2) English Gigaword - a large corpus of 5.7M news articles	TO VV DT NNS , PP VVD CD NN NN NP ( ) : DT NN NN NN NN NP NP : DT JJ NN IN JJ NN NNS	Fundamental	Basis	Positive
08-2008_6	Other researches attacked the problem of unsupervised extraction of world knowledge: Schubert ( ) reports a method for extracting general facts about the world from tree-banked Brown corpus	JJ VVZ VVN DT NN IN JJ NN IN NN NN NP ( ) VVZ DT NN IN VVG JJ NNS IN DT NN IN JJ NP NN	BackGround	GRelated	Neutral
08-2008_7	Schutze ( ) used bag-of-words contexts for sense discrimination	NP ( ) VVN NNS NNS IN NN NN	BackGround	GRelated	Neutral
08-2009_0	Another approach to fine-grained tagging captures grammatical structures with tree-based tags, such as "supertags" in the tree-adjoining grammar of Bangalore and Joshi ( )	DT NN TO JJ VVG NNS JJ NNS IN JJ NNS , JJ IN NN IN DT JJ NN IN NP CC NP ( )	BackGround	GRelated	Neutral
08-2009_1	Like previous Icelandic work ( ), morphological analyzers disambiguate words before statistical tagging in Arabic ( ) and Czech ( )	IN JJ JJ NN ( ) , JJ NNS VVP NNS IN JJ VVG IN NP ( ) CC JJ ( )	BackGround	GRelated	Neutral
08-2009_2	Given these challenges, the most successful tagger is IceTagger ( ), a linguistic rule based system with several linguistic resources: a morphological analyzer, a series of local rules and heuristics for handling PPs, verbs, and forcing agreement	VVN DT NNS , DT RBS JJ NN VBZ NN ( ) , DT JJ NN VVN NN IN JJ JJ NN DT JJ NN , DT NN IN JJ NNS CC NNS IN VVG NNS , NNS , CC VVG NN	BackGround	GRelated	Positive
08-2010_0	Our BoostedMERT should not be confused with other boosting algorithms such as ( )	PP$ NP MD RB VB VVN IN JJ VVG NNS JJ IN ( )	BackGround	GRelated	Neutral
08-2010_2	We used a standard phrase-based statistical MT system ( ) to generated N-best lists ( ) on  Developments,  Developments, and  Evaluation sub-sets	PP VVD DT JJ JJ JJ NP NN ( ) TO VVN JJ NNS ( ) IN NPS , NPS , CC NN NNS	Fundamental	Basis	Neutral
08-2011_0	These two models can be combined with the entity grid described by Lapata and Barzilay ( ) for significant improvement	DT CD NNS MD VB VVN IN DT NN NN VVN IN NP CC NP ( ) IN JJ NN	BackGround	GRelated	Positive
08-2011_0	These models typically view a sentence either as a bag of words ( ) or as a bag of entities associated with various syntactic roles ( )	DT NNS RB VVP DT NN CC IN DT NN IN NNS ( ) CC IN DT NN IN NNS VVN IN JJ JJ NNS ( )	BackGround	GRelated	Neutral
08-2011_0	Since the correct labeling depends on the coref-erence relationships between the NPs, we need some way to guess at this; we take all NPs with the same head to be coreferent, as in the non-coreference version of ( ) 2	IN DT JJ VVG VVZ IN DT NN NNS IN DT NP , PP VVP DT NN TO VV IN NN PP VVP DT NP IN DT JJ NN TO VB JJ , RB IN DT NN NN IN ( ) LS	Fundamental	Idea	Neutral
08-2011_0	As a baseline, we adopt the entity grid ( )	IN DT NN , PP VVP DT NN NN ( )	Fundamental	Basis	Neutral
08-2011_0	In the discrimination task ( ), a document is compared with a random permutation of its sentences, and we score the system correct ifitindicates the original as more coherent 4	IN DT NN NN ( ) , DT NN VBZ VVN IN DT JJ NN IN PP$ NNS , CC PP VVP DT NN JJ NNS DT NN RB RBR JJ CD	BackGround	SRelated	Neutral
08-2011_0	As mentioned, Barzilay and Lapata ( ) uses a coreference system to attempt to improve the entity grid, but with mixed results	IN VVN , NP CC NP ( ) VVZ DT NN NN TO VV TO VV DT NN NN , CC IN JJ NNS	BackGround	SRelated	Neutral
08-2011_1	Previous work has focused on the AIRPLANE corpus ( ), which contains short announcements of airplane crashes written by and for domain experts	JJ NN VHZ VVN IN DT NP NN ( ) , WDT VVZ JJ NNS IN NN NNS VVN IN CC IN NN NNS	BackGround	SRelated	Neutral
08-2011_2	Models of coherence have been used to impose an order on sentences for multidocument summarization ( ), to evaluate the quality of human-authored essays ( ), and to insert new information into existing documents ( )	NNS IN NN VHP VBN VVN TO VV DT NN IN NNS IN NN NN ( ) , TO VV DT NN IN JJ NNS ( ) , CC TO VV JJ NN IN JJ NNS ( )	BackGround	GRelated	Neutral
08-2011_4	Therefore we also test our systems on the task of insertion ( ), in which we remove a sentence from a document, then find the point of insertion which yields the highest coherence score	RB PP RB VVP PP$ NNS IN DT NN IN NN ( ) , IN WDT PP VV DT NN IN DT NN , RB VV DT NN IN NN WDT VVZ DT JJS NN NN	Fundamental	Basis	Neutral
08-2011_5	We construct a maximum-entropy classifier using syntactic and lexical features derived from Uryupina ( ), and a publicly available learning tool ( )	PP VV DT NN NN VVG JJ CC JJ NNS VVN IN NP ( ) , CC DT RB JJ VVG NN ( )	Fundamental	Basis	Neutral
08-2011_7	These patterns have been studied extensively, by linguists ( ) and in the field of coreference resolution	DT NNS VHP VBN VVN RB , IN NNS ( ) CC IN DT NN IN NN NN	BackGround	GRelated	Neutral
08-2011_7	Features such as full names, appositives, and restrictive relative clauses are associated with the introduction of unfamiliar entities into discourse ( )	NNS JJ IN JJ NNS , NNS , CC JJ JJ NNS VBP VVN IN DT NN IN JJ NNS IN NN ( )	BackGround	GRelated	Neutral
08-2011_7	Another issue is that NPs whose referents are familiar tend to resemble discourse-old NPs, even though they have not been previously mentioned ( )	DT NN VBZ IN/that NP WP$ NNS VBP JJ VVP TO VV JJ NP , RB IN PP VHP RB VBN RB VVN ( )	BackGround	SRelated	Neutral
08-2011_8	We use a model which probabilistically attempts to describe these preferences ( )	PP VVP DT NN WDT RB VVZ TO VV DT NNS ( )	Fundamental	Basis	Neutral
08-2011_8	(This takes more work than simply resolving the pronouns conditioned on the text.) The model of Ge et al.( ) provides the requisite probabilities: P (a i ,T i \d~ 1) =P (a i \h(a i ),m(a ,i)) Here h(a) is the Hobbs distance ( ), which measures distance between a pronoun and prospective antecedent, taking into account various factors, such as syntactic constraints on pronouns	NN VVZ JJR NN IN RB VVG DT NNS VVN IN DT NN DT NN IN NP NP NP ( ) VVZ DT JJ NN NN NP NP NP NP JJ JJ NN NN NP NP NP ) NN NN RB JJ VBZ DT NP NN ( ) , WDT VVZ NN IN DT NN CC JJ NN , VVG IN NN JJ NNS , JJ IN JJ NNS IN NNS	BackGround	SRelated	Neutral
08-2011_8	The model is trained using a small hand-annotated corpus first used in Ge et al.( )	DT NN VBZ VVN VVG DT JJ JJ NN RB VVN IN NP NP NP ( )	Fundamental	Basis	Neutral
