Papers reviewed and determined not to be word norm studies. Use the flag icon to report errors or suggest re-inclusion.
16504 papers
International audience
We present a study on two key characteristics of human syntactic annotations: anchoring and agreement. Anchoring is a well known cognitive bias in human decision making, where judgments are drawn towards pre-existing values. We study the influence of anchoring on a standard approach to creation of syntactic resources where syntactic annotations are obtained via human editing of tagger and parser output. Our experiments demonstrate a clear anchoring effect and reveal unwanted consequences, including overestimation of parsing performance and lower quality of annotations in comparison with human-based annotations. Using sentences from the Penn Treebank WSJ, we also report systematically obtained inter-annotator agreement estimates for English dependency parsing. Our agreement results control for parser bias, and are consequential in that they are on par with state of the art parsing performance for English newswire. We discuss the impact of our findings on strategies for future annotation efforts and parser evaluations.
We train one multilingual model for dependency parsing and use it to parse sentences in several languages. The parsing model uses (i) multilingual word clusters and embeddings; (ii) token-level language information; and (iii) language-specific features (fine-grained POS tags). This input representation enables the parser not only to parse effectively in multiple languages, but also to generalize across languages based on linguistic universals and typological similarities, making it more effective to learn from limited annotations. Our parser’s performance compares favorably to strong baselines in a range of data scenarios, including when the target language has a large treebank, a small treebank, or no treebank for training.
The PDTB Annotator is a tool for annotating and adjudicating discourse relations based on the annotation framework of the Penn Discourse TreeBank (PDTB). This demo describes the benefits of using the PDTB Annotator, gives an overview of the PDTB Framework and discusses the tool’s features, setup requirements and how it can also be used for adjudication.
In this paper, a German verb resource for verb-centered sentiment inference is introduced and evaluated. Our model specifies verb polarity frames that capture the polarity effects on the fillers of the verb's arguments given a sentence with that verb frame. Verb signatures and selectional restrictions are also part of the model. An algorithm to apply the verb resource to treebank sentences and the results of our first evaluation are discussed.
We investigated whether lines and shapes that present face-like features would be associated with emotions. In Experiment 1, participants associated concave, convex, or straight lines with the words happy or sad. Participants found it easiest to associate the concave line with happy and the convex line with sad. In Experiment 2, participants rated (valence, pleasantness, liking, and tension) and categorised (valence and emotion words) two convex and concave lines that were paired with six distinct pairs of eyes. The presence of eyes affected participants' valence ratings and response latencies; more congruent eye-mouth matches produced more consistent ratings and faster reaction times. In Experiment 3, we examined whether dots that resembled eyes would be associated with emotional words. Participants found it easier to match certain sets of dots with specific emotions. These results suggest that facial gestures that are associated with specific emotions can be captured using relatively simple shapes and lines.
This article attempts to place dependency annotation options on a solid theoretical and applied footing. By verifying the validity of some basic choices of the current dependency reference framework, Universal Dependencies (UD), in a perspective of general annotation principles, we show how some choices can lead to inconsistencies and discontinuities, partly due to UD's alternation between syntax and semantics. For some constructions, we propose better suited alternative structures with a clear-cut distinction of syntax and semantics. We propose a classification of conception-oriented, annotatororiented, and finally, treebank end-useroriented considerations to be used in the creation of new annotation schemes.
Abstract The automatic interpretation of 3D point clouds for building reconstruction is a challenging task. The interpretation process requires highly structured models representing semantics. Formal grammars can describe structures as well as the parameters of buildings and their parts. We propose a novel approach for the automatic learning of weighted attributed context‐free grammar rules for 3D building reconstruction, supporting the laborious manual design of rules. We separate structure from parameter learning. Specific Support Vector Machines (SVMs) are used to generate a weighted context‐free grammar and predict structured outputs such as parse trees. The grammar is extended by parameters and constraints, which are learned based on a statistical relational learning method using Markov Logic Networks (MLNs). MLNs enforce the topological and geometric constraints. MLNs address uncertainty explicitly and provide probabilistic inference. They are able to deal with partial observations caused by occlusions. Uncertain projective geometry is used to deal with the uncertainty of the observations. Learning is based on a large building database covering different building styles and façade structures. In particular, a treebank that has been derived from the database is employed for structure learning.
