Papers reviewed and determined not to be word norm studies. Use the flag icon to report errors or suggest re-inclusion.
18625 papers
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.
ABSTRACT. We investigated the efficacy of beam search parsing and deep parsing techniques in probabilistic HPSG parsing. We first tested the beam thresholding and iterative parsing. Next, we tested three techniques originally developed for deep parsing: quick check, large constituent inhibition, and hybrid parsing with a CFG chunk parser. The quick check, iterative parsing and hybrid parsing greatly contributed to total parsing performance. The accuracy and average parsing time for the Penn treebank were 87.2 % and 355 ms. Finally, we tested robustness and scalability of HPSG parsing on the MEDLINE corpus consisting of around 1.4 billion words. The entire corpus was parsed in 9 days with 340 CPUs. RÉSUMÉ. Nous avons étudié l’efficacité de l’analyse de beam search et des techniques de l’analyse profonde dans le probabiliste HPSG analyse. D’abord, nous avons examiné le beam thresholding et l’analyse itérative. Ensuite, nous avons examiné trois techniques développées originalement pour l’analyse profonde: quick check, large constituent inhibition, et l’analyse hybride avec la CFG chunk parser. Le quick check, l’analyse itérative et l’analyse hybride contribuaient considérablement à la performance de l’analyse totale. L’exactitude et le temps d’analyse moyen pour le Penn Treebank étaient 87.2 % et 355ms. Finalement, nous avons examiné la robustesse et la extensibilité de HPSG analyse sur le corpus de MEDLINE contenant presque 1.4 milliard de mots. Le corpus entier a été analysé en 9 jours avec 340 CPUs.
Unlexicalized probabilistic context-free parsing is a general and flexible approach that sometimes reaches competitive results in multilingual dependency parsing even if a minimum of language-specific information is supplied. Furthermore, integrating parser results (good at long dependencies) and tagger results (good at short range dependencies, and more easily adaptable to treebank peculiarities) gives competitive results in all languages.
Based on text chunking using HMM, transformation-based learning is made use of to improve the precision of chunk tags further. The training data and the test data are from Penn treebank 4.0, and 13 text chunks are used. Rules are learned automatically according to the rule templates. The precision is improved 4.48%. The detailed analysis that affects the text chunking is given. Different threshold, different scale of training data, different learning equation and different rule templates can affect the precision of the text chunking.
We report on a series of experiments with probabilistic context-free grammars predicting English and German syllable structure. The treebank-trained grammars are evaluated on a syllabification task. The grammar used by As she evaluates the grammar only for German, we reimplement the grammar and experiment with additional phonotactic features. Using bi-grams within the syllable, we can model the dependency from the previous consonant in the onset and coda. A 10fold cross validation procedure shows that syllabification can be improved by incorporating this type of phonotactic knowledge. Compared to the grammar of Mller (2002), syllable boundary accuracy increases from 95.8% to 97.2% for English, and from 95.9% to 97.2% for German. Moreover, our experiments with different syllable structures point out that there are dependencies between the onset on the nucleus for German but not for English. The analysis of one of our phonotactic grammars shows that interesting phonotactic constraints are learned. For instance, unvoiced consonants are the most likely first consonants and liquids and glides are preferred as second consonants in two-consonantal onsets.
TransBooster is a wrapper technology designed to improve the performance of wide-coverage machine translation \nsystems. Using linguistically motivated syntactic information, it automatically decomposes source language sentences into shorter and syntactically simpler chunks, and recomposes their translation to form target language sentences. This generally improves both the word order \nand lexical selection of the translation. To date, TransBooster has been successfully applied to rule-based MT, statistical MT, and multi-engine MT. This paper presents \nthe application of TransBooster to Example-Based Machine Translation. In an experiment conducted on test sets \nextracted from Europarl and the Penn II Treebank we show that our method can raise the BLEU score up to 3.8% relative \nto the EBMT baseline. We also conduct a manual evaluation, showing that TransBooster-enhanced EBMT produces \na better output in terms of fluency than the baseline EBMT in 55% of the cases and in terms of accuracy in 53% of the \ncases.
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.
The Jibiki platform ¡s an online generic environment for writing and querying all kinds of dictionaries: terminological glossaries, bilingual dictionaries, multilingual lexical databases, etc. It has been developed mainly by Mathieu Mangeot (Université de Savoie, France) and Gilles Sérasset (Université de Grenoble 1, France), thanks to research driven by the GETA team of the CLIPS laboratory in Grenoble, France. The platform allows one to lookup all the dictionaries available on the server and to display the results in the same window. The advanced query interface offers a combination of multiple search criteria. The writing of the entries is done directly online on the platform via a web browser. The writing interface is generated automatically from the description of the structure of the entries (an XML schema), thus allowing the edition of (almost) any type of dictionary entry.