08-2011_9	Centering theory ( ) describes additional constraints about which entities in a discourse can be pronominalized: if there are pronouns in a segment, they must include the backward-looking center	VVG NN ( ) VVZ JJ NNS IN WDT NNS IN DT NN MD VB JJ IN EX VBP NNS IN DT NN , PP MD VV DT JJ NN	BackGround	SRelated	Neutral
08-2011_10	As noted by studies since Hawkins ( ), there are marked syntactic differences between the two classes	IN VVN IN NNS IN NP ( ) , EX VBP JJ JJ NNS IN DT CD NNS	BackGround	SRelated	Neutral
08-2011_13	We evaluate our models using two tasks, both based on the assumption that a human-authored document is coherent, and uses the best possible ordering of its sentences (see Lapata ( ))	PP VVP PP$ NNS VVG CD NNS , CC VVN IN DT NN IN/that DT JJ NN VBZ JJ , CC VVZ DT RBS JJ VVG IN PP$ NNS NN NP ( NN	Fundamental	Basis	Neutral
08-2011_15	The system of Nenkova and McKeown ( ) works in the opposite direction	DT NN IN NP CC NP ( ) VVZ IN DT JJ NN	BackGround	GRelated	Neutral
08-2011_16	Classifiers in the literature include ( )	NNS IN DT NN VVP ( )	BackGround	GRelated	Neutral
08-2011_16	For the discourse-new classification task, the model's most important feature is whether the head word of the NP to be classified has occurred previously (as in Ng and Cardie ( ) and Vieira and Poesio ( ))	IN DT JJ NN NN , DT NNS RBS JJ NN VBZ IN DT NN NN IN DT NP TO VB VVN VHZ VVN RB NNS IN NP CC NP ( ) CC NP CC NP ( NN	BackGround	SRelated	Neutral
08-2011_20	This model outperforms a variety of word overlap and semantic similarity models, and is used as a component in the state-of-the-art system of Soricut and Marcu ( )	DT NN VVZ DT NN IN NN VVP CC JJ NN NNS , CC VBZ VVN IN DT NN IN DT JJ NN IN NP CC NP ( )	BackGround	SRelated	Positive
08-2011_20	Both of these models are very different from the lexical and entity-based models currently used for this task ( ), and are probably capable of improving the state of the art	CC IN DT NNS VBP RB JJ IN DT JJ CC JJ NNS RB VVN IN DT NN ( ) , CC VBP RB JJ IN VVG DT NN IN DT NN	BackGround	MRelated	Positive
08-2011_21	Our first model distinguishes discourse-new from discourse-old noun phrases, using features based on Uryupina ( )	PP$ JJ NN VVZ NN IN JJ NN NNS , VVG NNS VVN IN NP ( )	Fundamental	Basis	Neutral
08-2012_0	The b 3 scorer ( ) was proposed to overcome several shortcomings of the MUC scorer	DT SYM CD NN ( ) VBD VVN TO VV JJ NNS IN DT NP NN	BackGround	SRelated	Positive
08-2012_1	Other work on global models of coreference (as opposed to pairwise models) has included: Luo et al.( ) who used a Bell tree whose leaves represent possible partitionings of the mentions into entities and then trained a model for searching the tree; Mc-Callum and Wellner ( ) who defined several conditional random field-based models; Ng ( ) who took a reranking approach; and Culotta et al.( ) who use a probabilistic first-order logic model	JJ NN IN JJ NNS IN NN NNS VVN TO JJ NN VHZ JJ NP NP NP ( ) WP VVD DT NP NN WP$ NNS VVP JJ NNS IN DT VVZ IN NNS CC RB VVN DT NN IN VVG DT JJ NP CC NP ( ) WP VVD JJ JJ JJ JJ NN NP ( ) WP VVD DT JJ NN CC NP NP NP ( ) WP VVP DT JJ NN NN NN	BackGround	GRelated	Neutral
08-2012_2	Much work that followed improved upon this strategy, by improving the features ( ), the type of classifier ( ), and changing mention links to be to the most likely antecedent rather than the most recent positively labeled antecedent ( )	JJ NN WDT VVD VVN IN DT NN , IN VVG DT NNS ( ) , DT NN IN NN ( ) , CC VVG NN VVZ TO VB TO DT RBS JJ NN RB IN DT RBS JJ RB VVN NN ( )	BackGround	GRelated	Positive
08-2012_2	More recently, Denis and Baldridge ( ) utilized an integer linear programming (ILP) solver to better combine the decisions made by these two complementary classifiers, by finding the globally optimal solution according to both classifiers	RBR RB , NP CC NP ( ) VVD DT NN JJ NN NN NN TO RBR VV DT NNS VVN IN DT CD JJ NNS , IN VVG DT RB JJ NN VVG TO DT NNS	BackGround	GRelated	Positive
08-2012_2	When describing our model, we build upon the notation used by Denis and Baldridge ( )	WRB VVG PP$ NN , PP VVP IN DT NN VVN IN NP CC NP ( )	Fundamental	Basis	Neutral
08-2012_2	Prior work ( ) has generated training data for pairwise classifiers in the following manner	JJ NN ( ) VHZ VVN NN NNS IN JJ NNS IN DT VVG NN	BackGround	GRelated	Neutral
08-2012_2	The coref-ilp model of Denis and Baldridge ( ) took a different approach at test time: for each mention they would work backwards and add a link for all previous mentions which the classifier deemed coreferent	DT NN NN IN NP CC NP ( ) VVD DT JJ NN IN NN NN IN DT NN PP MD VV RB CC VV DT NN IN DT JJ VVZ WDT DT NN VVN NN	BackGround	SRelated	Neutral
08-2012_3	4 We added named entity (NE) tags to the data using the tagger of Finkel et al.( )	LS PP VVD VVN NN NN NNS TO DT NNS VVG DT NN IN NP NP NP ( )	Fundamental	Basis	Neutral
08-2012_4	In addition to the MUC and b 3 scorers, we also evaluate using cluster f-measure ( ), which is the standard f-measure computed over true/false coreference decisions for pairs of mentions; the Rand index ( ), which is pairwise accuracy of the clustering; and variation of information ( ), which utilizes the entropy of the clusterings and their mutual information (and for which lower values are better)	IN NN TO DT NP CC LS CD NNS , PP RB VV VVG NN NN ( ) , WDT VBZ DT JJ NN VVN IN JJ NN NNS IN NNS IN NN DT NP NN ( ) , WDT VBZ JJ NN IN DT NN CC NN IN NN ( ) , WDT VVZ DT NN IN DT NNS CC PP$ JJ NN NN IN WDT JJR NNS VBP JJ	Fundamental	Basis	Neutral
08-2012_8	Ng and Cardie ( ) and Ng ( ) highlight the problem of determining whether or not common noun phrases are anaphoric	NP CC NP ( ) CC NP ( ) VV DT NN IN VVG IN CC RB JJ NN NNS VBP JJ	BackGround	GRelated	Neutral
08-2012_13	Much recent work on coreference resolution, which is the task of deciding which noun phrases, or mentions, in a document refer to the same real world entity, builds on Soon et al.( )	JJ JJ NN IN NN NN , WDT VBZ DT NN IN VVG WDT NN NNS , CC VVZ , IN DT NN VVP TO DT JJ JJ NN NN , VVZ IN RB CC JJ ( )	BackGround	GRelated	Neutral
08-2012_13	Our feature set was simple, and included many features from ( ), including the pronoun, string match, definite and demonstrative NP, number and gender agreement, proper name and appositive features	PP$ NN NN VBD JJ , CC VVD JJ NNS IN ( ) , VVG DT NN , NN NN , JJ CC JJ NP , NN CC NN NN , JJ NN CC JJ NNS	Fundamental	Basis	Neutral
08-2012_13	Our Soon-style baseline used the same training and testing regimen as Soon et al.( )	PP$ NP NN VVD DT JJ NN CC NN NN IN RB CC JJ ( )	Fundamental	Basis	Neutral
