Papers reviewed and determined not to be word norm studies. Use the flag icon to report errors or suggest re-inclusion.
18265 papers
A grammatical method of combining two kinds of speech repair cues is presented. One cue, prosodic disjuncture, is detected by a decision tree-based ensemble classifier that uses acoustic cues to identify where normal prosody seems to be interrupted (Lickley, 1996). The other cue, syntactic parallelism, codifies the expectation that repairs continue a syntactic category that was left unfinished in the reparandum (Levelt, 1983). The two cues are combined in a Treebank PCFG whose states are split using a few simple tree transformations. Parsing performance on the Switchboard and Fisher corpora suggests that these two cues help to locate speech repairs in a synergistic way.
In this paper, we attempt to automatically annotate the Penn Chinese Treebank with semantic dependency structure. Ini-tially a small portion of the Penn Chinese Treebank was man-ually annotated with headword and semantic dependency re-lations. An initial investigation is then done using a Naive Bayesian Classifier and some handcrafted rules. The results show that the algorithms and proposed approach are effective at determining semantic dependency structure automatically. The Naive Bayesian Classifier makes a good baseline algo-rithm for future research.
Grammars can be induced from treebanks for a potential variety (e.g., parsing), but this task faces two major problems. One is the existence of annotation er-rors in the treebank which can affect the final system performance (cf. [4]). The other is that the sheer number of rules is overwhelming for most applications. As
Abstract When survey respondents rate the quality of life (QoL) associated with a health condition, they must not only evaluate the health condition itself, but must also interpret the meaning of the rating scale in order to assign a specific value. The way that respondents approach this task depends on subjective interpretations, resulting in inconsistent results across populations and tasks. In particular, patients and non-patients often give very different ratings to health conditions, a discrepancy that raises questions about the objectivity of either groups’ evaluations. In this study, we found that the perspective of the raters (i.e., their own current health relative to the health conditions they rated) influences the way they distinguish between different health states that vary in severity. Consistent with prospect theory, a mild and a severe lung disease scenario were rated quite differently by lung disease patients whose own health falls between the two scenarios, whereas healthy non-patients, whose own health was better than both scenarios, rated the two scenarios as much more similar. In addition, we found that the context of the rating task influences the way participants distinguish between mild and severe scenarios. Both patients and non-patients gave less distinct ratings to the two scenarios when each were presented in isolation than when they were presented alongside other scenarios that provided contextual information about the possible range of severity for lung disease. These results raise continuing concerns about the reliability and validity of subjective QoL ratings, as these ratings are highly sensitive to differences between respondent groups and the particulars of the rating task.
We built a morphological analyzer, which can be freely used by anyone for research purpose. In order to build a pratical system, a dictionary with reasonable size is necessary. The initial dictionary is built from the Penn Chinese Treebank corpus v4.0 and contains only 33,438 entries. Since the initial dictionary is quite small, unknown word detection methods are applied to a huge raw text in order to extract new words to be added into the system dictionary. We have successfully constructed a dictionary with 120,769 entries. Finally, we propose a two-layer morphological analyzer to cater for two sets of outputs. The first layer produces the minimal segmentation units defined by us, and the second layer transforms the output of the first layer to the original segmentation units defined by Penn Chinese Treebank.
PURPOSE: To determine observer performance in the detection of multiple sclerosis (MS) lesions on magnetic resonance (MR) images of the brain and to assess the dependence of observer performance on lesion size, parenchymal location, pulse sequence, and supratentorial versus infratentorial level. MATERIALS AND METHODS: This HIPAA-compliant protocol was approved by the institutional review board, and previously acquired MR data from a healthy volunteer and a patient with MS were used to derive parameter maps, with waiver of informed consent. Parameter maps and image simulator software were used to generate 320 phantom brain images with simulated supratentorial and infratentorial MS lesions. Images were displayed with T2-weighting or fluid-attenuated inversion recovery (FLAIR) contrast. Four readers independently evaluated the images, rating lesions on a five-point certainty scale. Observer performance was measured by using the area under the alternative free-response receiver operating characteristic curve (A(1)), and significance was determined with the z test. RESULTS: Pooled A(1) scores were significantly better for FLAIR imaging (0.96 +/- 0.01 [standard error]) than for T2-weighted MR imaging (0.89 +/- 0.04) supratentorially (P =.05) but were similar for FLAIR imaging (0.90 +/- 0.06) and T2-weighted MR imaging (0.88 +/- 0.05) infratentorially. A(1) scores for cortical, deep white matter, and periventricular lesions were 0.93 +/- 0.05, 0.97 +/- 0.02, and 0.89 +/- 0.04, respectively, for FLAIR imaging and 0.77 +/- 0.06, 0.99 +/- 0.01, and 0.89 +/- 0.05, respectively, for T2-weighted MR imaging. FLAIR scores were significantly higher than T2-weighted scores for cortical lesions. Linear correlation was found between A(1) and lesion size (r = 0.5). CONCLUSION: Supratentorially, performance was better with FLAIR imaging than with T2-weighted MR imaging. Infratentorially, performance was moderate with both modalities. Observers did better with FLAIR imaging in the detection of cortical lesions, and performance improved with increasing lesion size.
