Papers reviewed and determined not to be word norm studies. Use the flag icon to report errors or suggest re-inclusion.
18625 papers
Although some facial expressions provide clear information about people's emotions and intentions (happy, angry), others (surprise) are ambiguous because they can signal both positive (e.g., surprise party) and negative outcomes (e.g., witnessing an accident). Without a clarifying context, surprise is interpreted as positive by some and negative by others, and this valence bias is stable across time. When compared to fearful expressions, which are consistently rated as negative, surprise and fear share similar morphological features (e.g., widened eyes) primarily in the upper part of the face. Recently, we demonstrated that the valence bias was associated with a specific pattern of eye movements (positive bias associated with faster fixation to the lower part of the face). In this follow-up, we identified two participants from our previous study who had the most positive and most negative valence bias. We used their eye movements to create a moving window such that new participants viewed faces through the eyes of one our previous participants (subjects saw only the areas of the face that were directly fixated by the original participants in the exact order they were fixated; i.e., Simulated Eye-movement Experience). The input provided by these windows modulated the valence ratings of surprise, but not fear faces. These findings suggest there are meaningful individual differences in how people process faces, and that these differences impact our emotional perceptions. Furthermore, this study is unique in its approach to examining individual differences in emotion by creating a new methodology adapted from those used primarily in the vision/attention domain. (PsycINFO Database Record (c) 2018 APA, all rights reserved).
Increasingly, citizen scientists do work beyond the primary goal of the project (i.e., advanced work) such as writing articles. These activities often take place in discussion boards and have a set of linguistic norms for contributing. For newcomers, learning this language presents a challenge since there are no formal opportunities for them to learn the language and volunteers who join later need to learn more than volunteers who join earlier in a project life-cycle. In this poster, we examine how newcomers language use shifts over the course of two citizen science projects. We find that, although, newcomers joining later might face obstacles, newcomer language associated with advanced work increase over the project's life-cycle. The analysis can help the science team assess whether newcomers on the talk page have either adopted advanced terminologies or they need to have a more formal resource such as tutorial or blog posts.
Deep neural networks (DNNs) have achieved impressive predictive performance due to their ability to learn complex, non-linear relationships between variables. However, the inability to effectively visualize these relationships has led to DNNs being characterized as black boxes and consequently limited their applications. To ameliorate this problem, we introduce the use of hierarchical interpretations to explain DNN predictions through our proposed method, agglomerative contextual decomposition (ACD). Given a prediction from a trained DNN, ACD produces a hierarchical clustering of the input features, along with the contribution of each cluster to the final prediction. This hierarchy is optimized to identify clusters of features that the DNN learned are predictive. Using examples from Stanford Sentiment Treebank and ImageNet, we show that ACD is effective at diagnosing incorrect predictions and identifying dataset bias. Through human experiments, we demonstrate that ACD enables users both to identify the more accurate of two DNNs and to better trust a DNN's outputs. We also find that ACD's hierarchy is largely robust to adversarial perturbations, implying that it captures fundamental aspects of the input and ignores spurious noise.
The task of nuclearity recognition in Chinese discourse remains challenging due to the demand for more deep semantic information. In this paper, we propose a novel text matching network (TMN) that encodes the discourse units and the paragraphs by combining Bi-LSTM and CNN to capture both global dependency information and local n-gram information. Moreover, it introduces three components of text matching, the Cosine, Bilinear and Single Layer Network, to incorporate various similarities and interactions among the discourse units. Experimental results on the Chinese Discourse TreeBank show that our proposed TMN model significantly outperforms various strong baselines in both micro-F1 and macro-F1.
We present an architecture based on neural networks to generate natural language from unordered dependency trees. The task is split into the two subproblems of word order prediction and morphology inflection. We test our model gold corpus (the Italian portion of the Universal Dependency treebanks) and an automatically parsed corpus from the Web.