After parsing an important task is to determine the semantic structure of sentences. In this paper, we attempt to automatically annotate the Penn Chinese Treebank with semantic dependency structure. Initially a small portion of the Penn Chinese Treebank was manually annotated with headword and dependency relations. Two supervised machine learning algorithms with varying features were then used to learn the relations. Finally, a set of rules were created based on features of Chinese to solve some problem patterns that were found in the Penn Chinese Treebank dealing with ambiguous structures. The experimental results show that the algorithms and proposed approach are effective for determining semantic dependency structure automatically.
Incorporating semantic features from the WordNet lexical database is among one of the many approaches that have been tried to improve the predictive performance of text classifica-tion models. The intuition behind this is that keywords in the training set alone may not be extensive enough to enable generation of a universal model for a category, but if we in-corporate the word relationships in WordNet, a more accu-rate model may be possible. Other researchers have previ-ously evaluated the effectiveness of incorporating WordNet synonyms, hypernyms, and hyponyms into text classification models. Generally, they have found that improvements in accuracy using features derived from these relationships are dependent upon the nature of the text corpora from which the document collections are extracted. In this paper, we not only reconsider the role of WordNet synonyms, hypernyms, and hyponyms in text classification models, we also consider the role of WordNet meronyms and holonyms. Incorporating these WordNet relationships into a Coordinate Matching clas-sifier, a Naive Bayes classifier, and a Support Vector Machine classifier, we evaluate our approach on six document collec-tions extracted from the Reuters-21578, USENET, and Digi-Trad text corpora. Experimental results show that none of the WordNet relationships were effective at increasing the accu-racy of the Naive Bayes classifier. Synonyms, hypernyms, and holonyms were effective at increasing the accuracy of the Coordinate Matching classifier, and hypernyms were effec-tive at increasing the accuracy of the SVM classifier.
Traditional Constraint Grammar (CG) is a methodological, rather than a descriptive paradigm, designed for robust parsing, not the implementation of a specific linguistic theory. Therefore, if used for treebank generation, it is not immediately clear, which linguistic formalism would be easiest to
In this paper, we extend an existing parser to produce richer output annotated with function labels. We obtain state-of-the-art results both in function labelling and in parsing, by automatically relabelling the Penn Treebank trees. In particular, we obtain the best published results on semantic function labels. This suggests that current statistical parsing methods are sufficiently general to produce accurate shallow semantic annotation.
We investigated the performance efficacy of beam search parsing and deep parsing techniques in probabilistic HPSG parsing using the Penn treebank. We first tested the beam thresholding and iterative parsing developed for PCFG parsing with an HPSG. Next, we tested three techniques originally developed for deep parsing: quick check, large constituent inhibition, and hybrid parsing with a CFG chunk parser. The contributions of the large constituent inhibition and global thresholding were not significant, while the quick check and chunk parser greatly contributed to total parsing performance. The precision, recall and average parsing time for the Penn treebank (Section 23) were 87.85%, 86.85%, and 360 ms, respectively.
Chunk parsing is conceptually appealing but its performance has not been satisfactory for practical use. In this paper we show that chunk parsing can perform significantly better than previously reported by using a simple sliding-window method and maximum entropy classifiers for phrase recognition in each level of chunking. Experimental results with the Penn Treebank corpus show that our chunk parser can give high-precision parsing outputs with very high speed (14 msec/sentence). We also present a parsing method for searching the best parse by considering the probabilities output by the maximum entropy classifiers, and show that the search method can further improve the parsing accuracy.
We introduce a method for transferring annotation from a syntactically annotated corpus in a source language to a target language. Our approach assumes only that an (unannotated) text corpus exists for the target language, and does not require that the parameters of the mapping between the two languages are known. We outline a general probabilistic approach based on Data Augmentation, discuss the algorithmic challenges, and present a novel algorithm for sampling from a posterior distribution over trees.
Great progress has been made in parsing the Wall Street Journal portion of the Penn Treebank. Now parsing languages other than English is an intensive research area. Head-driven model is one of the best English parsing models. It has been successfully applied to Czech but failed to outperform a base-line model in parsing German. This paper attempts to parse Chinese with head-driven model. Promising experimental results demonstrate that head-driven model works well for Chinese. We propose a hybrid parsing strategy, which combines head-driven model with a Chinese base phrases parsing model. The combined model not only improves the performance but also makes the parser space and time efficient. We evaluate our method in PARSEVAL measures, and the combined model performances are at 79.88% precision, 81.97% recall.