Based on the analysis of the usage and the syntactic function of Chinese punctuations,this paper proposes a new hierarchical approach to parse the long Chinese sentences.In traditional parsing approaches,the parsing procedure is performed in an one-level way and the punctuation marks are not specially treated.Correspondingly,in our approach,the complex long Chinese sentences are broken into sub-sentences or units(say 'units' hereafter) by using punctuation marks with special functions,so that the original whole sentence is parsed unit by unit.This idea of 'divide-and-conquer' greatly reduces the difficulty in the traditional parsing approaches to recognize the syntactic relationship between the sub-sentences and phrases or inside the sub-sentences or phrases.And also,in our approach,the grammatical rules with punctuation marks and their probabilities are extracted from the large scale treebank,which are very beneficial to the syntactic disambiguation.Our experimental results have shown that comparing with the traditional Chart parsing algorithm,our approach can significantly reduce the time consumption and the numbers of ambiguous edges,and get about 7% of the correct rate and the recall rate increasing while parsing long Chinese sentences.
In recent years, research in parsing has extended in several new directions. One of these directions is concerned with parsing languages other than English. Treebanks have become available for many European languages, but also for Arabic, Chinese, or Japanese. However, it was shown that parsing results on these treebanks depend on the types of treebank annotations used [, ]. Another direction in parsing research is the development of dependency parsers. Dependency parsing profits from the non-hierarchical nature of dependency relations, thus lexical information can be included in the parsing process in a much more natural way. Especially machine learning based approaches are very successful (cf. e.g. [12, 13]). The results achieved by these dependency parsers are very competitive although comparisons are difficult because of the differences in annotation. For English, the Penn Treebank [11] has been converted to dependencies. For this version, Nivre et al. [14] report an accuracy rate of 86.3%, as compared to an F-score of 2.1 for Charniak’s parser [1]. The Penn Chinese Treebank [1 ] is also available in a constituent and a dependency representations. The best results reported for parsing experiments with this treebank give an F-score of 81.8 for the constituent version [2] and. % accuracy for the dependency version [14]. The general trend in comparisons between constituent and dependency parsers is that the dependency parser performs slightly worse than the constituent parser. The only exception occurs for German, where F-scores for constituent plus grammatical function parses range between 51.4 and 5.3, depending on the treebank, NEGRA [1 ] or TuBa-D/Z [1 ]. The dependency parser based on a converted version of Tuba-D/Z, in contrast, reached an accuracy of 3.4% [14], i.e. 12 percent points better than the best constituent analysis including grammatical functions.
This paper explores several important issues in developing syntactically annotated Korean corpora for higher-level language processing, including semantic-discourse parsing, question-answering, machine translation, information retrieval, etc.In particular, we compare the Penn Korean Treebank (PKT) and the Korean Treebank of the 21st Century Sejong Project (ST) and discuss four critical issues in syntactic annotation.We argue for the use of more sophisticated morphosyntactic information, and based on our comparative study, we propose revisions in the syntactic annotation schemes of the existing Korean Treebanks in order to improve the quality of annotated corpora and their usability both for conducting theoretical research and for developing computational tools.The results of our comparative study reveal four significant issues in syntactic annotations: the syntactic analysis of verbal complexes, the hierarchical structure of noun phrases, the representation of traces, and the marking of zero elements.These factors may trigger erroneous syntactic representations for certain linguistic phenomena, and they may increase difficulties in data search and lessen reliability in computational processing.Thus, evaluating and improving the syntactic annotation of Treebanks is an important task for aspects of both theoretical and computational linguistics.
Nous presentons, ici, une implementation d’un systeme qui n’extrait pas seulement une grammaire lexicalisee (LTAG), mais aussi une grammaire LTAG avec traits (FB-LTAG) a partir d’un corpus arbore. Nous montrons les experimentations pratiques ou nous extrayons les grammaires TAG a partir du Sejong Treebank pour le coreen. Avant tout, les 57 etiquettes syntaxiques et les analyses morphologiques dans le corpus SJTree nous permettent d’extraire les traits syntaxiques automatiquement. De plus, nous modifions le corpus pour l’extraction d’une grammaire lexicalisee et convertissons les grammaires lexicalisees en schemas d’arbre pour resoudre le probleme de la couverture lexicale limitee des grammaires lexicalisees extraites.