Cross-domain sentiment classifiers aim to predict the polarity, namely the sentiment orientation of target text documents, by reusing a knowledge model learned from a different source domain. Distinct domains are typically heterogeneous in language, so that transfer learning techniques are advisable to support knowledge transfer from source to target. Distributed word representations are able to capture hidden word relationships without supervision, even across domains. Deep neural networks with memory (MemDNN) have recently achieved the state-of-the-art performance in several NLP tasks, including cross-domain sentiment classifica- tion of large-scale data. The contribution of this work is the massive experimentations of novel outstanding MemDNN architectures, such as Gated Recurrent Unit (GRU) and Differentiable Neural Computer (DNC) both in cross-domain and in-domain sentiment classification by using the GloVe word embeddings. As far as we know, only GRU neural networks have been applied in cross-domain sentiment classification. Senti- ment classifiers based on these deep learning architectures are also assessed from the viewpoint of scalability and accuracy by gradually increasing the training set size, and showing also the effect of fine-tuning, an ex- plicit transfer learning mechanism, on cross-domain tasks. This work shows that MemDNN based classifiers improve the state-of-the-art on Amazon Reviews corpus with reference to document-level cross-domain sen- timent classification. On the same corpus, DNC outperforms previous approaches in the analysis of a very large in-domain configuration in both binary and fine-grained document sentiment classification. Finally, DNC achieves accuracy comparable with the state-of-the-art approaches on the Stanford Sentiment Treebank dataset in both binary and fine-grained single-sentence sentiment classification.
This paper demonstrates an end-to-end Chinese discourse parser. We propose a unified framework based on recursive neural network (RvNN) to jointly model the subtasks including elementary discourse unit (EDU) segmentation, tree structure construction, center labeling, and sense labeling. Experimental results show our parser achieves the state-of-the-art performance in the Chinese Discourse Treebank (CDTB) dataset. We release the source code with a pre-trained model for the NLP community. To the best of our knowledge, this is the first open source toolkit for Chinese discourse parsing. The standalone toolkit can be integrated into subsequent applications without the need of external resources such as syntactic parser.
The paper describes the conversion of an LFG treebank of Polish into enhanced Universal Dependencies, and—more generally—identifies the kinds of information lost in translation from LFG to UD. The paper also presents the resulting UD treebank of Polish and compares it to the previous UD treebank of Polish.
Web queries with question intent manifest a complex syntactic structure and the processing of this structure is important for their interpretation. However, their algorithms rely on resources typically not available outside of big web corporates. We propose a new BiLSTM query parser that: (1) Explicitly accounts for the unique grammar of web queries; and (2) Utilizes named entity (NE) information from a BiLSTM NE tagger, that can be jointly trained with the parser. In order to train our model we annotate the query treebank of When trained on 2500 annotated queries our parser achieves UAS of 83.5% and segmentation F1score of 84.5, substantially outperforming existing state-of-the-art parsers. 1
The present study features speech errors Italian students make in translation as the most complex form of speech activity. The examined speech errors were made by both students at the Higher School of Translation and budding traslators. The low quality of technical translation and the large number of translation errors determined the scope of the present research: to draw the scientists’ attention to the problem in question, to generalize the translation practice, and to work out recommendations on preventing translation errors. The research employed general scientific methods (generalization, analysis, synthesis) and empirical research methods. The author puts forward some possible causes of error making, e.g. interlingual and cross-language interference, loan translations, etc. The author gives examples of the most typical student errors, indicates the specific reasons behind them, and offers recommendations for error prevention. The results of this study can be used for training and practical purposes. The author comes to the conclusion that it is impossible not to make errors due to the fact that the students cannot overcome interference while in their native linguistic environment. However, it is quite possible to reduce the number of errors. The article is composed of three parts. The introduction specifies the aims and tasks of the work. The main body of the article introduces some specific examples of student errors. In the conclusion, the author sums up recommendations for effective work in the learning process.
This paper describes the system of our team Phoenix for participating CoNLL 2018 Shared Task: Multilingual Parsing from Raw Text to Universal Dependencies. Given the annotated gold standard data in CoNLL-U format, we train the tokenizer, tagger and parser separately for each treebank based on an open source pipeline tool UDPipe. Our system reads the plain texts for input, performs the preprocessing steps (tokenization, lemmas, morphology) and finally outputs the syntactic dependencies. For the low-resource languages with no training data, we use cross-lingual techniques to build models with some close languages instead. In the official evaluation, our system achieves the macro-averaged scores of 65.61%, 52.26%, 55.71% for LAS, MLAS and BLEX respectively.