08-2012_14	We also added part of speech (POS) tags to the data using the tagger of Toutanova et al.( ), and used the tags to decide if mentions were plural or singular	PP RB VVD NN IN NN NN NNS TO DT NNS VVG DT NN IN NP NP NP ( ) , CC VVD DT NNS TO VV IN NNS VBD JJ CC JJ	Fundamental	Basis	Neutral
08-2012_15	The MUC scorer ( ) is a popular coreference evaluation metric, but we found it to be fatally flawed	DT NP NN ( ) VBZ DT JJ NN NN JJ , CC PP VVD PP TO VB RB VVN	BackGround	SRelated	Negative
08-2013_0	System utterances were generated using a simple template-based algorithm and synthesised using the speech synthesis system Cerevoice ( ), which has been shown to be intelligible to older users ( )	NN NNS VBD VVN VVG DT JJ JJ NN CC VVD VVG DT NN NN NN NN ( ) , WDT VHZ VBN VVN TO VB JJ TO JJR NNS ( )	BackGround	SRelated	Positive
08-2013_1	Older people are a user group with distinct needs and abilities ( ) that present challenges for user modelling	JJR NNS VBP DT NN NN IN JJ NNS CC NNS ( ) IN/that JJ NNS IN NN VVG	BackGround	GRelated	Neutral
08-2013_2	They are also in line with findings of tests of deployed Interactive Voice Response systems with younger and older users ( ), which show the diversity of older people's behaviour	PP VBP RB IN NN IN NNS IN NNS IN VVN NP NP NP NNS IN JJR CC JJR NNS ( ) , WDT VVP DT NN IN JJR NNS NN	BackGround	SRelated	Neutral
08-2013_3	In order to learn good policies, the behaviour of the SUs needs to cover the range of variation seen in real users ( )	IN NN TO VV JJ NNS , DT NN IN DT NP VVZ TO VV DT NN IN NN VVN IN JJ NNS ( )	BackGround	GRelated	Neutral
08-2013_3	Our data comes from a fully annotated corpus of 447 interactions of older and younger users with a Wizard-of-Oz (WoZ) appointment scheduling system ( )	PP$ NN VVZ IN DT RB VVN NN IN CD NNS IN JJR CC JJR NNS IN DT NP NN NN NN NN ( )	Fundamental	Basis	Neutral
08-2013_3	Currently one of the standard methods for evaluating the quality of a SU is to run a user simulation on a real corpus and measure how often the action generated by the SU agrees with the action observed in the corpus ( )	RB CD IN DT JJ NNS IN VVG DT NN IN DT NP VBZ TO VV DT NN NN IN DT JJ NN CC VV WRB RB DT NN VVN IN DT NP VVZ IN DT NN VVD IN DT NN ( )	BackGround	GRelated	Neutral
08-2013_3	Given a history of system and user actions (n-1 actions) the SU generates an action based on a probability distribution learned from the training data ( )	VVN DT NN IN NN CC NN NNS JJ NN DT NP VVZ DT NN VVN IN DT NN NN VVN IN DT NN NNS ( )	BackGround	SRelated	Neutral
08-2013_3	The actions generated by our SUs were compared to the actions observed in the corpus using five metrics proposed in the literature ( ): perplexity (PP), precision, recall, expected precision and expected recall	DT NNS VVN IN PP$ NP VBD VVN TO DT NNS VVD IN DT NN VVG CD NNS VVN IN DT NN ( JJ NN NN , NN , NN , VVN NN CC VVN NN	Fundamental	Basis	Neutral
08-2013_4	A detailed description of the corpus design, statistics, and annotation scheme is provided in ( )	DT JJ NN IN DT NN NN , NNS , CC NN NN VBZ VVN IN ( )	BackGround	SRelated	Neutral
08-2013_4	There are 28 distinct user speech acts ( )	EX VBP CD JJ NN NN NNS ( )	BackGround	SRelated	Neutral
08-2013_5	This finding supports the principle of "inclusive design" ( ): designers should consider a wide range of users when developing a product for general use	DT NN VVZ DT NN IN JJ NN ( JJ NNS MD VV DT JJ NN IN NNS WRB VVG DT NN IN JJ NN	BackGround	SRelated	Positive
08-2013_6	The only statistical spoken dialogue system for older people we are aware of is Nursebot, an early application of statistical methods (POMDPs) within the context of a medication reminder system ( )	DT RB JJ VVN NN NN IN JJR NNS PP VBP JJ IN VBZ NP , DT JJ NN IN JJ NNS JJ IN DT NN IN DT NN NN NN ( )	BackGround	SRelated	Positive
08-2013_7	We then evaluate these models using standard metrics ( ) and compare our findings with the results of statistical corpus analysis	PP RB VV DT NNS VVG JJ NNS ( ) CC VV PP$ NNS IN DT NNS IN JJ NN NN	Fundamental	Basis	Neutral
08-2014_0	Finally, a linear model is trained using a variation of the averaged perceptron ( ) algorithm	RB , DT JJ NN VBZ VVN VVG DT NN IN DT VVN NN ( ) NN	Fundamental	Basis	Neutral
08-2014_1	We use a linear classifier trained with a regularized perceptron update rule ( ) as implemented in SNoW, ( )	PP VVP DT JJ NN VVN IN DT VVN NN NN NN ( ) RB VVN IN NP , ( )	Fundamental	Basis	Neutral
08-2014_2	( ) bootstrap with a classifier used interchangeably with an un-supervised temporal alignment method	( ) NN IN DT NN VVN RB IN DT JJ JJ NN NN	BackGround	GRelated	Neutral
08-2014_2	We evaluated our approach in two settings; first, we compared our system to a baseline system described in ( )	PP VVD PP$ NN IN CD NN JJ , PP VVD PP$ NN TO DT NN NN VVN IN ( )	Compare	Compare	Neutral
08-2014_3	Previous works usually take a generative approach, ( )	JJ VVZ RB VV DT JJ NN , ( )	BackGround	GRelated	Neutral
08-2014_4	The idea of selectively sampling training samples has been wildly discussed in machine learning theory ( ) and has been applied successfully to several NLP applications ( )	DT NN IN RB VVG NN NNS VHZ VBN RB VVN IN NN VVG NN ( ) CC VHZ VBN VVN RB TO JJ NP NNS ( )	BackGround	GRelated	Positive
08-2014_7	Other approaches exploit similarities in aligned bilingual corpora; for example, ( ) combine two unsupervised methods	JJ NNS VVP NNS IN VVN JJ NN IN NN , ( ) VV CD JJ NNS	BackGround	GRelated	Neutral
08-2015_0	DictEx We extend the phrase table with entries from a manually created dictionary - the English glosses of the Buckwalter Arabic morphological analyzer ( )	NP PP VVP DT NN NN IN NNS IN DT RB VVN NN : DT JJ NNS IN DT NP NP JJ NN ( )	Fundamental	Basis	Neutral
08-2015_1	Our transliteration system is rather simple: it uses the transliteration similarity measure described by Freeman et al.( ) to select a best match from a large list of possible names in English	PP$ NN NN VBZ RB JJ PP VVZ DT NN NN NN VVN IN NP NP NP ( ) TO VV DT JJS NN IN DT JJ NN IN JJ NNS IN NP	Fundamental	Basis	Neutral
08-2015_2	More details are available in a technical report ( )	JJR NNS VBP JJ IN DT JJ NN ( )	BackGround	SRelated	Neutral
08-2015_4	We tokenize using the Mada morphological disambiguation system ( ), and Tokan, a general Arabic tokenizer ( )	PP VVP VVG DT NP JJ NN NN ( ) , CC NP , DT JJ NP NN ( )	Fundamental	Basis	Neutral
08-2015_5	There is much work on name transliteration and its integration in larger MT systems ( )	EX VBZ JJ NN IN NN NN CC PP$ NN IN JJR NP NNS ( )	BackGround	GRelated	Neutral