In this work, we apply statistical learning algorithms to Lexicalized Tree Adjoining Grammar (LTAG) parsing, as an effort toward statistical analysis over deep structures. LTAG parsing is a well known hard problem. Statistical methods successfully applied to LTAG parsing could also be used in many other structure prediction problems in NLP. For the purpose of achieving accurate and efficient LTAG parsing, we will investigate two aspects of the problem, the data structure and the algorithm. 1. We introduce LTAG-spinal, a variant of LTAG with very desirable linguistic, computational and statistical properties. It can be shown that LTAG-spinal with adjunction constraints is weakly equivalent to the traditional LTAG. For the purpose of statistical processing, we extract an LTAG-spinal treebank from the Penn Treebank with Propbank annotation. 2. We not only explore various parsing strategies, but also investigate the reranking approach. (a) We first propose a left-to-right incremental parser for LTAG-spinal, as an attempt to dynamically incorporate supertagging and dependency analysis. A perceptron like discriminative learning algorithm is used for training. We further investigate a bidirectional dependency parser for LTAG-spinal, in order to overcome the limitation of left-to-right processing. We propose a novel algorithm for graph-based incremental construction, and apply this algorithm to LTAG style dependency parsing. (b) We also explore learning algorithms for parse reranking, as well as other NLP problems, e.g. Machine Translation. We propose a novel reranking strategy, Ordinal Regression with Uneven Margins (ORUM), which achieves state-of-the-art performance on parse reranking for CFG parsing and MT reranking. To sum up, we have accomplished the following achievements. (i) A new formalism, LTAG-spinal, which is weakly equivalent to LTAG. (ii) An LTAG-spinal Treebank extracted from the PTB with the Propbank annotation. (iii) A left-to-right incremental parser for LTAG-spinal. (iv) A bidirectional LTAG-spinal dependency parser. (v) A novel graph-based incremental construction algorithm, which could be applied to many structure prediction problem in NLP, e.g. semantic role labeling. (vi) A novel discriminative reranking algorithm, ORUM, which has been successfully applied to parse reranking as well as other tasks, e.g. MT reranking.
Abstract The aim of this study was to examine associations between children's temperament, parent–child goodness-of-fit, and the emotional content of parent–child conversations about past events. Fifty one New Zealand 5- and 6-year-old children and their parents discussed 4 emotional past events. Parents rated children's temperament along 15 dimensions associated with effortful control, extroversion, and negative affect. Parents then rated their own expectations of children's temperament, from which parent–child goodness-of-fit was assessed. Children rated as higher in effortful control were involved in more emotional past-event conversations. Children whose negative affect ratings corresponded more closely with parental expectations were also involved in more emotional conversations. Our findings provide preliminary support for an association between temperament and narrative content.
In this paper, we present a system which can extract syntactic feature structures from a Korean Treebank (Sejong Treebank) to develop a Feature-based Lexicalized
In this paper, we present an alternative approach to the problem of contextual relevance feedback in web-based information retrieval. Our approach utilises a rich contextual model that exploits a user's implicit and explicit data. Each user's implicit data are gathered from their Internet search histories on their local machine. The user's explicit data are captured from a lexical database, a shared contextual knowledge base and domain-specific concepts using data mining techniques and a relevance feedback approach. This data is later used by our approach to modify queries to more accurately reflect the user's interests as well as to continually build the user's contextual profile and a shared contextual knowledge base. Finally, the approach retrieves personalised or contextual search results from the search engine using the modified/expanded query. Preliminary experiments indicate that our approach has the potential to not only aid in the contextual relevance feedback but also contribute towards the long term goal of intelligent relevance feedback in web-based information retrieval.
The amygdala is closely linked to basal ganglia circuitry and plays a key role in danger detection and fear-potentiated startle. Based on recent findings of amygdalar abnormalities in Parkinson's disease, we hypothesized that non-demented patients with this illness would show blunted reactivity during aversive/unpleasant events, as indexed by diminished emotional modulation of the startle eyeblink response. To test this hypothesis, 23 idiopathic patients with Parkinson's disease and 17 controls viewed standardized sets of aversive, pleasant and neutral pictures for 6 s each. During this time, white noise bursts (50 ms, 95 db) were binaurally presented to elicit startle eyeblink responses, measured from electrodes over the orbicularis oculi. After viewing each picture, subjects provided ratings of valence and arousal. The Parkinson's disease patients were in the early to middle stages of their disease, not demented or depressed, and were tested 'on' dopaminergic medication. The two groups were similar in age, education, gender and cognitive screening status. The control group had larger startle responses when viewing negative, aversive pictures than neutral or pleasant pictures. As predicted, startle enhancement during aversive pictures was significantly muted in the Parkinson's disease patients. This blunting was not due to abnormalities in the mechanics of the startle eyeblink per se. Nor was it related to depression symptoms, medications (psychotropics), or failure to perceive/appreciate the negative meaning of aversive pictures (i.e. normal valence ratings). Reduced startle reactivity in the disease group was related to disease severity (Hoehn-Yahr) and occurred in the context of reduced arousal ratings of aversive pictures. These findings of blunted startle reactivity add to the literature on emotional changes associated with Parkinson's disease. The basis for this muted reactivity is unknown but may involve an amygdala-based translational defect whereby the results of cognitive appraisal are not appropriately transcoded into somato-motor-arousal responses normally associated with an aversive motivational state. This may arise from faulty dopaminergic gating of the amygdala, resulting in 'inhibition' of the amygdala in the manner described by Marowsky et al. (Marowsky A, Yanagawa Y, Obata K, Vogt E. Neuron 2005; 48: 1025-37). More broadly, the findings of muted reactivity to aversive stimuli may reflect a 'bradylimbic' affective disturbance in patients with Parkinson's disease. Future studies are needed to address whether the physiologic blunting observed here might be a useful correlate of apathy.