This chapter reports on the comprehensive reform of an early childhood teacher education program aiming to substantively address the improvement of educational circumstances of culturally and linguistically diverse children. This project engages sociocultural theory, research, and praxis and expands a funds of knowledge perspective. This reform calls for new contexts for action, new forms of family and community engagement, and new circulation of literacies across households, neighborhoods, and U.S. schools. Overarching design goals for the program are focused on challenging the following: (a) narrow conceptions of language literacy and stories; (b) bounded or contained views of learning contexts; (c) hegemonic cultural and linguistic norms, matters regarding U.S. language and educational policies and power; (d) issues of voice and silencing; (e) quantitative and static views of “resources”; and (f) the limited attention to the agency, identities, and strategic actions of diverse students and their families as they traverse home, community, and institutional contexts.
In this paper, we focus on parsing rare and non-trivial constructions, in particular ellipsis. We report on several experiments in enrichment of training data for this specific construction, evaluated on five languages: Czech, English, Finnish, Russian and Slovak. These data enrichment methods draw upon self-training and tri-training, combined with a stratified sampling method mimicking the structural complexity of the original treebank. In addition, using these same methods, we also demonstrate small improvements over the CoNLL-17 parsing shared task winning system for four of the five languages, not only restricted to the elliptical constructions.
With the recent growth in attention to transgender people’s experiences, language has become a focused site of both anxiety and contestation. This chapter focuses on three challenges for trans-inclusive language: how to select gendered labels and pronouns, how to make language more gender-neutral when gender isn’t relevant, and how to talk about gender more precisely when it is relevant, such as in discussions of identity, social inequality, physiology, or socialization. In each case, concrete strategies are presented that reflect the way transgender people themselves reshape language to work in more inclusive and affirming ways. These include more direct negotiation regarding appropriate identifying terms, the selection of gender-neutral or gender-inclusive forms, and avoiding the conflation of different aspects of gender. Although transphobia will not be eliminated through linguistic reform alone, changes to linguistic norms that delegitimize trans identities and normalize cis (non-trans) identities are a critical part of trans liberation.
The purpose of this study was to evaluate the relationship between emotional responses to sounds, hearing acuity, and isolation, specifically objective isolation (social disconnectedness) and subjective isolation (loneliness). It was predicted that ratings of valence in response to pleasant and unpleasant stimuli would influence the relationship between hearing loss and isolation. Participants included 83 adults, without depression, who were categorized into three groups (young with normal hearing, older with normal hearing, and adults with mild-to-moderately severe hearing loss). Participants made ratings of valence in response to pleasant and unpleasant nonspeech sounds, presented at a moderate overall level in the laboratory. Participants also completed questionnaires related to social disconnectedness and loneliness. Data were analyzed using multiple regression with questionnaire scores as dependent variables. Independent variables were age, gender, degree of hearing loss, perceived hearing handicap, number of depressive symptoms, mean valence rating in response to unpleasant sounds, and mean valence rating in response to pleasant sounds. Emotional responses to pleasant sounds explained significant variability in scores of both social disconnectedness and loneliness. Depressive symptoms also explained variability in loneliness scores. Hearing loss was not significantly related to social disconnectedness or loneliness, although it was the only variable significantly related to ratings of valence in response to pleasant sounds. Emotional responses to pleasant sounds are related to disconnectedness and loneliness. Although not related to isolation in this study, hearing loss was related to emotional responses. Thus, emotional responses should be considered in future models of isolation and hearing loss.
We propose a novel way to create categorized discourse lexicons for multiple languages. We combine information from the Penn Discourse Treebank with statistical machine translation techniques on the Europarl corpus. Using gender profiling as an application, we evaluate our approach by comparing it with an approach using features from a knowledge-based lexicon and with an Rhetorical structure theory (RST) discourse parser. Our experiments are performed on corpora for three languages (English, Dutch, and German) in two genres (news and blogs). We include a feature analysis in which we look for (in)consistencies of discourse features related to male and female authors between the different experimental settings.