Due to the constant increasing of electronic textual information, modern society needs for the automatic processing of natural language (NL). The main purpose of NL automatic text processing systems is to analyze and create texts and represent their content. The purpose of the paper is the development of linguistic and software bases of an automatic system for processing English publicistic texts. This article discusses the examples of different approaches to the creation of linguistic databases for processing systems. The author gives a detailed description of basic building blocks for a new linguistic processor: lexical-semantic, syntactical and semantic-syntactical. The main advantage of the processor is using special semantic codes in the alphabetical dictionary. The semantic codes have been developed in accordance with a lexical-semantic classification. It helps to precisely define semantic functions of the keywords that are situated in parsing groups and allows the automatic system to avoid typical mistakes. The author also represents the realization of a developed linguistic database in the form of a training computer program.
OBJECTIVE: To investigate the effect of transcranial direct current stimulation (tDCS) on food craving, intake, binge eating desire, and binge eating frequency in individuals with binge eating disorder (BED). METHOD: N = 30 adults with BED or subthreshold BED received a 20-min 2 milliampere (mA) session of tDCS targeting the dorsolateral prefrontal cortex (DLPFC; anode right/cathode left) and a sham session. Food image ratings assessed food craving, a laboratory eating test assessed food intake, and an electronic diary recorded binge variables. RESULTS: tDCS versus sham decreased craving for sweets, savory proteins, and an all-foods category, with strongest reductions in men (p < 0.05). tDCS also decreased total and preferred food intake by 11 and 17.5%, regardless of sex (p < 0.05), and reduced desire to binge eat in men on the day of real tDCS administration (p < 0.05). The reductions in craving and food intake were predicted by eating less frequently for reward motives, and greater intent to restrict calories, respectively. DISCUSSION: This proof of concept study is the first to find ameliorating effects of tDCS in BED. Stimulation of the right DLPFC suggests that enhanced cognitive control and/or decreased need for reward may be possible functional mechanisms. The results support investigation of repeated tDCS as a safe and noninvasive treatment adjunct for BED. © 2016 Wiley Periodicals, Inc.(Int J Eat Disord 2016; 49:930-936).
In this paper, we study novel neural network structures to better model long term dependency in sequential data. We propose to use more memory units to keep track of more preceding states in recurrent neural networks (RNNs), which are all recurrently fed to the hidden layers as feedback through different weighted paths. By extending the popular recurrent structure in RNNs, we provide the models with better short-term memory mechanism to learn long term dependency in sequences. Analogous to digital filters in signal processing, we call these structures as higher order RNNs (HORNNs). Similar to RNNs, HORNNs can also be learned using the back-propagation through time method. HORNNs are generally applicable to a variety of sequence modelling tasks. In this work, we have examined HORNNs for the language modeling task using two popular data sets, namely the Penn Treebank (PTB) and English text8 data sets. Experimental results have shown that the proposed HORNNs yield the state-of-the-art performance on both data sets, significantly outperforming the regular RNNs as well as the popular LSTMs.
Content-basis image retrieval is important option to prevail within the difficulties of previous works and contains attracted an excellent concentration in past decades. The models according to graph-based ranking were mostly analysed and extensively functional in file recovery area. Within our work we concentrate on the novel in addition to efficient graph-based model for content based image retrieval, designed for out-of-sample recovery on extensive databases. We advise a scalable graph-based ranking representation referred to as effective Manifold Ranking, which address weak points of Manifold Ranking from two most significant viewpoints for example scalable graph construction in addition to effective ranking computation. We concentrate on a famous graph-based model known Manifold Ranking that is a well-known graph-based ranking representation that ranks data samples relevant to intrinsic geometrical structure uncovered with a huge data. The suggested model includes two separate stages just like an offline stage for structuring of ranking model plus an online stage for controlling of recent query. Using the suggested system, we are able to handle database by a million images and perform online retrieval inside a short instance.
<span>This work reports on ongoing research aimed at modeling a metonymic relationship in the FrameNet <span>Brasil database. This paper is based on a case study with the <span>Teams <span>frame. Both the frame and the <span>corpus consulted are part of a frame-based trilingual (Portuguese – Spanish – English) electronic <span>dictionary covering the soccer, tourism and World Cup domains developed by FrameNet Brasil. The<br /><span>basic infrastructure, analytical categories and methodology used were those developed for FrameNet <span>(Fillmore et al. 2003, Baker et al. 2003, Ruppenhofer et al. 2010), which can be defined as an<br /><span>application of Frame Semantics to practical lexicography.</span></span></span></span></span></span><br /></span></span></span>
This paper provides a binary, token-based classification of German particle verbs (PVs) into literal vs. non-literal usage. A random forest improving standard features (e.g., bagof-words; affective ratings) with PV-specific information and abstraction over common nouns significantly outperforms the majority baseline. In addition, PV-specific classification experiments demonstrate the role of shared particle semantics and semantically related base verbs in PV meaning shifts.