We propose a method for labelling prepositional phrases according to two different semantic role classifications, as contained in the Penn treebank and the CoNLL 2004 Semantic Role Labeling data set. Our results illustrate the difficulties in determining preposition semantics, but also demonstrate the potential for PP semantic role labelling to improve the performance of a holistic semantic role labelling system.
This document describes the information used for summarization-inspired temporal-relation extraction [Dorr and Gaasterland, 2007]. We present a set of tense/aspect extraction templates that are applied to a Penn Treebank-style analysis of the input sentence. We also present an analysis of tense-pair combinations for different temporal connectives based on a corpus analysis of complex tense structures in Treebank-3. Finally, we include analysis charts and temporal relation tables for all combinations of intervals/points for each legal BTS combinations.
The trees in the Penn Treebank have a standard representation that involves complete balanced bracketing. In this article, an alternative for this standard representation of the tree bank is proposed. The proposed representation for the trees is loss-less, but it reduces the total number of brackets by 28%. This is possible by omitting the redundant pairs of special brackets that encode initial and final embedding, using a technique proposed by Krauwer and des Tombe (1981). In terms of the paired brackets, the maximum nesting depth in sentences decreases by 78%. The 99.9% coverage is achieved with only five non-top levels of paired brackets. The observed shallowness of the reduced bracketing suggests that finite-state based methods for parsing and searching could be a feasible option for tree bank processing.
research-article Free Access Share on The treebanks used in the shared taskCoNLL-X '06: Proceedings of the Tenth Conference on Computational Natural Language LearningJune 2006 Pages 165Online:08 June 2006Publication History 0citation107DownloadsMetricsTotal Citations0Total Downloads107Last 12 Months35Last 6 weeks0 Get Citation AlertsNew Citation Alert added!This alert has been successfully added and will be sent to:You will be notified whenever a record that you have chosen has been cited.To manage your alert preferences, click on the button below.Manage my AlertsNew Citation Alert!Please log in to your account Save to BinderSave to BinderCreate a New BinderNameCancelCreateExport CitationPublisher SiteeReaderPDF
We exploit the resources in the Arabic Treebank (ATB) for the novel task of automatically creating lexical semantic verb classes for Modern Standard Arabic (MSA). Verbs are clustered into groups that share semantic elements of meaning as they exhibit similar syntactic behavior. The results of the clustering experiments are compared with a gold standard set of classes, which is approximated by using the noisy English translations provided in the ATB to create Levin-like classes for MSA. The quality of the clusters is found to be sensitive to the inclusion of information about lexical heads of the constituents in the syntactic frames, as well as parameters of the clustering algorithm. The best set of parameters yields an Fβ=1 score of 0.501, compared to a random baseline with an Fβ=1 score of 0.37.
We describe experiments in parsing the German TIGER Treebank. In parsing the complete treebank, 86.44 % of the sentences receive full parses; 13.56 % receive fragment parses. We discuss the methods used to enhance coverage and parsing quality and we present an evaluation on a gold standard, to our knowledge the first one for a deep grammar of German. Considering the selection performed by our current version of a stochastic disambiguation component, we achieve an f-score of 84.2%, the upper and lower bounds being 87.4% and 82.3 % respectively. 1.
In this paper, we construct a biomedical semantic role labeling (SRL) system that can be used to facilitate relation extraction. First, we construct a proposition bank on top of the popular biomedical GENIA treebank following the PropBank annotation scheme. We only annotate the predicate-argument structures (PAS's) of thirty frequently used biomedical predicates and their corresponding arguments. Second, we use our proposition bank to train a biomedical SRL system, which uses a maximum entropy (ME) model. Thirdly, we automatically generate argument-type templates which can be used to improve classification of biomedical argument types. Our experimental results show that a newswire SRL system that achieves an F-score of 86.29% in the newswire domain can maintain an F-score of 64.64% when ported to the biomedical domain. By using our annotated biomedical corpus, we can increase that F-score by 22.9%. Adding automatically generated template features further increases overall F-score by 0.47% and adjunct arguments (AM) F-score by 1.57%, respectively.