Answering questions formulated in natural language is a long standing quest in Artificial Intelligence. However, even formulating the problem in precise terms has proven to be too challenging, which lead many researchers to focus on Multiple-Choice Question Answering problems. One particularly interesting type of the latter problem is solving standardized tests such as university entrance exams. The Exame Nacional do Ensino Médio (ENEM) is a High School level exam widely used by Brazilian universities as entrance exam, and the world's second biggest university entrance examination in number of registered candidates. In this work we tackle the problem of answering purely textual multiple-choice questions from the ENEM. We build on a previous solution that formulated the problem as a text information retrieval problem. In particular, we investigate how to enhance these methods by text augmentation using Word Embedding and WordNet, a structured lexical database where words are connected according to some relations like synonymy and hypernymy. We also investigate how to boost performance by building ensembles of weakly correlated solvers. Our approaches obtain accuracies ranging from 26% to 29.3%, outperforming the previous approach.
We propose a novel approach to Vietnamese word segmentation. Our approach is based on the Single Classification Ripple Down Rules methodology (Compton and Jansen, 1990), where rules are stored in an exception structure and new rules are only added to correct segmentation errors given by existing rules. Experimental results on the benchmark Vietnamese treebank show that our approach outperforms previous state-of-the-art approaches JVnSegmenter, vnTokenizer, DongDu and UETsegmenter in terms of both accuracy and performance speed. Our code is open-source and available at: https://github.com/datquocnguyen/RDRsegmenter.
A dependency parser generates both a syntactic structure and a shallow semantic structure of a sentence. It is a fundamental component of natural language processing (NLP) based pipelines, which are critical to facilitate research using the Electronic Health Records (EHR). However, current works mainly apply parsers developed in the general English domain to clinical text. There are no formal evaluations and comparisons of deep learning based dependency parsers in the medical domain. No state-of-the-art dependency parsing performance has been established on clinical text, either. In this study, we investigated the performance of four state-ofthe-art deep learning based dependency parsers, Stanford parser, Bist-parser, dependency_tf parser and jPTDP parser, respectively. Experiments for evaluation are conducted on two datasets: (1) The MiPACQ Treebank and (2) A Treebank of progress notes. Our results showed that the original parsers achieved lower performance in clinical text compared to general English text. After retraining on the clinical Treebank, all parsers obtained better performance. Besides, using word embeddings from Gigaword and MIMICIII yielded comparable performance. Interestingly, the transition-based parsers demonstrated stronger generalizability on different treebanks than the graph-based parsers. Overall, Bist-parser achieved the best performance on MiPACQ (88.95% UAS, 92.69% LS, 86.10% LAS). Stanford parser achieved the best performance on progress notes (84.01% UAS, 89/97% LS, 80.72% LAS).
This paper describes our approach to developing the Turkish PropBank by adopting the semantic role-labeling guidelines of the original PropBank and using the translation of the English Penn-TreeBank as a resource. We discuss the semantic annotation process of the PropBank and language-specific cases for Turkish, the tools we have developed for annotation, and quality control for multiuser annotation. In the current phase of the project, more than 9500 sentences are semantically analyzed and predicate-argument information is extracted for 1330 verbs and 1914 verb senses. Our plan is to annotate 17,000 sentences by the end of 2017.
The contents and structure of semantic memory have been the focus of much recent research, with major advances in the development of distributional models, which use word co-occurrence information as a window into the semantics of language. In parallel, connectionist modeling has extended our knowledge of the processes engaged in semantic activation. However, these two lines of investigation have rarely been brought together. Here, we describe a processing model based on distributional semantics in which activation spreads throughout a semantic network, as dictated by the patterns of semantic similarity between words. We show that the activation profile of the network, measured at various time points, can successfully account for response times in lexical and semantic decision tasks, as well as for subjective concreteness and imageability ratings. We also show that the dynamics of the network is predictive of performance in relational semantic tasks, such as similarity/relatedness rating. Our results indicate that bringing together distributional semantic networks and spreading of activation provides a good fit to both automatic lexical processing (as indexed by lexical and semantic decisions) as well as more deliberate processing (as indexed by ratings), above and beyond what has been reported for previous models that take into account only similarity resulting from network structure.
An important research field in the area of text mining is text categorization. Most of the real world documents are multi-label in nature. In this paper we have proposed a novel method for automated and effective categorization of multi-label text documents. The proposed method is based on lexical and semantics concepts. Tokens are identified in the text documents using standard IEEE taxonomy. To analyze the semantic relationships between tokens, standard lexical database WordNet is used. The proposed method is tested on a dataset of 150 research articles of computer science domain from IEEE Xplore digital library. It has shown a significantly good performance with an accuracy of 75%.