Recently, shallow parsing has been applied to various information processing systems, such as information retrieval, information extraction, question answering, and automatic document summarization. A shallow parser is suitable for online applications, because it is much more efficient and less demanding than a full parser. In this research, we formulate shallow parsing as a sequential tagging problem and use a supervised machine learning technique, Maximum Entropy (ME), to build a Chinese shallow parser. The major features of the ME-based shallow parser are POSs and the context words in a sentence. We adopt the shallow parsing results of Sinica Treebank as our standard, and select 30,000 and 10,000 sentences from Sinica Treebank as the training set and test set respectively. We then test the robustness of the shallow parser with noisy data. The experiment results show that the proposed shallow parser is quite robust for sentences with unknown proper nouns. 1.
Contemporary neuropsychological studies have stressed the widely distributed and multicomponential nature of human affective processes. Here, we examined facial electromyographic (EMG) (zygomaticus and corrugator muscle activity), autonomic (skin conductance and heart rate) and subjective measures of affective valence and arousal in patient TG, a 30 year-old man with left anterior mediotemporal and left orbitofrontal lesions resulting from a traumatic brain injury. Both TG and a normal control group were exposed to hedonically valenced visual and olfactory stimuli. In contrast with control subjects, facial EMG and electrodermal activity in TG did not differentiate among pleasant, unpleasant and neutral pictures. In addition, the controls reacted spontaneously with larger corrugator EMG activity and higher skin conductance to unpleasant odors. By contrast, the subjective feeling states (pleasure and arousal ratings) remained preserved in TG. The covariation between facial and self-report measures of negative valence was also a function of the nature of the olfactory task in the patient only. Taken together, the data suggest a functional dissociation between brain substrates supporting generation of emotion and those supporting representation of emotion.
All studies that examined the co-occurrence of pleasure and displeasure revealed at least some reports of mixed feelings (i.e., reports of concurrent pleasure and displeasure). Some researchers attribute these reports to measurement error, whereas others regard them as valid. This article examined response latencies of affect ratings to test the validity of reported mixed feelings. First, I demonstrate that respondents need more time to indicate the presence of an affect than the absence of an affect on unipolar response formats. Then, I demonstrate that mixed feelings are reported even when response latencies show a unipolar pattern for displeasure ratings, while the results for pleasure ratings were more mixed, but additional evidence suggests that they also are valid reflections of pleasure. In addition, I demonstrate that pleasure and displeasure ratings are independent of item-order and item-spacing. These results provide further support for the validity of reported mixed feelings and two-dimensional representations of pleasure and displeasure.
In data-driven approaches to natural language processing, a common problem is the lack of data for many languages. Within the project Stochastic Dependency Grammars for Natural Language Parsing at Växjö University, we (Joakim Nivre, Johan Hall and Jens Nilsson) are developing a deterministic data-driven dependency parser, which is language independent. In this project we intend to enlarge the data resources for our parser. For the moment, we have only tested our parser on small Swedish treebank converted to dependency structure, and on English using Penn Treebank converted to dependency trees. Since we do not have more Swedish dependency treebanks at hand, we want to broaden our view towards treebanks for other languages, especially the bigger ones, to investigate the influence of data size. Primarily, we are focusing on the Danish Dependency Treebank (DDT) and the Prague Dependency Treebank (PDT). These treebanks are not in a format that we can use for our parser and therefore we have to convert them to MALT-XML, a format which our parser can handle.
INTRODUCTION: In this work, we introduce the concept of semantic role labeling to the medical domain. We report first results of porting and adapting an existing resource, Propbank, to the medical field. Propbank is an adjunct to Penn Treebank that provides semantic annotation of predicates and the roles played by their arguments. The main aim of this work is the applicability of the Propbank frame files to predicates typically encountered in the medical literature. METHODS: We analyzed a target corpus of 610,100 abstracts, which was selected by searching for publication type "case reports". From this target corpus, we randomly selected 10,000 sample abstracts to estimate the predicate distribution, and matched the predicates from this sample to the predicates in Propbank. RESULTS: Of the 1998 unique verbs in our sample, 76% were represented in Propbank. This included the 40 most frequent verbs, which represented 49% of all predicate instances in our sample and which matched the Propbank usage in a study of representative sentences. We propose extensions to Propbank that handle medical predicates, which are not adequately covered by Propbank. CONCLUSION: We believe that semantic role labeling using Propbank is a valid approach to capture predicate relations in the medical literature.