While much of the research and labor in treebanks has focused on modern languages, recent scholarship has also seen the rise of treebanks for historical languages as well, such as Middle English (Kroch and Taylor [15]), Early Modern English (Kroch et al. [16]), Old English (Taylor et al. [28]), Early New High German (Demske et al. [11]) and Medieval Portuguese (Rocio et al. [27]). Like their modern counterparts, these historical treebanks serve two distinct ends and often two different audiences: they provide crucial datasets for NLP projects such as automatic parsing and grammar induction while also providing a valuable corpus for scholars researching the state of a language and its progression across time. Historical treebanks, however, also offer one additional benefit over modern treebanks: they provide an annotated set of texts that scholars actually care about. When linguists of modern languages base theories on corpus evidence, their analysis is generally directed toward the language at large; few, if any, pore over the Wall Street Journal examining its use of an arcane literary device. If the corpus is Vergil, however, we do. The sheer volume of Latin texts available electronically1 not to mention the enormous mass still locked in print is much larger than the small community of scholars and students who can read it. This alone justifies a treebank as a resource for those attempting to learn the language, but it also highlights the need for automatic methods of parsing and machine translation. To this end a Latin treebank will well serve the NLP community, which has a long history of applying such research to modern languages.2 Classical scholars, however, largely operate on a fixed canon of texts. The value of a treebank for them is not so much in training
In Corpus Linguistics: Readings in a Widening Discipline, Geoffrey Sampson and Diana McCarthy present a diverse yet accessible collection of forty-two previously published papers, which introduce, discuss and exploit the methods of corpus linguistics. The volume can be divided into those papers that investigate language based on the analysis of corpora and those papers that discuss how a corpus should be compiled and analysed. The most straightforward introduction to corpus construction is John Sinclair’s (1987) paper Corpus Creation, which considers the significance of a variety of factors such as medium, corpus size and sample size. A more technical treatment of representativeness in corpus design is found in a 1992 paper by Douglas Biber. Papers that discuss actual corpora include W. Nelson Francis’s 1965 description of the Brown Corpus, the first electronic corpus of English; Burnage and Dunlop’s outline of the compilation and annotation of the 100-million-words British National Corpus; and Tent and Mugler’s 1996 defence of the Corpus of Fijian English, a component of the International Corpus of English. The Editors also include Adam Kilgarriff’s 2001 paper Web as Corpus – an approach to corpus linguistics which will undoubtedly become more common in future research. The majority of papers included in this collection, however, report the results of research based on corpus analyses, and the majority of these papers report the results of grammatical corpus studies. The volume opens with an excerpt from Charles C. Fries’s 1957 The Structure of English – an early empirical description of English grammar, based on a manual analysis of a corpus of recorded telephone conversations. The collection also includes F.G.A.M. Aarts’ 1971 study which demonstrates that English object noun phrases tend to be more complex than subject noun phrases – a result that challenges the assumption, inherent in standard phrase structure rules, that the position of a noun phrase is structurally insignificant. The collection also includes more recent papers which discuss grammatical issues such as treebanks and automatic parsing techniques. Geoffrey Sampson, an editor of the collection, includes an introductory paper, which argues for the importance of treebanks – a corpus where each sentence is associated with a parse tree. The construction of the Penn
Algorithms designed to support users in retrieving relevant information base their relevance computations on user profiles, in which representations of the users interests are maintained. This paper focuses on the use of supervised machine learning techniques to induce user profiles for Intelligent Information Access. The access must be personalized by profiles allowing users to retrieve information on the basis of conceptual content. To address this issue, we propose a method to learn sense-based user profiles based on WordNet, a lexical database. I. WORDNET-BASED DOCUMENT REPRESENTATION Personalization is an important method for digital libraries to take a more active role in dynamically tailoring its information and service offer to individuals. Novel solutions for personalized information access exploit machine learning algorithms to induce a structured model of a user’s interests, referred to as user profile, from text documents. We propose a strategy to induce semantic user profiles in which keywords are replaced with their meanings, defined in lexicons or ontologies. We adopted WordNet [1] as a reference lexicon. As regards document representation, we propose a model called bag of synsets (BOS), in which the senses corresponding to the words in the documents are considered as features. The adopted sense repository is WordNet (version 1.7.1). In WordNet, nouns, verbs, adjectives and adverbs are organized into synsets (synonym sets), each representing one underlying lexical concept. The problem of determining which of the senses of an ambiguous word is invoked in a particular use of the word (word sense disambiguation, WSD) has to be solved [2]. We propose a WSD strategy based on the idea that semantic similarity between synsets a and b is inversely proportional to the distance between them in the WordNet IS-A hierarchy, measured by the number of nodes in the path from a to b. More details about the WSD procedure are reported in [3]. A document is mapped into a list of WordNet synsets according to the following three rules: a) each monosemous word w in a slot of a document d is mapped into the corresponding WordNet synset; b) for each pair of words 〈noun, noun〉 or 〈adjective, noun〉, a search in WordNet is made in order to verify if at least one synset exists for the bigram 〈w1, w2〉. In the positive case, the WSD strategy is applied on the bigram, otherwise it is applied separately on w1 and w2; c) each polysemous unigram w is disambiguated using a set of words surrounding w as the context of w. The WSD strategy has been used to process documents in the EachMovie dataset, a collection made of 1,628 movie descriptions1. The BOS representation of a document consists of a list of synsets recognized from the words in the document. II. THE ITEM RECOMMENDER PROFILING SYSTEM ITem Recommender (ITR) is a content-based profiling system able to induce semantic user profiles as naive Bayesian classifiers by learning from documents represented according to the BOS model [4]. Given a set of documents in which each document is labeled by a specific user with a rating representing his degree of interest on that document, the system is able to learn the user profiles as a Bayesian text classifier able to classify a new document as interesting or uninteresting for that user. Naive Bayes is a probabilistic approach to inductive learning. The learned probabilistic model estimates the a posteriori probability, P (cj |di), of document di belonging to class cj. This estimation is based on the a priori probability, P (cj), i.e. the probability of observing a document in class cj, P (di|cj), that is the probability of observing document di given cj, and P (di), the probability of observing the instance di at all. Using these probabilities, Bayesian classifiers apply Bayes theorem to calculate P (cj |di). To classify a document di, the class with the highest probability is selected. As a working model for the naive Bayes classifier, we use the multinomial event model [5]: P (cj |di) = P (cj) ∏
In this paper, we describe an empirical study of Chinese chunking on a corpus, which is extracted from UPENN Chinese Treebank-4 (CTB4). First, we compare the performance of the state-of-the-art machine learning models. Then we propose two approaches in order to improve the performance of Chinese chunking. 1) We propose an approach to resolve the special problems of Chinese chunking. This approach extends the chunk tags for every problem by a tag-extension function. 2) We propose two novel voting methods based on the characteristics of chunking task. Compared with traditional voting methods, the proposed voting methods consider long distance information. The experimental results show that the SVMs model outperforms the other models and that our proposed approaches can improve performance significantly.