We present Poly-GrETEL, an online tool which enables syntactic querying in parallel treebanks and which is based on the monolingual GrETEL environment. We provide online access to the Europarl parallel treebank for Dutch and English, allowing users to query the treebank using either an XPath expression or an example sentence in order to look for similar constructions. We provide automatic alignments between the nodes. By combining example-based query functionality with node alignments, we limit the need for users to be familiar with the query language and the structure of the trees in the source and target language, thus facilitating the use of parallel corpora for comparative linguistics and translation studies.
The question of the type of text used as primary data in treebanks is of certain importance. First, it has an influence at the discourse level: an article is not organized in the same way as a novel or a technical document. Moreover, it also has consequences in terms of semantic interpretation: some types of texts can be easier to interpret than others. We present in this paper a new type of treebank which presents the particularity to answer to specific needs of experimental linguistic. It is made of short texts (book backcovers) that presents a strong coherence in their organization and can be rapidly interpreted. This type of text is adapted to short reading sessions, making it easy to acquire physiological data (e.g. eye movement, electroencepholagraphy). Such a resource offers reliable data when looking for correlations between computational models and human language processing.
The article presents methodological analyses of topical ideas of the famous modern linguist – E.Cosseriu. The authors argue that incorporation of theoretical ideas of E.Cosseriu could substentially extand the euristic potential of the conept of norm in the sphere of linguistics. The key to solvation of the problem lies in the necessity of changing of modern theoretical context of the question. The authors of the article consider the history of operationalization of the concept of norm in linguistics from the point of view of a specific hermeneutic approach in relation to other linguistic techniques. In the course of study of the role of linguistic norms in interaction of content and expression the article presents examples of extrapolation of the concept of norm from one discipline to another. In case of extrapolation of the concept of norm from the other disciplines into linguistic investigations the structural and functional dichotomy of language turns out that it is impossible to avoid considering spiritual as the world of objects, so that speech and language are considered as two different things. The rapid development of information technologies made possible to calculate many of the aspects of Humboldtian ideas. Computer statistics created conditions where the idea of "language picture of the world" and of the "inner form of the language" is gradually losing its original romantic charge and turns in a very trivial thing. Yet despite the fact that global standardization significantly enhances the processing and automatic addition of translation, which once gave beginning to hermeneutics the hermeneutic potential of linguistic norm still preseves many promissing prospects from the epistemological point of view.
Les ressources linguistiques permettant aux études cross-langues de se développer sont très importantes pour les langues minoritaires telles que l’irlandais, car elles favorisent le partage des ressources\npour palier au problème du manque de données. Le projet «Universal Dependencies » (UD) a pour\nbut de faciliter les études cross-langues des arbres syntaxiques, des structures linguistiques et de\nl’analyse syntaxique. L’objectif principal de ce projet est de former un ensemble harmonieux d’arbres\nsyntaxiques en utilisant un schéma d’annotations universelles. Dans cet article, nous présentons\nla transformation de l’arbre de dépendance syntaxique irlandais (IDT) (Lynn, 2016) au schéma\nd’annotations universelles du projet UD, suivie d’une description claire des changements structurels\nnécessaires à cette conversion. Le nouvel arbre est ainsi appelé « Irish Universal Dependency\nTreebank » ( IUDT ).\n\nLanguage resources that enable cross-lingual studies have become increasingly valuable for lesserresourced languages such as Irish, as they allow for easier sharing of resources, thus overcoming\nthe problem of data scarcity. The Universal Dependencies (UD) Project1\nis an initiative aimed at\ncross-lingual studies of treebanks, linguistic structures and parsing. Its goal is to create a set of\nmultilingual harmonised treebanks that are designed according to a universal annotation scheme. In\nthis paper, we report on the conversion of the Irish Dependency Treebank (IDT) (Lynn, 2016) to a\nUD version of the treebank which we term the Irish Universal Dependency Treebank (IUDT). We\nreport on the mapping of the IDT labelling scheme to the UD scheme, along with a clear description\nof the structural changes required in this conversion.