08-2015_6	We decode using Pharaoh ( )	PP VVP VVG NN ( )	Fundamental	Basis	Neutral
08-2015_7	Word alignment is done with GIZA++ ( )	NN NN VBZ VVN IN NP ( )	Fundamental	Basis	Neutral
08-2015_8	The first 200 sentences in the 2002 MTEval test set were used for Minimum Error Training (MERT) ( )	DT JJ CD NNS IN DT CD NP NN NN VBD VVN IN NP NP NP NN ( )	BackGround	SRelated	Neutral
08-2015_9	Okuma et al.( ) describe a dictionary-based technique for translating OOV words in SMT	NP NP NP ( ) VV DT JJ NN IN VVG NP NNS IN NP	BackGround	GRelated	Neutral
08-2015_10	6 We report results in terms of case-insensitive 4-gram BLEU ( ) scores	CD PP VVP NNS IN NNS IN JJ NP NP ( ) NNS	Fundamental	Basis	Neutral
08-2015_11	This is especially true for languages with rich morphology such as Spanish, Catalan, and Serbian ( ) and Arabic ( )	DT VBZ RB JJ IN NNS IN JJ NN JJ IN NP , NP , CC JJ ( ) CC NP ( )	BackGround	GRelated	Neutral
08-2015_12	We address some of these challenges in our baseline system by removing all diacritics, normalizing Alif and Ya forms, and tokenizing Arabic text in the highly competitive Arabic Treebank scheme ( )	PP VVP DT IN DT NNS IN PP$ JJ NN IN VVG DT NNS , VVG NP CC NP NNS , CC VVG NP NN IN DT RB JJ NP NP NN ( )	Fundamental	Basis	Neutral
08-2015_13	All evaluated systems use the same surface trigram language model, trained on approximately 340 million words from the English Gigaword corpus ( ) using the SRILM toolkit ( )	DT VVN NNS VVP DT JJ NN NN NN NN , VVN IN RB CD CD NNS IN DT NP NP NN ( ) VVG DT NP NN ( )	Fundamental	Basis	Neutral
08-2015_14	Vilar et al.( ) address spelling-variant OOVs in MT through online re-tokenization into letters and combination with a word-based system	NP NP NP ( ) VV JJ NP IN NP IN JJ NN IN NNS CC NN IN DT JJ NN	BackGround	GRelated	Neutral
08-2015_15	Some previous approaches anticipate OOV words that are potentially morphologically related to in-vocabulary (INV) words ( )	DT JJ NNS VVP NP NNS WDT VBP RB RB VVN TO JJ JJ NNS ( )	BackGround	GRelated	Neutral
08-2016_1	Jewish Law documents written in Hebrew are known to be rich in ambiguous abbreviations ( )	JJ NN NNS VVN IN NP VBP VVN TO VB JJ IN JJ NNS ( )	BackGround	GRelated	Neutral
08-2016_1	In our previous research ( ), we developed a prototype abbreviation disambiguation system for Jewish Law documents written in Hebrew, without using any ML method	IN PP$ JJ NN ( ) , PP VVD DT NN NN NN NN IN JJ NN NNS VVN IN NP , IN VVG DT NP NN	BackGround	SRelated	Neutral
08-2016_1	The numerical sum of the numerical values attributed to the Hebrew letters forming the abbreviation ( )	DT JJ NN IN DT JJ NNS VVD TO DT JJ NNS VVG DT NN ( )	BackGround	SRelated	Neutral
08-2016_2	The one sense per discourse hypothesis (OS) was introduced by Gale et al.( )	DT CD NN IN NN NN NN VBD VVN IN NP NP NP ( )	BackGround	SRelated	Neutral
08-2016_3	Systems developed by Pakhomov ( ), Yu et al.( ) and Gaudan et al.( ) achieved 84% to 98% accuracy	NNS VVN IN NP ( ) , NP NP NP ( ) CC NP NP NP ( ) VVD CD TO CD NN	BackGround	GRelated	Positive
08-2016_7	Yosef ( )	NP ( )	NULL	NULL	NULL
08-2017_0	Active learning (AL) can be employed to reduce the costs of corpus annotation ( )	JJ VVG NN MD VB VVN TO VV DT NNS IN NN NN ( )	BackGround	GRelated	Neutral
08-2017_0	The easiest solution is to normalize by sentence length, as has been done previously ( )	DT JJS NN VBZ TO VV IN NN NN , RB VHZ VBN VVN RB ( )	BackGround	SRelated	Positive
08-2017_0	We follow Engelson & Dagan ( ) in the implementation of vote entropy for sentence selection using these models	PP VVP NP CC NP ( ) IN DT NN IN NN NN IN NN NN VVG DT NNS	Fundamental	Idea	Neutral
08-2017_1	In the context of parse tree annotation, Hwa ( ) estimates cost using the number of constituents needing labeling and Osborne & Baldridge ( ) use a measure related to the number of possible parses	IN DT NN IN VV NN NN , NP ( ) VVZ VV VVG DT NN IN NNS VVG VVG CC NP CC NP ( ) VV DT NN VVN TO DT NN IN JJ VVZ	BackGround	GRelated	Neutral
08-2017_2	One exception is ( ) (discussed later) which compares the cost of manual rule writing with AL-based annotation for noun phrase chunking	CD NN VBZ ( ) JJ NN WDT VVZ DT NN IN JJ NN VVG IN JJ NN IN NN NN NN	BackGround	GRelated	Neutral
08-2017_2	In contrast to the model presented by Ngai and Yarowsky ( ), which predicts monetary cost given time spent, this model estimates time spent from characteristics of a sentence	IN NN TO DT NN VVN IN NP CC NP ( ) , WDT VVZ JJ NN VVN NN VVD , DT NN VVZ NN VVN IN NNS IN DT NN	Compare	Compare	Neutral
08-2017_4	We also consider another selection algorithm introduced in ( ) that eliminates the overhead of entropy computations altogether by estimating per-sentence uncertainty with 1 ?P(|), where f is the Viterbi (best) tag sequence	PP RB VVP DT NN NN VVN IN ( ) WDT VVZ DT NN IN NN NNS RB IN VVG NN NN IN CD SENT NP , WRB SYM VBZ DT NP NN NN NN	Fundamental	Basis	Neutral
08-2017_5	In prior work, we describe such a cost model for POS annotation on the basis of the time required for annotation ( )	IN JJ NN , PP VVP PDT DT NN NN IN NP NN IN DT NN IN DT NN VVN IN NN ( )	BackGround	SRelated	Neutral
08-2017_6	Perhaps the best known are Query by Committee (QBC) ( ) and uncertainty sampling (or Query by Uncertainty, QBU) ( )	RB DT RBS JJ VBP NP IN NP NP ( ) CC NN NN NN NP IN NN , NP ( )	BackGround	GRelated	Neutral
08-2017_8	Our implementation of QBC employs a committee of three MEMM taggers to balance computational cost and diversity, following Tomanek et al.( )	PP$ NN IN NP VVZ DT NN IN CD NP NN TO VV JJ NN CC NN , VVG NP NP NP ( )	Fundamental	Idea	Neutral
08-2017_9	The features used in this work are typical for modern MEMM POS tagging and are mostly based on work by Toutanova and Manning ( )	DT NNS VVN IN DT NN VBP JJ IN JJ NP NP VVG CC VBP RB VVN IN NN IN NP CC NP ( )	Fundamental	Basis	Neutral
08-2018_0	Though there has been a growing interest in semi-supervised learning ( ), it is in an early phase of development	IN EX VHZ VBN DT VVG NN IN JJ NN ( ) , PP VBZ IN DT JJ NN IN NN	BackGround	GRelated	Neutral
08-2018_1	Previous domain resources include WordNet ( ) and HowNet ( ), among others	JJ NN NNS VVP NP ( ) CC NP ( ) , IN NNS	BackGround	GRelated	Neutral
08-2018_3	In this study, the 12 domains in Table 1 are used following ( ) (H&K hereafter) 1 	IN DT NN , DT CD NNS IN JJ CD VBP VVN VVG ( ) NN NP CD	Fundamental	Idea	Neutral
08-2018_4	Previous text categorization methods like Joachims ( ) and Schapire and Singer ( ) are mostly based on machine learning	JJ NN NN NNS IN NP ( ) CC NP CC NP ( ) VBP RB VVN IN NN NN	BackGround	GRelated	Neutral