Abstract Attempting to automatically learn to identify verb complements from natural language corpora without the help of sophisticated linguistic resources like grammars, parsers or treebanks leads to a significant amount of noise in the data. In machine learning terms, where learning from examples is performed using class-labelled feature-value vectors, noise leads to an imbalanced set of vectors: assuming that the class label takes two values (in this work complement/non-complement), one class (complements) is heavily underrepresented in the data in comparison to the other. To overcome the drop in accuracy when predicting instances of the rare class due to this disproportion, we balance the learning data by applying one-sided sampling to the training corpus and thus by reducing the number of non-complement instances. This approach has been used in the past in several domains (image processing, medicine, etc) but not in natural language processing. For identifying the examples that are safe to remove, we use the value difference metric, which proves to be more suitable for nominal attributes like the ones this work deals with, unlike the Euclidean distance, which has been used traditionally in one-sided sampling. We experiment with different learning algorithms which have been widely used and their performance is well known to the machine learning community: Bayesian learners, instance-based learners and decision trees. Additionally we present and test a variation of Bayesian belief networks, the COr-BBN (Class-oriented Bayesian belief network). The performance improves up to 22% after balancing the dataset, reaching 73.7% f-measure for the complement class, having made use only a phrase chunker and basic morphological information for preprocessing.
This paper presents the English valency lexicon EngValLex, built within the Functional Generative Description framework. The form of the lexicon, as well as the process of its semi-automatic creation is described. The lexicon describes valency for verbs and also includes links to other lexical sources, namely PropBank. Basic statistics about the lexicon are given. The lexicon will be later used for annotation of the Wall Street Journal section of the Penn Treebank in Praguian formalisms.
We present an improved approach for learning dependency parsers from tree-bank data. Our technique is based on two ideas for improving large margin training in the context of dependency parsing. First, we incorporate local constraints that enforce the correctness of each individual link, rather than just scoring the global parse tree. Second, to cope with sparse data, we smooth the lexical parameters according to their underlying word similarities using Laplacian Regularization. To demonstrate the benefits of our approach, we consider the problem of parsing Chinese treebank data using only lexical features, that is, without part-of-speech tags or grammatical categories. We achieve state of the art performance, improving upon current large margin approaches.
The eye-blink startle reflex can be modulated by attentional and emotional processes. The reflex is facilitated during stimuli that engage attention. A linear pattern of emotional modulation has also been consistently demonstrated: the reflex is facilitated during unpleasant stimuli and attenuated during pleasant stimuli. However, during anticipation of pleasant or unpleasant stimuli it is unclear whether emotion or attention drives startle reflex modulation. This study used a differential learning procedure to investigate whether startle modulation during anticipation of a salient stimulus reflected emotional or attentional processes. In acquisition, a CS+ was paired with a pleasant or unpleasant US and a CS- was presented alone. In extinction, blink startle magnitude was measured during CS+ and CS-. Post-acquisition valence ratings and affective priming showed that CS+ had acquired the same affective value as the pleasant or unpleasant US with which it was paired. No differences in modulation of blink startle reflexes during pleasant CS+ and unpleasant CS+ were found throughout extinction. Blink startle facilitation occurred during CS+ but not CS- across the first third of extinction. Thus, attentional rather than emotional processes appeared to facilitate blink startle during anticipation of salient stimuli.
There are many methods to improve performances of statistical parsers. Among them, resolving structural ambiguities is a major task. In our approach, the parser produces a set of n-best trees based on a feature-extended PCFG grammar and then selects the best tree structure based on association strengths of dependency word-pairs. However, there is no sufficiently large Treebank producing reliable statistical distributions of all word-pairs. This paper aims to provide a self-learning method to resolve the problems. The word association strengths were automatically extracted and learned by parsing a giga-word corpus. Although the automatically learned word associations were not perfect, the built structure evaluation model improved the bracketed f-score from 83.09 % to 86.59%. We believe that the above iterative learning processes can improve parsing performances automatically by learning word-dependence knowledge continuously from web. 1.