Annotated linguistic databases are widely used in linguistic research and in language technology development. These annotations are typically hierarchical, and represent the nested structure of syntactic and prosodic constituents. Recently, the LPath language has been proposed as a convenient path-based language for querying linguistic trees. We establish the formal expressiveness of LPath relative to the XPath family of languages. We also extend LPath to permit simple closures, resulting in a first-order complete language which we believe is sufficiently expressive for the majority of linguistic tree query needs. 1.
One of the main motivations for building treebanks is that they facilitate the development of syntactic parsers, by providing realistic data for evaluation as well as inductive learning. In this paper we present what we believe to be the first robust data-driven parser for Bulgarian, trained and evaluated on data from BulTreeBank (Simov et al., 2002). The parser uses dependency-based representations and employs a deterministic algorithm to construct dependency structures in a single pass over the input string, guided by a memory-based classifier at each nondeterministic choice point, as described in Nivre et al. (2004). Since the original BulTreeBank annotation is based on HPSG, it has been necessary to extract dependency structures from the original annotation both for training and evaluating the parser. The paper is structured as follows. Section 2 introduces the MaltParser system used in the experiments to induce parsers from treebank data. Section 3 presents BulTreeBank, and section 4 shows how the BulTreeBank annotation can be converted into a dependency structure annotation of the kind required by MaltParser. Section 5 describes the experimental conditions, and section 6 discusses the results of the experiments. Section 7 contains our conclusions.
MONA is an automata toolkit providing a compiler for compiling formulae of monadic second order logic on strings or trees into string automata or tree automata. In this paper, we evaluate the option of using MONA as a treebank query tool. Unfortunately, we find that MONA is not an option. There are several reasons why the main being unsustainable query answer times. If the treebank contains larger trees with more than 100 nodes, then even the processing of simple queries may take hours.
We propose a framework that enables the acquisition of annotation-heavy resources such as syntacfic dependency tree corpora for low-resource languages by importing linguistic annotations from high-quality English resources. We present a large-scale experiment showing that Chinese dependency trees can be induced by using an English parser, a word alignment package, and a large corpus of sentence-aligned bilingual text. As a part of the experiment, we evaluate the quality of a Chinese parser trained on the induced dependency treebank. We find that a parser trained in this manner out-performs some simple baselines inspite of the noise in the induced treebank. The results suggest that projecting syntactic structures from English is a viable option for acquiring annotated syntactic structures quickly and cheaply. We expect the quality of the induced treebank to improve when more sophisticated filtering and error-correction techniques are applied.
We explore the application of memorybased learning to morphological analysis and part-of-speech tagging of written Arabic, based on data from the Arabic Treebank. Morphological analysis -- the construction of all possible analyses of isolated unvoweled wordforms -- is performed as a letter-by-letter operation prediction task, where the operation encodes segmentation, part-of-speech, character changes, and vocalization. Part-of-speech tagging is carried out by a bi-modular tagger that has a subtagger for known words and one for unknown words. We report on the performance of the morphological analyzer and part-of-speech tagger. We observe that the tagger, which has an accuracy of 91.9% on new data, can be used to select the appropriate morphological analysis of words in context at a precision of 64.0 and a recall of 89.7.
Development and testing of a large range of NLP applications presuppose corpora annotated at levels more advanced than those of part–of–speech and shallow syntax. Therefore, multi–layered annotation schemes have been designed in order to provide deeper representations of intra – and inter–sentential structure and meaning.
We describe a parallel annotation approach for PubMed abstracts. It includes both entity/relation annotation and a treebank containing syntactic structure, with a goal of mapping entities to constituents in the treebank. Crucial to this approach is a modification of the Penn Treebank guidelines and the characterization of entities as relation components, which allows the integration of the entity annotation with the syntactic structure while retaining the capacity to annotate and extract more complex events.
We present a heuristic technique for converting a constituency treebank into a dependency treebank. In particular, we comment on our experience in converting the Spanish treebank Cast3LB. We extract a context-free grammar from the treebank, automatically identify the head in each rule, and use this information for constructing the dependency tree. Our heuristics have 99 % precision and 80 % recall in identifying the head in the rules, which gives 92% accuracy in identifying dependencies between words.