08-2018_5	Ko and Seo ( ) automatically collect training data using a large amount of unlabeled data and a small amount of seed information	NP CC NP ( ) RB VV NN NNS VVG DT JJ NN IN JJ NNS CC DT JJ NN IN NN NN	BackGround	GRelated	Neutral
08-2018_6	This is consistent with Kornai et al.( ), who claim that only positive evidence matter in categorization	DT VBZ JJ IN NP NP NP ( ) , WP VVP IN/that RB JJ NN NN IN NN	Compare	Compare	Neutral
08-2018_7	Liu et al.( ) prepare representative words for each class, by which they collect initial training data to build classifier	NP NP NP ( ) VV JJ NNS IN DT NN , IN WDT PP VVP JJ NN NNS TO VV NN	BackGround	GRelated	Neutral
08-2018_8	Magnini et al.( ) show the effectiveness of domain information for WSD	NP NP NP ( ) VV DT NN IN NN NN IN NP	BackGround	GRelated	Neutral
08-2018_9	Piao et al.( ) use domain tags to extract MWEs	NP NP NP ( ) VV NN NNS TO VV NP	BackGround	GRelated	Neutral
08-2019_0	We have implemented a Mixture Model POMDP architecture as a multi-state version of the DIPPER "Information State Update" dialogue manager ( )	PP VHP VVN DT NN NP NP NN IN DT NNS NN IN DT NP NN NP NP NN NN ( )	BackGround	SRelated	Neutral
08-2020_0	An alternative is to apply automatically learned reordering rules to the test sentences before decoding ( )	DT NN VBZ TO VV RB VVD VVG NNS TO DT NN NNS IN VVG ( )	BackGround	SRelated	Neutral
08-2020_1	Our translation system is based on the CMU SMT decoder as described in ( )	PP$ NN NN VBZ VVN IN DT NP NP NN IN VVN IN ( )	Fundamental	Basis	Neutral
08-2020_2	We used the Pharaoh/Moses package ( ) to extract and score phrase pairs using the grow-diag-final extraction method	PP VVD DT NNS NN ( ) TO VV CC VVP NN NNS VVG DT JJ NN NN	Fundamental	Basis	Neutral
08-2020_3	To accelerate the training of word alignment models we implemented a distributed version of GIZA++ ( ), based on the latest version of GIZA++ and a parallel version developed at Peking University ( )	TO VV DT NN IN NN NN NNS PP VVD DT VVN NN IN NP ( ) , VVN IN DT JJS NN IN NP CC DT JJ NN VVN IN NP NP ( )	Fundamental	Basis	Neutral
08-2020_6	In this paper we report results using the BLEU metric ( ), however as the evaluation criterion in GALE is HTER ( ), we also report in TER ( )	IN DT NN PP VVP NNS VVG DT NP NN ( ) , RB IN DT NN NN IN NP VBZ JJ ( ) , PP RB VVP IN NP ( )	Fundamental	Basis	Neutral
08-2020_7	Every outgoing edge of a node is scored with the relative frequency of the pattern used on the following sub path (For details see ( ))	DT JJ NN IN DT NN VBZ VVN IN DT JJ NN IN DT NN VVN IN DT VVG NN NN NN NNS VVP ( NN	BackGround	SRelated	Neutral
08-2020_10	We trained separate open vocabulary language models for each source and interpolated them using the SRI Language Modeling Toolkit ( )	PP VVN JJ JJ NN NN NNS IN DT NN CC VVD PP VVG DT NP NP NP NP ( )	Fundamental	Basis	Neutral
08-2020_11	Our preprocessing steps include tokenization on the English side and for Chinese: automatic word segmentation using the revised version of the Stanford Chinese Word Segmenter 2 ( ) from 2007, replacement of traditional by simplified Chinese characters and 2-byte to 1-byte ASCII character normalization	PP$ NN NNS VVP NN IN DT JJ NN CC IN NP JJ NN NN VVG DT VVN NN IN DT NP NP NP NP CD ( ) IN CD , NN IN JJ IN VVN JJ NNS CC JJ TO JJ NP NN NN	Fundamental	Basis	Neutral
08-2020_12	In order to find an optimal set of weights, we use MER training as described in ( ), which uses rescoring of the top n hypotheses to maximize an evaluation metric like BLEU or TER	IN NN TO VV DT JJ NN IN NNS , PP VVP JJ NN IN VVN IN ( ) , WDT VVZ NN IN DT JJ NN NNS TO VV DT NN JJ IN NP CC NP	Fundamental	Basis	Neutral
08-2021_0	Note that our contributions in this paper could be applied to arbitrary lattice topologies.) For example, Bangalore et al.( ) show how to build a confusion network following a multistring alignment procedure of several MT outputs	NN IN/that PP$ NNS IN DT NN MD VB VVN TO JJ NN NN IN NN , NP NP NP ( ) VV WRB TO VV DT NN NN VVG DT JJ NN NN IN JJ NP NNS	BackGround	GRelated	Neutral
08-2021_1	The BLEU oracle sentences were found using the dynamic-programming algorithm given in Dreyer et al.( ) and measured using Philipp Koehn'seval-uation script	DT NP NN NNS VBD VVN VVG DT JJ NN VVN IN NP NP NP ( ) CC VVN VVG NP NP NN	Fundamental	Basis	Neutral
08-2022_0	In Model II, the semi-supervised setup, the training data is used to initialize the Expectation-Maximization (EM) algorithm ( ) and the unlabeled data, described in Section 3.1, updates the initial estimates	IN NP NP , DT JJ NN , DT NN NN VBZ VVN TO VV DT NN NN NN ( ) CC DT JJ NNS , VVN IN NP CD , NNS DT JJ NNS	Fundamental	Basis	Neutral
08-2022_1	225 words were selected for manual annotation as homograph or non-homograph by random sampling of words that were on the above list and used in prior psycholinguistic studies of homographs ( ) or on the Academic Word List ( )	CD NNS VBD VVN IN JJ NN IN NN CC NN IN JJ NN IN NNS WDT VBD IN DT JJ NN CC VVN IN JJ JJ NNS IN NNS ( ) CC IN DT NP NP NP ( )	BackGround	GRelated	Neutral
08-2022_2	However, making fine-grained sense distinctions for words with multiple closely-related meanings is a subjective task ( ), which makes it difficult and error-prone	RB , VVG JJ NN NNS IN NNS IN JJ JJ NNS VBZ DT JJ NN ( ) , WDT VVZ PP JJ CC JJ	BackGround	GRelated	Negative
08-2022_4	Lexical ambiguity resolution is an important research problem for the fields of information retrieval and machine translation ( )	JJ NN NN VBZ DT JJ NN NN IN DT NNS IN NN NN CC NN NN ( )	BackGround	GRelated	Positive
08-2022_6	Fine-grained sense distinctions aren't necessary for many tasks, thus a possibly-simpler alternative is lexical disambiguation at the level of homographs ( )	JJ NN NNS NN JJ IN JJ NNS , RB DT NN NN VBZ JJ NN IN DT NN IN NNS ( )	BackGround	GRelated	Neutral
08-2023_0	The above two work was further advanced by Bunescu and Mooney ( ) who argued that the information to extract a relation between two entities can be typically captured by the shortest path between them in the dependency graph	DT JJ CD NN VBD RBR JJ IN NP CC NP ( ) WP VVD IN/that DT NN TO VV DT NN IN CD NNS MD VB RB VVN IN DT JJS NN IN PP IN DT NN NN	BackGround	GRelated	Positive
08-2023_1	In the feature-based framework, Kambhatla ( ) employed ME models to combine diverse lexical, syntactic and semantic features derived from word, entity type, mention level, overlap, dependency and parse tree	IN DT JJ NN , NP ( ) VVN JJ NNS TO VV JJ JJ , JJ CC JJ NNS VVN IN NN , NN NN , NN NN , VVP , NN CC VV NN	BackGround	GRelated	Neutral