Mood affects both memory accuracy and memory distortions. However, some aspects of this relation are still poorly understood: (1) whether valence and arousal equally affect false memory production, and (2) whether retrieval-related processes matter; the extant literature typically shows that mood influences memory performance when it is induced before encoding, leaving unsolved whether mood induced before retrieval also impacts memory. We examined how negative, positive, and neutral mood induced before retrieval affected inferential false memories and related subjective memory experiences. A recognition-memory paradigm for photographs depicting script-like events was employed. Results showed that individuals in both negative and positive moods-similar in arousal levels-correctly recognized more target events and endorsed fewer false memories (and these errors were linked to remember responses less frequently), compared to individuals in neutral mood. This suggests that arousal (but not valence) predicted memory performance; furthermore, we found that arousal ratings provided by participants were more adequate predictors of memory performance than their actual belonging to either positive, negative or neutral mood groups. These findings suggest that arousal has a primary role in affecting memory, and that mood exerts its power on true and false memory even when induced at retrieval.
Study 2 Image Rating
The first-line psychological treatment for anxiety disorders is exposure therapy, which can be modeled in the laboratory using fear extinction. In healthy women, estradiol levels predict return of fear following extinction, whereas low levels are associated with greater return of fear. Investigating whether estradiol is similarly associated with extinction in clinically anxious women may provide insight to mechanisms underlying symptom relapse following exposure therapy. In the present study, women with spider phobia and healthy women participated in a 2-day fear conditioning and extinction procedure during a period of high or low estradiol levels. Skin conductance responses, shock expectancy, and valence ratings were measured throughout. Women exhibited comparable decreases in physiological arousal from conditioning to the end of extinction training on Day 1. However, compared to women with high estradiol, and irrespective of clinical status, women with low estradiol exhibited significant return of physiological arousal at extinction recall on Day 2, despite accurate ratings regarding the likelihood of shock. Low estradiol women also reported heightened threat expectancy and physiological responding during presentation of safety cues. These results may point to novel means of enhancing exposure therapy in women by timing treatment delivery during periods of higher estradiol levels. (PsycINFO Database Record
In this work, we focuses on the assisted exploitation of lexical databases designed according to the LMF standard (Lexical Markup Framework) ISO-24613. The proposed system is a service-oriented solution which relies on a requirement-based lexical web service generation approach that expedites the task of engineers when developing NLP (Natural Language Processing) systems. Using this approach, the developer will neither deal with the database content or its structure nor use any language query. Furthermore, this approach will promote a largescale reuse of LMF lexical databases by generating lexical web services for all languages. For evaluating this approach we have tested it on the Arabic language.
In this work, we propose a novel method to incorporate corpus-level discourse information into language modelling. We call this larger-context language model. We introduce a late fusion approach to a recurrent language model based on long shortterm memory units (LSTM), which helps the LSTM unit keep intra-sentence dependencies and inter-sentence dependencies separate from each other. Through the evaluation on four corpora (IMDB, BBC, Penn TreeBank, and Fil9), we demonstrate that the proposed model improves perplexity significantly. In the experiments, we evaluate the proposed approach while varying the number of context sentences and observe that the proposed late fusion is superior to the usual way of incorporating additional inputs to the LSTM. By analyzing the trained larger-context language model, we discover that content words, including nouns, adjectives and verbs, benefit most from an increasing number of context sentences. This analysis suggests that larger-context language model improves the unconditional language model by capturing the theme of a document better and more easily. * Recently,