We describe several improvements to the method of treebank-based LFG induction for Spanish from the Cast3LB treebank (O’Donovan et al., 2005). We discuss the different categories of problems encountered and present the solutions adopted. Some of the problems involve a simple adoption of existing linguistic analyses, as in our treatment of clitic doubling and null subjects. In other cases there is no standard LFG account for the phenomenon \nwe wish to model and we adopt a compromise, conservative solution. This is exemplified by our treatment of Spanish periphrastic constructions. In yet another case, the less configurational nature of Spanish means that the LFG annotation algorithm has to rely mostly on Cast3LB function tags, and consequently a reliable method of adding those tags to parse trees had to be developed. This method achieves over 6% improvement over the baseline for the \nCast3LB-function-tag assignment task, and over 3% improvement over the baseline for LFG f-structure construction from function-tag-enriched trees.
This report explores the question of compatibility between annotation projects including translating annotation formalisms to each other or to common forms. Compatibility issues are crucial for systems that use the results of multiple annotation projects. We hope that this report will begin a concerted effort in the field to track the compatibility of annotation schemes for part of speech tagging, time annotation, treebanking, role labeling and other phenomena.
Approximately 60 kinds of lexical relations have been recognized in languages of the world (Grimes & Grimes, 1993). In this paper, I present evidence for a wide range of lexical relations in the Ilokano language. Grimes explains the meaning of lexical relations in terms of the way two words are related but differ in meaning, giving examples such as write and writer, row and rower. I will exemplify some of the types of lexical relations attested in Ilokano: 1) Verbs with an incorporated nominal, e.g. ag-diram'os 'to wash one's face' where the implied noun is 'face'; aginnaw 'wash dishes', implied noun, 'dishes'. There is no word for 'face' in agdiram'os, nor word for 'dishes' in aginnaw. 2) Derived nouns expressing an agentive relation, e.g. from the verb agsugal 'to gamble' the derived noun is mannugal 'gambler', agsurat 'to write', mannurat 'writer'. 3) Reduplication of a noun describing a condition of that noun, e.g. saka 'foot', saka-saka 'barefoot'; ima 'hand', ima-ima 'emptyhanded'. 4) Derived verbs denoting animal vocalization, e.g. aso 'dog', agtaol (phonation) 'to bark', ul'ul'ol (onomatopoeia); 5) Derived verbs denoting a quantum, e.g. sangalilig a sua 'one section of a pomelo'; 6) Complements, e.g. biag ken patay 'life and death'; 7) Derived verbs denoting function, e.g. karayan 'river', agayos 'to flow', sabong flower', agukrad 'to bloom'. I will then show how these derivations are handled in the Ilokano Lexical Database where they are listed making use of the band format.
Motivationally relevant stimuli have been shown to receive prioritized processing compared to neutral stimuli at distinct processing stages. This effect has been related to the evolutionary importance of rapidly detecting dangers and potential rewards and has been shown to be modulated by the distance between an organism and a faced stimulus. Similarly, recent studies showed degrees of emotional modulation of autonomic responses and subjective arousal ratings depending on stimulus size. In the present study, affective modulation of pictures presented in different sizes was investigated by measuring event-related potentials during a two-choice categorization task. Results showed significant emotional modulation across all sizes at both earlier and later stages of processing. Moreover, affective modulation of earlier processes was reduced in smaller compared to larger sizes, whereas no changes in affective modulation were observed at later stages.
Most of the work on treebank-based statistical parsing exclusively uses the Wall-Street-Journal part of the Penn treebank for evaluation purposes. Due to the presence of this quasi-standard, the question of to which degree parsing results depend on the properties of treebanks was often ignored. In this paper, we use two similar German treebanks, TüBa-D/Z and NeGra, and investigate the role that different annotation decisions play for parsing. For these purposes, we approximate the two treebanks by gradually taking out or inserting the corresponding annotation components and test the performance of a standard PCFG parser on all treebank versions. Our results give an indication of which structures are favorable for parsing and which ones are not.
We have introduced a new methodology that maps designs to human perceptions. Perceptions are adjectives/adverbs, phrases or sentences expressed in natural language. We used the lexical database WordNet to compute semantical relationships (or distances) between these perceptions. We partitioned the set of perceptions into k clusters that represent the classes for a further classification task. We have developed a new classifier called “structural hidden Markov model” (SHMM) that combines probability and distances in a seamless way. SHMM enables to learn and predict user perceptions given object designs. We have applied this approach to Kansei engineering in order to map car external contours (shapes) to customer perceptions. The accuracy obtained using the SHMM is 90%. This model has outperformed the neural network and the k-nearest-neighbor classifiers.