08-2023_2	Zelenko et al ( ) proposed a kernel over two parse trees, which recursively matched nodes from roots to leaves in a top-down manner	NP NP NP ( ) VVN DT NN IN CD VVP NNS , WDT RB VVD NNS IN NNS TO NNS IN DT NN NN	BackGround	GRelated	Neutral
08-2023_3	Based on his work, Zhou et al ( ) further incorporated the base phrase chunking information and semi-automatically collected country name list and personal relative trigger word list	VVN IN PP$ NN , NP NP NP ( ) RBR VVN DT JJ NN NN NN CC RB VVN NN NN NN CC JJ JJ NN NN NN	BackGround	GRelated	Neutral
08-2023_4	Later, Zhang et al ( ) developed a composite kernel that combined parse tree kernel with entity kernel and Zhou et al ( ) experimented with a context-sensitive kernel by automatically determining context-sensitive tree spans	RBR , NP NP NP ( ) VVN DT JJ NN WDT VVD VV NN NN IN NN NN CC NP NP NP ( ) VVN IN DT JJ NN IN RB VVG JJ NN NNS	BackGround	GRelated	Neutral
08-2023_5	Che et al ( ) defined an improved edit distance kernel over the original Chinese string representation around particular entities	NP NP NP ( ) VVN DT VVD VV NN NN IN DT JJ JJ NN NN IN JJ NNS	BackGround	GRelated	Positive
08-2024_1	It is not difficult to list all of those characters that have the same or similar pronunciations, e.g., "t^AI" and ""fAI", if we have a machine readable lexicon that provides information about pronunciations of characters and when we ignore special patterns for tone sandhi in Chinese ( )	PP VBZ RB JJ TO VV DT IN DT NNS WDT VHP DT JJ CC JJ NNS , FW , NN CC NN , IN PP VHP DT NN JJ NN WDT VVZ NN IN NNS IN NNS CC WRB PP VVP JJ NNS IN NN NN IN JJ ( )	BackGround	SRelated	Negative
08-2024_2	Bong-Foo Chu, selected a set of 24 basic elements in Chinese characters, and proposed a set of rules to decompose Chinese characters into elements that belong to this set of building blocks ( )	NP NP , VVD DT NN IN CD JJ NNS IN JJ NNS , CC VVD DT NN IN NNS TO VV JJ NNS IN NNS WDT VVP TO DT NN IN NN NNS ( )	BackGround	SRelated	Neutral
08-2024_4	Figure 4 illustrates possible layouts of the components in Chinese characters that were adopted by the Cangjie method ( )	NN CD VVZ JJ NNS IN DT NNS IN JJ NNS WDT VBD VVN IN DT NP NN ( )	Fundamental	Basis	Neutral
08-2024_5	There are more than 22000 different characters in large corpus of Chinese documents ( ), so directly computing the similarity between images of these characters demands a lot of computation	EX VBP JJR IN CD JJ NNS IN JJ NN IN JJ NNS ( ) , RB RB VVG DT NN IN NNS IN DT NNS NNS DT NN IN NN	BackGround	SRelated	Neutral
08-2024_6	Taft, Zhu, and Peng ( ) investigated the effects of positions of radicals on subjects' lexical decisions and naming responses	NP , NP , CC NP ( ) VVN DT NNS IN NNS IN NNS IN NP JJ NNS CC VVG NNS	BackGround	GRelated	Neutral
08-2024_7	Yeh and Li ( ) studied how similar characters influenced the judgments made by skilled readers of Chinese	NP CC NP ( ) VVN WRB JJ NNS VVD DT NNS VVN IN JJ NNS IN NP	BackGround	GRelated	Neutral
08-2025_0	In this case, the "best" answer may be chosen by the votes, or alternatively by automatically predicting answer quality (e.g., ( ) or ( ))	IN DT NN , DT JJ NN MD VB VVN IN DT NNS , CC RB IN RB VVG NN NN NN , ( ) CC ( NN	BackGround	GRelated	Neutral
08-2025_0	Category Features: We hypothesized that user behavior (and asker satisfaction) varies by topical question category, as recently shown in reference ( )	JJ NP PP VVN WDT NN NN NN NN NN VVZ IN JJ NN NN , RB RB VVN IN NN ( )	BackGround	SRelated	Neutral
08-2025_2	While information seeker satisfaction has been studied in ad-hoc IR context (see ( ) for an overview), previous studies have been limited by the lack of realistic user feedback	IN NN NN NN VHZ VBN VVN IN NN NN NN NN ( ) IN DT NN , JJ NNS VHP VBN VVN IN DT NN IN JJ NN NN	BackGround	SRelated	Neutral
08-2025_3	In our recent work ( ) we have introduced a general model for predicting asker satisfaction in community question answering	IN PP$ JJ NN ( ) PP VHP VVN DT JJ NN IN VVG NN NN IN NN NN NN	BackGround	SRelated	Neutral
08-2025_3	For more detailed treatment of user interactions in CQA see ( )	IN JJR JJ NN IN NN NNS IN NP VVP ( )	BackGround	SRelated	Neutral
08-2025_3	We now briefly review our ASP (Asker Satisfaction Prediction) framework that learns to classify whether a question has been satisfactorily answered, originally introduced in ( )	PP RB RB VV PP$ NP NN NP NP NN WDT VVZ TO VV IN DT NN VHZ VBN RB VVN , RB VVN IN ( )	Fundamental	Basis	Positive
08-2025_4	Furthermore, while automatic complex QA has been an active area of research, ranging from simple modification to factoid QA technique (e.g., ( )) to knowledge intensive approaches for specialized domains, the technology does not yet exist to automatically answer open domain, complex, and subjective questions	RB , IN JJ NN NP VHZ VBN DT JJ NN IN NN , VVG IN JJ NN TO NN NP NN NN , ( NN TO NN JJ NNS IN JJ NNS , DT NN VVZ RB RB VV TO RB VV JJ NN , JJ , CC JJ NNS	BackGround	GRelated	Neutral
08-2025_5	Classification Algorithms: We experimented with a variety of classifiers in the Weka framework ( )	NN NN PP VVD IN DT NN IN NNS IN DT NP NN ( )	Fundamental	Basis	Neutral
08-2026_0	However more recent results have shown that it can indeed improve parser performance ( )	RB RBR JJ NNS VHP VVN IN/that PP MD RB VV NN NN ( )	BackGround	GRelated	Positive
08-2026_0	Two previous papers would seem to address this issue: the work by Bacchiani et al.( ) and McClosky et al.( )	CD JJ NNS MD VV TO VV DT NN DT NN IN NP NP NP ( ) CC NP NP NP ( )	BackGround	GRelated	Neutral
08-2026_1	2 A close second (1% behind) was the parser of Bikel ( )	LS DT JJ JJ NN NN VBD DT NN IN NP ( )	BackGround	GRelated	Neutral
08-2026_2	While self-training has worked in several domains, the early results on self-training for parsing were negative ( )	IN NN VHZ VVN IN JJ NNS , DT JJ NNS IN NN IN VVG VBD JJ ( )	BackGround	GRelated	Negative
08-2026_3	Section three describes our main experiment on standard test data ( )	NN CD VVZ PP$ JJ NN IN JJ NN NNS ( )	Fundamental	Basis	Neutral
08-2026_3	Clegg and Shepherd ( ) do not provide separate precision and recall numbers	NP CC NP ( ) VVP RB VV JJ NN CC NN NNS	BackGround	SRelated	Negative
08-2026_4	In contrast, the out-of-vocabulary rate of biomedical abstracts given the same lexicon is significantly higher at about 25% ( )	IN NN , DT JJ NN IN JJ NNS VVN DT JJ NN VBZ RB JJR IN IN CD ( )	BackGround	GRelated	Neutral