Notre recherche se situe en lexicographie computationnelle, et concerne non seulement le support informatique aux ressources lexicales utiles pour la TA (traduction automatique) et la THAM (traduction humaine aidée par la machine), mais aussi l'architecture linguistique des bases lexicales supportant ces ressources, dans un contexte opérationnel (thèse CIFRE avec L&M).Nous commençons par une étude de l'évolution des idées, depuis l'informatisation des dictionnaires classiques jusqu'aux plates-formes de construction de vraies "bases lexicales" comme JIBIKI-1 [Mangeot, M. et al., 2003; Sérasset, G., 2004] et JIBIKI-2 [Zhang, Y. et al., 2014]. Le point de départ a été le système PIVAX-1 [Nguyen, H.-T. et al., 2007; Nguyen, H. T. & Boitet, C., 2009] de bases lexicales pour systèmes de TA hétérogènes à pivot lexical supportant plusieurs volumes par "espace lexical" naturel ou artificiel (UNL). En prenant en compte le contexte industriel, nous avons centré notre recherche sur certains problèmes, informatiques et lexicographiques.Pour passer à l'échelle, et pour profiter des nouvelles fonctionnalités permises par JIBIKI-2, dont les "liens riches", nous avons transformé PIVAX-1 en PIVAX-2, et réactivé le projet GBDLEX-UW++ commencé lors du projet ANR TRAOUIERO, en réimportant toutes les données (multilingues) supportées par PIVAX-1, et en les rendant disponibles sur un serveur ouvert.Partant d'un besoin de L&M concernant les acronymes, nous avons étendu la "macrostructure" de PIVAX en y intégrant des volumes de "prolexèmes", comme dans PROLEXBASE [Tran, M. & Maurel, D., 2006]. Nous montrons aussi comment l'étendre pour répondre à de nouveaux besoins, comme ceux du projet INNOVALANGUES. Enfin, nous avons créé un "intergiciel de lemmatisation", LEXTOH, qui permet d'appeler plusieurs analyseurs morphologiques ou lemmatiseurs, puis de fusionner et filtrer leurs résultats. Combiné à un nouvel outil de création de dictionnaires, CREATDICO, LEXTOH permet de construire à la volée un "mini-dictionnaire" correspondant à une phrase ou à un paragraphe d'un texte en cours de "post-édition" en ligne sous IMAG/SECTRA, ce qui réalise la fonctionnalité d'aide lexicale proactive prévue dans [Huynh, C.-P., 2010]. On pourra aussi l'utiliser pour créer des corpus parallèles "factorisés" pour construire des systèmes de TA en MOSES.
Purpose: The purpose of this study was to evaluate the effects of hearing loss and age on subjective ratings of emotional valence and arousal in response to nonspeech sounds. Method: Three groups of adults participated: 20 younger listeners with normal hearing (M = 24.8 years), 20 older listeners with normal hearing (M = 55.8 years), and 20 older listeners with mild-to-severe acquired hearing loss (M = 65.6 years). Stimuli were presented via headphones at either 35 and 65 dB SPL or 50 and 80 dB SPL on the basis of random assignment within each group. Participants rated the emotional valence and arousal for previously normed nonspeech auditory stimuli. Results: Linear mixed model analyses were conducted separately for ratings of valence and arousal. Results revealed that listeners with hearing loss exhibited a reduced range of emotional ratings. Furthermore, for stimuli presented at 80 dB SPL, valence ratings from listeners with hearing loss were significantly lower than ratings from listeners with normal hearing. Conclusions: Acquired hearing loss, not increased age, affected emotional responses by reducing the range of subjective ratings and by reducing the reported valence of the highest intensity stimuli. These results have potentially important clinical implications for aural rehabilitation.
This paper introduces the implementation and integration of a sentiment analysis pipeline into the ongoing open source cross-media analysis framework. The pipeline includes the following components; chat room cleaner, NLP and sentiment analyzer. Before the integration, we also compare two broad categories of sentiment analysis methods, namely lexicon-based and machine learning approaches. We mainly focus on finding out which method is appropriate to detect sentiments from forum discussion posts. In order to conduct our experiments, we use the apache-hadoop framework with its lexicon-based sentiment prediction algorithm and Stanford coreNLP library with the Recursive Neural Tensor Network (RNTN) model. The lexicon-based uses sentiment dictionary containing words annotated with sentiment labels and other basic lexical features, and the later one is trained on Sentiment Treebank with 215,154 phrases, labeled using Amazon Turk. Our overall performance evaluation shows that RNTN outperforms the lexicon-based by 9.88% accuracy on variable length positive, negative, and neutral comments. However, the lexicon-based shows better performance on classifying positive comments. We also found out that the F1-score values of the Lexicon-based is greater by 0.16 from the RNTN.
This paper introduces Graph Convolutional Recurrent Network (GCRN), a deep learning model able to predict structured sequences of data. Precisely, GCRN is a generalization of classical recurrent neural networks (RNN) to data structured by an arbitrary graph. Such structured sequences can represent series of frames in videos, spatio-temporal measurements on a network of sensors, or random walks on a vocabulary graph for natural language modeling. The proposed model combines convolutional neural networks (CNN) on graphs to identify spatial structures and RNN to find dynamic patterns. We study two possible architectures of GCRN, and apply the models to two practical problems: predicting moving MNIST data, and modeling natural language with the Penn Treebank dataset. Experiments show that exploiting simultaneously graph spatial and dynamic information about data can improve both precision and learning speed.