Sentence similarity measures play an increasingly important role in text-related research and applications in areas such as text mining, Web page retrieval, and dialogue systems. Existing methods for computing sentence similarity have been adopted from approaches used for long text documents. These methods process sentences in a very high-dimensional space and are consequently inefficient, require human input, and are not adaptable to some application domains. This paper focuses directly on computing the similarity between very short texts of sentence length. It presents an algorithm that takes account of semantic information and word order information implied in the sentences. The semantic similarity of two sentences is calculated using information from a structured lexical database and from corpus statistics. The use of a lexical database enables our method to model human common sense knowledge and the incorporation of corpus statistics allows our method to be adaptable to different domains. The proposed method can be used in a variety of applications that involve text knowledge representation and discovery. Experiments on two sets of selected sentence pairs demonstrate that the proposed method provides a similarity measure that shows a significant correlation to human intuition
This report contains the guidelines for the annotation of discourse relations in the Penn Discourse Treebank (http://www.seas.upenn.edu/~pdtb), PDTB. Discourse relations in the PDTB are annotated in a bottom up fashion, and capture both lexically realized relations as well as implicit relations. Guidelines in this report are provided for all aspects of the annotation, including annotation explicit discourse connectives, implicit relations, arguments of relations, senses of relations, and the attribution of relations and their arguments. The report also provides descriptions of the annotation format representation.
Grammar induction is one of the most important research areas of the natural language processing. The lack of a large Treebank, which is required in supervised grammar induction, in some natural languages such as Persian encouraged us to focus on unsupervised methods. We have found the Inside-Outside algorithm, introduced by Lari and Young, as a suitable platform to work on, and augmented IO with a history notion. The result is an improved unsupervised grammar induction method called History-based IO (HIO). Applying HIO to two very divergent natural languages (i.e., English and Persian) indicates that inducing more conditioned grammars improves the quality of the resultant grammar. Besides, our experiments on ATIS and WSJ show that HIO outperforms most current unsupervised grammar induction methods.
In the present paper, we examined the effects of autobiographically induced mood and music on emotional evaluations of and psychophysiological responses to music in 48 subjects. Participants listened to music after a mood induction. Both music and induction varied on the dimensions of valence (pleasant - unpleasant) and arousal (high - low). During mood induction and listening to music, psychophysiological responses were measured continuously to assess physiological arousal (indexed by electrodermal activity) and the valence of the emotional state (indexed by facial muscle activity) of the participant. After listening to music, participants evaluated the music using pictorial scales for valence and arousal. As expected, subjects were in a more positive emotional state during listening to pleasant than unpleasant music and also evaluated the music more positively after a pleasant compared to an unpleasant pre-existing mood. As also expected, high-arousal music and pre-existing mood generated both higher physiological arousal and higher arousal ratings compared to low-arousal pre-existing mood and music. We found no support for the principle of mood-congruency, which posits that individuals preferentially process emotional stimuli that are congruent in emotional tone with their current mood state.
CIDOC-CRM is a new standard for encoding a wide range of information for Cultural Heritage (CH). At present, existing CH collections are stored using all sorts of formats, sometimes proprietary, often defined roughly, which \nmakes it difficult to share or access heterogeneous information among the CH community. There is a need for a tool to map diverse formats into CIDOC-CRM, assisted by another tool using intelligent language technology to help the mapping whenever fields are underspecified or loosely described, both tools being complementary. In some cases, it may even be better to build fragments of a CIDOC database directly from informal descriptions in natural \nlanguage only, as the CH community may be reluctant to switch to new formats of data entry. Therefore, this paper focus primarily on the mapping of CH data described in natural language into CIDOC-CRM triples, the building blocks of the full CIDOC-CRM ontology. The methods exploits the propositional nature of CIDOC-CRM triples. Using WordNet as a lexical database and the WEB as corpus, we first extract triples from examples provided in the CIDOC-CRM literature, and then from text describing the medieval city of Wolfenbüttel. We show the strong points of the system and suggest where and how it could be improved. Although the triples extracted automatically from texts do not provide a full picture of the CIDOC-CRM structure buried in the textual description, our results indicate that it provides a sound initial working basis for the mapping/translation process, saving time on what would otherwise have to be done by hand.
The paper focuses on the description of the process of “mining” lexical semantic information from published dictionaries of Brazilian Portuguese language. Specifically, it is described the manual approach of compiling and filtering hyperonymy/hyponymy and holonymy/meronym logic-conceptual relations of concrete nouns. These relations, filtering from the dictionaries, will be use to organize part of the nouns of Wordnet.Br lexical database.
This paper explores techniques to take advantage of the fundamental difference in structure between hidden Markov models (HMM) and hierarchical hidden Markov models (HHMM). The HHMM structure allows repeated parts of the model to be merged together. A merged model takes advantage of the recurring patterns within the hierarchy, and the clusters that exist in some sequences of observations, in order to increase the extraction accuracy. This paper also presents a new technique for reconstructing grammar rules automatically. This work builds on the idea of combining a phrase extraction method with HHMM to expose patterns within English text. The reconstruction is then used to simplify the complex structure of an HHMM The models discussed here are evaluated by applying them to natural language tasks based on CoNLL-2004 1 and a sub-corpus of the Lancaster Treebank 2.