08-2026_4	Lease and Charniak ( ) achieve their results using small amounts of hand-annotated biomedical part-of-speech-tagged data and also explore other possible sources or information	NP CC NP ( ) VV PP$ NNS VVG JJ NNS IN JJ JJ JJ NNS CC RB VV JJ JJ NNS CC NN	BackGround	MRelated	Neutral
08-2026_5	Marcus et al., 1993)) and then self-training on a second type of data in order to adapt the parser to the second domain	NP NP NP , JJ CC RB VVG IN DT JJ NN IN NNS IN NN TO VV DT NN TO DT JJ NN	BackGround	GRelated	Neutral
08-2026_6	So, for example, McClosky et al.( ) found that the data from the handannotated WSJ data should be considered at least five times more important than NANC data on an event by event level	RB , IN NN , NP NP NP ( ) VVD IN/that DT NNS IN DT JJ NP NNS MD VB VVN IN JJS CD NNS RBR JJ IN NP NNS IN DT NN IN NN NN	BackGround	SRelated	Neutral
08-2026_8	However, several very good current parsers were not available when this paper was written (e.g., the Berkeley Parser ( ))	RB , JJ RB JJ JJ NNS VBD RB JJ WRB DT NN VBD VVN NN , DT NP NP ( NN	BackGround	GRelated	Neutral
08-2026_9	Bacchiani and Roark train the Roark parser ( ) on trees from the Brown treebank and then self-train and test on data from Wall Street Journal	NP CC NP VVP DT NP NN ( ) IN NNS IN DT NP NN CC RB NN CC NN IN NNS IN NP NP NP	BackGround	GRelated	Neutral
08-2026_10	1 Steedman et al.( ) generally found that self-training does not work, but found that it does help if the baseline results were sufficiently bad	CD NP NP NP ( ) RB VVD IN/that NN VVZ RB VV , CC VVD IN/that PP VVZ VV IN DT NN NNS VBD RB JJ	BackGround	GRelated	Neutral
08-2027_0	Speech repairs are common in spontaneous speech - one study found 30% of dialogue turns contained repairs ( ) and another study found one repair every 4.8 seconds ( )	NN NNS VBP JJ IN JJ NN : CD NN VVD CD IN NN VVZ VVN NNS ( ) CC DT NN VVD CD NN DT CD NNS ( )	BackGround	GRelated	Neutral
08-2027_2	Recent advances in recognizing spontaneous speech with repairs ( ) have used parsing approaches on transcribed speech to account for the structure inherent in speech repairs at the word level and above	JJ NNS IN VVG JJ NN IN NNS ( ) VHP VVN VVG NNS IN VVN NN TO VV IN DT NN JJ IN NN NNS IN DT NN NN CC JJ	BackGround	GRelated	Positive
08-2027_2	I I Figure 1 : Standard tree repair structure, with -UNF propagation as in ( ) shown in brackets	PP PP NP CD : JJ NN NN NN , IN NN NN IN IN ( ) VVN IN NNS	Fundamental	Basis	Neutral
08-2027_2	Figure 1 also shows, in brackets, the augmented annotation used by Hale et al.( )	NN CD RB VVZ , IN NNS , DT JJ NN VVN IN NP NP NP ( )	Fundamental	Basis	Neutral
08-2027_2	The approach used by ( ) works because the information about the transition to an error state is propagated up the tree, in the form of the -UNF tags	DT NN VVN IN ( ) VVZ IN DT NN IN DT NN TO DT NN NN VBZ VVN RP DT NN , IN DT NN IN DT NN NNS	BackGround	SRelated	Neutral
08-2027_2	With this representation, the problem noticed by Hale and colleagues ( ) has been solved in a different way, by incrementally building up left-branching rather than right-branching structure, so that only a single special error rule is required at the end of the constituent	IN DT NN , DT NN VVN IN NP CC NNS ( ) VHZ VBN VVN IN DT JJ NN , IN RB VVG RP VVG RB IN VVG NN , RB IN/that RB DT JJ JJ NN NN VBZ VVN IN DT NN IN DT NN	BackGround	SRelated	Neutral
08-2027_2	To make a fair comparison to the CYK baseline of ( ), the recognizer was given correct part-of-speech tags as input along with words	TO VV DT JJ NN TO DT NP NN IN ( ) , DT NN VBD VVN JJ NN NNS IN NN IN IN NNS	Compare	Compare	Neutral
08-2027_3	The evaluation of this system was performed on the Switchboard corpus, using the mrg annotations in directories 2 and 3 for training, and the files sw4004.mrg to sw4153.mrg in directory 4 for evaluation, following Johnson and Charniak( )	DT NN IN DT NN VBD VVN IN DT NP NN , VVG DT NN NNS IN NNS CD CC CD IN NN , CC DT NNS NN TO NN IN NN CD IN NN , VVG NP CC NP )	Fundamental	Idea	Neutral
08-2027_3	The TAG system ( ) achieves a higher EDIT-F score, largely as a result of its explicit tracking of overlapping words between reparanda and alterations	DT NP NN ( ) VVZ DT JJR NP NN , RB IN DT NN IN PP$ JJ NN IN JJ NNS IN NNS CC NNS	BackGround	SRelated	Positive
08-2027_4	In order to obtain a linguistically plausible right-corner transform representation of incomplete constituents, the Switchboard corpus is subjected to a pre-process transform to introduce binary-branching nonterminal projections, and fold empty categories into nonterminal symbols in a manner similar to that proposed by Johnson ( ) and Klein and Manning ( )	IN NN TO VV DT RB JJ NN VV NN IN JJ NNS , DT NN NN VBZ VVN TO DT NN VV TO VV VVG JJ NNS , CC VV JJ NNS IN JJ NNS IN DT NN JJ TO DT VVN IN NP ( ) CC NP CC NP ( )	Fundamental	Idea	Neutral
08-2027_8	The speech repair terminology used here follows that of Shriberg ( )	DT NN NN NN VVN RB VVZ IN/that IN NP ( )	Fundamental	Idea	Neutral
08-2027_9	The incomplete constituent categories created by the right corner transform are similar in form and meaning to non-constituent categories used in Combinatorial Categorial Grammars (CCGs) ( )	DT JJ NN NNS VVN IN DT JJ NN VV VBP JJ IN NN CC VVG TO JJ NNS VVN IN NP NP NP NN ( )	Fundamental	Idea	Neutral
08-2028_0	The classifiers generally mimic human judgements in that accuracy is much lower in the three-way classification task - a pattern concurring with past observations (cf.Esuli and Sebastiani ( ); Andreevskaia and Bergler ( ))	DT NNS RB VVP JJ NNS IN DT NN VBZ RB JJR IN DT JJ NN NN : DT NN VVG IN JJ NNS NNS CC NP ( NP NP CC NP ( NN	Compare	Compare	Positive
08-2028_0	Non-neutral adjectives were extracted from WordNet and assigned fuzzy sentiment category member-ship/centrality scores and tags in Andreevskaia and Bergler ( )	JJ NNS VBD VVN IN NP CC VVN JJ NN NN NN NNS CC NNS IN NP CC NP ( )	BackGround	GRelated	Neutral
08-2028_1	The semi-supervised learning method in Esuli and Sebastiani ( ) involves constructing a training set of non-neutral words using WordNet synsets, glosses and examples by iteratively adding syn- and antonyms to it and learning a term classifier on the glosses of the terms in the training set	DT JJ VVG NN IN NP CC NP ( ) VVZ VVG DT NN VVN IN JJ NNS VVG NP NNS , NNS CC NNS IN RB VVG NN CC NNS TO PP CC VVG DT NN NN IN DT NNS IN DT NNS IN DT NN NN	BackGround	GRelated	Neutral
08-2028_2	Esuli and Sebastiani ( ) used the method to cover objective (n) cases	NP CC NP ( ) VVN DT NN TO VV JJ NN NNS	BackGround	GRelated	Neutral