We introduce a simple semi-supervised approach to improve implicit discourse relation identification. This approach harnesses large amounts of automatically extracted discourse connectives along with their arguments to construct new distributional word representations. Specifically, we represent words in the space of discourse connectives as a way to directly encode their rhetorical function. Experiments on the Penn Discourse Treebank demonstrate the effectiveness of these task-tailored representations in predicting implicit discourse relations. Our results indeed show that, despite their simplicity, these connective-based representations outperform various off-the-shelf word embeddings, and achieve state-of-the-art performance on this problem.
In this work we describe the system built for the three English subtasks of the Se-mEval 2016 Task 3 by the Department of Computer Science of the University of Houston (UH) and the Pattern Recognition and Human Language Technology (PRHLT) research center -Universitat Politcnica de Valncia: UH-PRHLT. Our system represents instances by using both lexical and semantic-based similarity measures between text pairs. Our semantic features include the use of distributed representations of words, knowledge graphs generated with the BabelNet multilingual semantic network, and the FrameNet lexical database. Experimental results outperform the random and Google search engine baselines in the three English subtasks. Our approach obtained the highest results of subtask B compared to the other task participants.
Generalising what is learned about one stimulus to other but perceptually related stimuli is a basic behavioural phenomenon. We evaluated whether a rule learning mechanism may serve to explain such generalisation. To this end, we assessed whether inference rules communicated through verbal instructions affect generalisation. Expectancy ratings, but not valence ratings, proved sensitive to this manipulation. In addition to revealing a role for inference rules in generalisation, our study has clinical implications as well. More specifically, we argue that targeting inference rules might prove to be an effective strategy to affect the excessive generalisation that is often observed in psychopathology.
Within the field of functional magnetic resonance imaging (fMRI) neurofeedback, most studies provide subjects with instructions or suggest strategies to regulate a particular brain area, while other neuro-/biofeedback approaches often do not. This study is the first to investigate the hypothesis that subjects are able to utilize fMRI neurofeedback to learn to differentially modulate the fMRI signal from the bilateral amygdala congruent with the prescribed regulation direction without an instructed or suggested strategy and apply what they learned even when feedback is no longer available. Thirty-two subjects were included in the analysis. Data were collected at 3 Tesla using blood oxygenation level dependent (BOLD)-sensitivity optimized multi-echo EPI. Based on the mean contrast between up- and down-regulation in the amygdala in a post-training scan without feedback following three neurofeedback sessions, subjects were able to regulate their amygdala congruent with the prescribed directions with a moderate effect size of Cohen's d = 0.43 (95% conf. int. 0.23-0.64). This effect size would be reduced, however, through stricter exclusion criteria for subjects that show alterations in respiration. Regulation capacity was positively correlated with subjective arousal ratings and negatively correlated with agreeableness and susceptibility to anger. A learning effect over the training sessions was only observed with end-of-block feedback (EoBF) but not with continuous feedback (trend). The results confirm the above hypothesis. Further studies are needed to compare effect sizes of regulation capacity for approaches with and without instructed strategies.
Syntactic parsing of web queries is important for query understanding. However, web queries usually do not observe the grammar of a written language, and no labeled syntactic trees for web queries are available. In this paper, we focus on a query's clicked sentence, i.e., a well-formed sentence that i) contains all the tokens of the query, and ii) appears in the query's top clicked web pages. We argue such sentences are semantically consistent with the query. We introduce algorithms to derive a query's syntactic structure from the dependency trees of its clicked sentences. This gives us a web query treebank without manual labeling. We then train a dependency parser on the treebank. Our model achieves much better UAS (0.86) and LAS (0.80) scores than state-of-the-art parsers on web queries.
This study examines how preadolescent African American students in Washington, D.C., used a linguistic practice called ‘joning,’ a style of verbal play similar to ritual insults, in peer interactions. Sociolinguists have focused on how children socialize each other into vernacular styles appropriate for peer group use but often assume that they disalign with social and linguistic norms for classroom behavior. Drawing from a nine‐month ethnographic study that the author conducted in an after‐school program, this article analyzes the structure and function of joning as a vernacular style of African American Vernacular English and its uses in constructing classroom identities. Joning often facilitated student learning, but it was perceived as a socially and physically risky linguistic practice because of its uses as conflict talk in the local community. Focusing on preadolescence as a key stage of language socialization, this article shows how minority students modify peer‐learned linguistic practices to pursue academic success on their own terms.