In this paper, we describe the automatic annotation of the Cast3LB Treebank with LFG f-structures for the subsequent extraction of Spanish probabilistic grammar and lexical resources. We adapt the approach and methodology of Cahill et al. (2004), O’Donovan et al. (2004) and elsewhere for English to Spanish and the Cast3LB treebank encoding. We report on the quality and coverage of the automatic f-structure annotation. Following the pipeline and integrated models of Cahill et al. (2004), we extract wide-coverage \nprobabilistic LFG approximations and parse unseen Spanish text into f-structures. We also extend Bikel’s (2002) Multilingual Parse Engine to include a Spanish language module. Using the retrained Bikel parser in the pipeline model gives the best results against a manually constructed gold standard (73.20% predsonly f-score). We also extract Spanish lexical resources: 4090 semantic form types with 98 frame types. Subcategorised prepositions and particles are included in the frames.
Discourse connectives can be analyzed as discourse level predicates which projectpredicate-argument structure on a par with verbs at the sentence level. The PennDiscourse Treebank (PDTB) reflects this view in its design providing annotation ofthe discourse connectives and their arguments. Like verbs, discourse connectiveshave multiple senses. We present a set of manual sense annotation studies for threeconnectives whose arguments have been annotated in the PDTB. Using syntacticfeatures computed from the Penn Treebank and a simple MaxEnt model, we haveachieved some success in automatically disambiguating among their senses.
Recent years have seen a revived interst in semantic parsing by applying statistical and machinelearning methods to semantically annotated corpora such as the FrameNet and the Proposition Bank. So far much of the research has been focused on English due to the lack of semantically annotated resources in other languages. In this paper, we report first results on semantic role labeling using a pre-release version of the Chinese Proposition Bank. Since the Chinese Proposition Bank is superimposed on top of the Chinese Treebank, i.e., the semantic role labels are assigned to constituents in a treebank parse tree, we start by reporting results on experiments using the handcrafted parses in the treebank. This will give us a measure of the extent to which the semantic role labels can be bootstrapped from the syntactic annotation in the treebank. We will then report experiments using a fully automatic Chinese parser that integrates word segmentation, POS-tagging and parsing. This will gauge how successful semantic role labeling can be done for Chinese in realistic situations. We show that our results using hand-crafted parses are slightly higher than the results reported for the state-of-the-art semantic role labeling systems for English using the Penn English Proposition Bank data, even though the Chinese Proposition Bank is smaller in size. When
The annotations of the Penn Discourse Treebank (PDTB) include (1) discourse connectives and their arguments, and (2) attribution of each argument of each connective and of the relation it denotes. Because the PDTB covers the same text as the Penn TreeBank WSJ corpus, syntactic and discourse annotation can be compared. This has revealed significant differences between syntactic structure and discourse structure, in terms of the arguments of connectives, due in large part to attribution. We describe these differences, an algorithm for detecting them, and finally some experimental results. These results have implications for automating discourse annotation based on syntactic annotation.
The paper investigates the use of richer syntactic dependencies in the structured language model (SLM). We present two simple methods of enriching the dependencies in the syntactic parse trees used for initializing the SLM. We evaluate the impact of both methods on the perplexity (PPL) and word-error-rate (WER, N-best rescoring) performance of the SLM. We show that the new model achieves an improvement in PPL and WER over the baseline results reported using the SLM on the UPenn Treebank and Wall Street Journal (WSJ) corpora, respectively.
Computer-based studies usually produce log files as raw data. These data cannot be analyzed adequately with conventional statistical software. The Chemnitz LogAnalyzer provides tools for quick and comfortable visualization and analyses of hypertext navigation behavior by individual users and for aggregated data. In addition, it supports analogous analyses of questionnaire data and reanalysis with respect to several predefined orders of nodes of the same hypertext. As an illustration of how to use the Chemnitz LogAnalyzer, we give an account of one study on learning with hypertext. Participants either searched for specific details or read a hypertext document to familiarize themselves with its content. The tool helped identify navigation strategies affected by these two processing goals and provided comparisons, for example, of processing times and visited sites. Altogether, the Chemnitz LogAnalyzer fills the gap between log files as raw data of Web-based studies and conventional statistical software.
We present a method for improving dependency structure analysis of Chinese. Our bottom-up deterministic analyzer adopt Nivre’s algorithm (Nivre and Scholz, 2004). Support Vector Machines (SVMs) are utilized to determine the word dependency relations. We find that there are two problems in our analyzer and propose two methods to solve them. One problem is that some operations cannot be solved only using local feature. We utilize the global features to solve this. The other problem is that this bottom-up analyzer doesn’t use top-down information. We supply the top-down information by constructing SVMs based root node finder to solve this problem. Experimental evaluation on the Penn Chinese Treebank Corpus shows that the proposed extensions improve the parsing accuracy significantly. 1