08-2028_3	Hatzivassiloglou and McKeown ( ) clustered adjectives into (+) and (-) sets based on conjunction constructions, weighted similarity graphs, minimum-cuts, supervised learning, and clustering	NP CC NP ( ) VVN NNS IN NN CC NN NNS VVN IN NN NNS , JJ NN NNS , NNS , JJ NN , CC VVG	BackGround	GRelated	Neutral
08-2028_5	Kaji and Kitsuregawa ( ) describe a method for harvesting sentiment words from non-neutral sentences extracted from Japanese web documents based on structural layout clues	NP CC NP ( ) VV DT NN IN NN NN NNS IN JJ NNS VVN IN JJ NN NNS VVN IN JJ NN NNS	BackGround	GRelated	Neutral
08-2028_6	Since not all constituents are of equal importance, the sentiment salience of each subconstituent is estimated using a subset of the grammatical polarity rankings and compositional processes proposed in Moilanen and Pulman ( )	IN RB DT NNS VBP IN JJ NN , DT NN NN IN DT NN VBZ VVN VVG DT NN IN DT JJ NN NNS CC JJ NNS VVN IN NP CC NP ( )	Fundamental	Basis	Neutral
08-2028_7	Riloff et al.( ) mined subjective nouns from unannotated texts with two bootstrapping algorithms that exploit lexico-syntactic extraction patterns and manually-selected subjective seeds	NP NP NP ( ) VVN JJ NNS IN JJ NNS IN CD NN NNS WDT VVP JJ NN NNS CC JJ JJ NNS	BackGround	GRelated	Neutral
08-2028_8	Takamura et al.( ) apply to words' polarities a physical spin model inspired by the behaviour of electrons with a (+) or (-) direction, and an iterative term-neighbourhood matrix which models magnetisation	NP NP NP ( ) VV TO NP NNS DT JJ NN NN VVN IN DT NN IN NNS IN DT NN CC NN NN , CC DT JJ NN NN WDT NNS NN	BackGround	GRelated	Neutral
08-2028_9	A popular, more general unsupervised method was introduced in Turney and Littman ( ) which induces the polarity of a word from its Pointwise Mutual Information (PMI) or Latent Semantic Analysis (LSA) scores obtained from a web search engine against a few paradigmatic (+) and (-) seeds	DT JJ , RBR JJ JJ NN VBD VVN IN NP CC NP ( ) WDT VVZ DT NN IN DT NN IN PP$ NP NP NP NP CC NP NP NP NN NNS VVN IN DT NN NN NN IN DT JJ JJ NN CC NN NNS	BackGround	GRelated	Positive
08-2028_10	Strong adjectival subjectivity clues were mined in Wiebe ( ) with a distributional similarity-based word clustering method seeded by hand-labelled annotation	JJ JJ NN NNS VBD VVN IN NP ( ) IN DT JJ JJ NN VVG NN VVN IN JJ NN	BackGround	GRelated	Neutral
08-2029_0	This is the reason why our kernels on linguistic structures improve it by 63%, which is a remarkable result for an IR task ( )	DT VBZ DT NN WRB PP$ NNS IN JJ NNS VV PP IN CD , WDT VBZ DT JJ NN IN DT JJ NN ( )	BackGround	SRelated	Positive
08-2029_1	Question Answering (QA) is an IR task where the major complexity resides in question processing and answer extraction ( ) rather than document retrieval (a step usually carried out by off-the shelf IR engines)	NN NN NN VBZ DT JJ NN WRB DT JJ NN VVZ IN NN NN CC NN NN ( ) RB IN NN NN NN NN RB VVD RP IN JJ NN NN NN	BackGround	SRelated	Neutral
08-2029_2	representations: (1)linear kernels on the bag-of-words (BOW) or bag-of-POS-tags (POS) features, (2)the String Kernel (SK) ( ) on word sequences (WSK) and POStag sequences (POSSK), (3)the Syntactic Tree Kernel (STK) ( ) on syntactic parse trees (PTs), (4)the Shallow Semantic Tree Kernel (SSTK) ( ) and the Partial Tree Kernel (PTK) ( ) on PASs	JJ JJ NNS IN DT NNS NN CC NNS NN NNS , VVG NP NP NN ( ) IN NN NNS NN CC NN NNS JJ , JJ NP NP NP NN ( ) IN JJ VVP NNS JJ , JJ NP NP NP NP NN ( ) CC DT NP NP NP NN ( ) IN NP	Fundamental	Basis	Neutral
08-2029_6	Then, two PAS-based trees: Shallow Semantic Trees for SSTK and Shallow Semantic Trees for PTK, both based on PropBank structures ( ) are automatically generated by our SRL system ( )	RB , CD JJ NN NP NP NP IN NP CC NP NP NP IN NP , CC VVN IN NP NNS ( ) VBP RB VVN IN PP$ NP NN ( )	Fundamental	Basis	Neutral
08-2029_7	The experimental datasets were created by submitting the 138 TREC 2001 test questions labeled as "description" in ( ) to our basic QA system, YourQA ( ) and by gathering the top 20 answer paragraphs	DT JJ NNS VBD VVN IN VVG DT CD NP CD NN NNS VVN IN NN IN ( ) TO PP$ JJ NP NN , NP ( ) CC IN VVG DT JJ CD NN NNS	Fundamental	Basis	Neutral
08-2029_8	As a kernel operator, we applied the sum between kernels 5 that yields the joint feature space of the individual kernels ( )	IN DT NN NN , PP VVD DT NN IN NNS CD WDT VVZ DT JJ NN NN IN DT JJ NNS ( )	Fundamental	Basis	Neutral
08-2029_9	Although typical methods are based exclusively on word similarity between query and answer, recent work, e.g.( ) has shown that shallow semantic information in the form of predicate argument structures (PASs) improves the automatic detection of correct answers to a target question	IN JJ NNS VBP VVN RB IN NN NN IN NN CC NN , JJ NN , FW ( ) VHZ VVN IN/that JJ JJ NN IN DT NN IN JJ NN NNS NN VVZ DT JJ NN IN JJ NNS TO DT NN NN	BackGround	GRelated	Positive
08-2030_0	Since the dictionary we use, BAMA ( ), also includes diacritics (orthographic marks not usually written), we extend this approach to the diacritization task in ( )	IN DT NN PP VVP , NP ( ) , RB VVZ NNS JJ NNS RB RB JJ , PP VVP DT NN TO DT NN NN IN ( )	Fundamental	Basis	Neutral
08-2030_2	Hajic et al.( ) implement the approach of Hajic ( ) for Arabic	NP NP NP ( ) VV DT NN IN NP ( ) IN NP	BackGround	GRelated	Neutral
08-2030_3	Hajic ( ) is the first to use a dictionary as a source of possible morphological analyses (and hence tags) for an inflected word form	NP ( ) VBZ DT JJ TO VV DT NN IN DT NN IN JJ JJ NNS NN RB JJ IN DT VVN NN NN	BackGround	GRelated	Positive
08-2030_4	We use an implementation of a Downhill Simplex Method in many dimensions based on the method developed by Nelder and Mead ( ) to tune the weights applied to each feature	PP VVP DT NN IN DT NP NP NP IN JJ NNS VVN IN DT NN VVN IN NP CC NP ( ) TO VV DT NNS VVN TO DT NN	Fundamental	Basis	Neutral
08-2030_5	We also build 4-gram lexeme models using an open-vocabulary language model with Kneser-Ney smoothing, by means of the SRILM toolkit ( )	PP RB VVP JJ NN NNS VVG DT JJ NN NN IN NP VVG , IN NNS IN DT NP NN ( )	Fundamental	Basis	Neutral
08-2030_6	These 19 features consist of the 14 morphological features shown in Figure 1, which MADA predicts using 14 distinct Support Vector Machines trained on ATB3-Train (as defined by Zitouni et al.( ))	DT CD NNS VVP IN DT CD JJ NNS VVN IN NP CD , WDT NP VVZ VVG CD JJ NP NP NP VVN IN NP NNS VVN IN NP NP NP ( NN	Fundamental	Basis	Neutral