edition) as a disorder that merits further research. The diagnostic criteria are based on those for Substance Use Disorder and Gambling Disorder. Excessive gamblers and persons with Substance Use Disorder show attentional biases towards stimuli related to their addictions. We investigated whether excessive Internet gamers show a similar attentional bias, by using two established experimental paradigms. Methods We measured reaction times of excessive Internet gamers and non-gamers (N = 51, 23.7 ± 2.7 years) by using an addiction Stroop with computer-related and neutral words, as well as a visual probe with computer-related and neutral pictures. Mixed design analyses of variance with the between-subjects factor group (gamer/non-gamer) and the within-subjects factor stimulus type (computer-related/neutral) were calculated for the reaction times as well as for valence and familiarity ratings of the stimulus material. Results In the addiction Stroop, an interaction for group × word type was found: Only gamers showed longer reaction times to computer-related words compared to neutral words, thus exhibiting an attentional bias. In the visual probe, no differences in reaction time between computer-related and neutral pictures were found in either group, but the gamers were faster overall. Conclusions An attentional bias towards computer-related stimuli was found in excessive Internet gamers, by using an addiction Stroop but not by using a visual probe. A possible explanation for the discrepancy could lie in the fact that the visual probe may have been too easy for the gamers.
ABSTRACT This study investigates whether and when differences in the credit rating agencies' methodologies result in differences in rating properties. In particular, this study focuses on differences in information processing constraints between a rating agency that utilizes qualitative analysis and direct access to borrowers' management in its rating process (Standard & Poor's) compared to one that does not (Egan Jones Ratings Company) and how these differences affect rating quality. We find that as information uncertainty about borrowers increases, Egan Jones's rating accuracy, informativeness, and timeliness decrease relative to Standard & Poor's. Our findings suggest that Egan Jones's more restricted rating methodology can lead to limitations in information processing and, thus, reductions in Egan Jones's rating quality advantage for borrowers with greater information uncertainty. JEL Classifications: G10; G24.
Many sequential processing tasks require complex nonlinear transition functions from one step to the next. However, recurrent neural networks with 'deep' transition functions remain difficult to train, even when using Long Short-Term Memory (LSTM) networks. We introduce a novel theoretical analysis of recurrent networks based on Gersgorin's circle theorem that illuminates several modeling and optimization issues and improves our understanding of the LSTM cell. Based on this analysis we propose Recurrent Highway Networks, which extend the LSTM architecture to allow step-to-step transition depths larger than one. Several language modeling experiments demonstrate that the proposed architecture results in powerful and efficient models. On the Penn Treebank corpus, solely increasing the transition depth from 1 to 10 improves word-level perplexity from 90.6 to 65.4 using the same number of parameters. On the larger Wikipedia datasets for character prediction (text8 and enwik8), RHNs outperform all previous results and achieve an entropy of 1.27 bits per character.
Multiple treebanks annotated under heterogeneous standards give rise to the research question of best utilizing multiple resources for improving statistical models. Prior research has focused on discrete models, leveraging stacking and multi-view learning to address the problem. In this paper, we empirically investigate heterogeneous annotations using neural network models, building a neural network counterpart to discrete stacking and multiview learning, respectively, finding that neural models have their unique advantages thanks to the freedom from manual feature engineering. Neural model achieves not only better accuracy improvements, but also an order of magnitude faster speed compared to its discrete baseline, adding little time cost compared to a neural model trained on a single treebank.
The growing work in multi-lingual parsing faces the challenge of fair comparative evaluation and performance analysis across languages and their treebanks. The difficulty lies in teasing apart the properties of treebanks, such as their size or average sentence length, from those of the annotation scheme, and from the linguistic properties of languages. We propose a method to evaluate the effects of word order of a language on dependency parsing performance, while controlling for confounding treebank properties. The method uses artificially-generated treebanks that are minimal permutations of actual treebanks with respect to two word order properties: word order variation and dependency lengths. Based on these artificial data on twelve languages, we show that longer dependencies and higher word order variability degrade parsing performance. Our method also extends to minimal pairs of individual sentences, leading to a finer-grained understanding of parsing errors.
This paper presents a novel latent variable recurrent neural network architecture for jointly modeling sequences of words and (possibly latent) discourse relations between adjacent sentences. A recurrent neural network generates individual words, thus reaping the benefits of discriminatively-trained vector representations. The discourse relations are represented with a latent variable, which can be predicted or marginalized, depending on the task. The resulting model can therefore employ a training objective that includes not only discourse relation classification, but also word prediction. As a result, it outperforms state-of-the-art alternatives for two tasks: implicit discourse relation classification in the Penn Discourse Treebank, and dialog act classification in the Switchboard corpus. Furthermore, by marginalizing over latent discourse relations at test time, we obtain a discourse informed language model, which improves over a strong LSTM baseline.
People intuitively match basic tastes to sounds of different pitches, and the matches that they make tend to be consistent across individuals. It is, though, not altogether clear what governs such crossmodal mappings between taste and auditory pitch. Here, we assess whether variations in taste intensity influence the matching of taste to pitch as well as the role of emotion in mediating such crossmodal correspondences. Participants were presented with 5 basic tastants at 3 concentrations. In Experiment 1, the participants rated the tastants in terms of their emotional arousal and valence/pleasantness, and selected a musical note (from 19 possible pitches ranging from C2 to C8) and loudness that best matched each tastant. In Experiment 2, the participants made emotion ratings and note matches in separate blocks of trials, then made emotion ratings for all 19 notes. Overall, the results of the 2 experiments revealed that both taste quality and concentration exerted a significant effect on participants' loudness selection, taste intensity rating, and valence and arousal ratings. Taste quality, not concentration levels, had a significant effect on participants' choice of pitch, but a significant positive correlation was observed between individual perceived taste intensity and pitch choice. A significant and strong correlation was also demonstrated between participants' valence assessments of tastants and their valence assessments of the best-matching musical notes. These results therefore provide evidence that: 1) pitch-taste correspondences are primarily influenced by taste quality, and to a lesser extent, by perceived intensity; and 2) such correspondences may be mediated by valence/pleasantness.
Natural language parsing is known to potentially produce a high number of syntactic interpretations for a sentence. Some of them may contain multiword expressions (MWEs) and achieving them faster than compositional alternatives proved efficient in symbolic parsing (see below). We propose to apply this strategy to symbolic LTAG (Lexicalized Tree Adjoining Grammar) parsing using an architecture adaptable to probabilistic parsing. We are particularly interested in LTAGs because, according to (Abeille and Schabes 1989), they show several advantages with respect to parsing MWEs. Firstly, unification constraints on feature structures attached to tree nodes allow one to naturally express dependencies between arguments at different depths in the elementary trees (as in NP 0 vider DET sac 'to express one's secret thoughts', where the determiner DET embedded in the direct object must agree in person and number with the subject NP 0). Secondly, the so-called extended domain of locality offers a natural framework for representing two different kinds of discontinuities. Namely, discontinuities coming from the internal structure of a MWE are directly visible in elementary trees and are handled in parsing mostly by substitution. Disconti-nuities coming from insertion of modifiers (e.g. a bunch of NP, a whole bunch of NP) are invisible in elementary trees but are handled in parsing by adjunction. Consider the sentence in example (1). (1) Acid rains in Ghana are equally grim. When it is being scanned by a left-to-right parser, two competing interpretations are syntactically valid for the first 4 words. One of them considers rains as a verb whose subject is acid while, according to the other, rains is the head noun of the NN compound acid rains. Our objective is to propose a parsing strategy which would promote the latter interpretation due the fact that it contains a known MWE. More precisely, the parser should: (i) trivially, admit only grammar-compliant analyses of a sentence, (ii) achieve MWE-oriented interpretations more rapidly than potential compositional interpretations, (iii) eliminate no grammar-compliant interpretations. Note that all these conditions could rather easily be met for sentence (1) in a pre-processing-based approach in which potential MWEs are identified prior to parsing and conflated into word-with-spaces tokens. Such an approach might however lead to a parsing failure in the case of sentence (2) if the two initial tokens are wrongly merged into a nominal compound in the pre-parsing step. In order to avoid errors of this kind, MWE identification and parsing should be performed jointly. (2) Hunger strikes the civilians since 2001. Seminal works, such as (Finkel and Manning 2009, Green et al. 2011, 2013, Constant et al. 2013), show that the results of probabilis-tic MWE identification and/or parsing are improved when both tasks are performed simultaneously. (Wehrli et al. 2010) point out that such an improvement (also within further parsing-based applications, e.g. machine translation) occurs in symbolic parsing (here: in a Chomskian grammar-based approach) when the knowledge about a potential occurrence of MWEs guides the parsing process. Our goal is to apply a similar strategy to the one in (Wehrli et al. 2010), i.e. to systematically promote MWE-oriented interpretations, within LTAG parsing 1 We additionally wish to design the parser architecture in such a way that corpus-based probabilities about MWE contexts can be 1 The parsing algorithm should of course abstract away from the way the input LTAG grammar was obtained (manually crafted, generated from a metagram-mar, or learned from a treebank).