The transfer or share of knowledge between languages is a popular solution to resource scarcity in NLP. However, the effectiveness of cross-lingual transfer can be challenged by variation in syntactic structures. Frameworks such as Universal Dependencies (UD) are designed to be cross-lingually consistent, but even in carefully designed resources trees representing equivalent sentences may not always overlap. In this paper, we measure cross-lingual syntactic variation, or anisomorphism, in the UD treebank collection, considering both morphological and structural properties. We show that reducing the level of anisomorphism yields consistent gains in cross-lingual transfer tasks. We introduce a source language selection procedure that facilitates effective cross-lingual parser transfer, and propose a typologically driven method for syntactic tree processing which reduces anisomorphism. Our results show the effectiveness of this method for both machine translation and cross-lingual sentence similarity, demonstrating the importance of syntactic structure compatibility for boosting cross-lingual transfer in NLP.
We present a complete, automated, and efficient approach for utilizing valency analysis in making dependency parsing decisions. It includes extraction of valency patterns, a probabilistic model for tagging these patterns, and a joint decoding process that explicitly considers the number and types of each token's syntactic dependents. On 53 treebanks representing 41 languages in the Universal Dependencies data, we find that incorporating valency information yields higher precision and F1 scores on the core arguments (subjects and complements) and functional relations (e.g., auxiliaries) that we employ for valency analysis. Precision on core arguments improves from 80.87 to 85.43. We further show that our approach can be applied to an ostensibly different formalism and dataset, Tree Adjoining Grammar as extracted from the Penn Treebank; there, we outperform the previous stateof-the-art labeled attachment score by 0.7. Finally, we explore the potential of extending valency patterns beyond their traditional domain by confirming their helpfulness in improving PP attachment decisions. 1
Standard neural network architectures are non-linear only by virtue of a simple element-wise activation function, making them both brittle and excessively large. In this paper, we consider methods for making the feed-forward layer more flexible while preserving its basic structure. We develop simple drop-in replacements that learn to adapt their parameterization conditional on the input, thereby increasing statistical efficiency significantly. We present an adaptive LSTM that advances the state of the art for the Penn Treebank and WikiText-2 word-modeling tasks while using fewer parameters and converging in less than half as many iterations.
This paper chooses two news-genre dependency treebanks, one in Chinese and one in English and examines the synergetics or the interrelations among dependency tree widths, heights and sentence lengths in the framework of dependency grammar. When sentences grow longer, the dependency trees grow both taller and higher. The growths of heights and widths are competing with each other, resulting in a minimization of dependency distance. The product of the widest layer and its width corresponds to the sentence length. These correlations are integrated into a tentative synergetic syntactic model based on the framework of dependency grammar.
Sarcasm is a sophisticated form of sentiment expression where speaker express their opinions opposite of what they mean. Sarcasm detection and Emotion detection from social net-working sites has been a great field of study. With the growth of e-services such as e-commerce, e-tourism and e-business, the companies are very keen on exploiting emotion and sarcasm analysis for their marketing strategies in order to evaluate the public attitudes towards their brand. Thus efficient emotion and sarcasm modeling system can be a good solution to the above problem. This work aims at developing a system that groups posts based on emotions, sentiment and find sarcastic posts, if present. The proposed system is to develop a prototype that help to come to an inference about the emotions of the posts namely anger, surprise, happy, fear, sorrow, trust, anticipation and disgust with three sentic levels in each. This helps in better understanding of the posts when compared to the approaches which senses the polarity of the posts and gives just their sentiments i.e., positive, negative or neutral. The posts handling these emotions might be sarcastic too. The Sentiment & emotion identification module identifies the sentiment or emotion of the post by evaluating score of each word in the comment which is used by different sarcasm detection methods to detect sarcasm. The emotion identification module uses the lexical databases WordNet, SentiWordNet to find the right sentiment scores for the words with respect to each emotion. It also uses Sarcasm detection algorithms like Emoticon sarcasm detection, Hybrid sarcasm detection, Hashtag Processing, Interjection Word Start (IWT).
Abstract This paper illustrates how enriched diachronic treebank data can shed new light on an old and vexed topic, even when that topic is primarily morphological and semantic in nature rather than syntactic. The topic is the rise of the Russian po delimitatives, a change seen as crucial in most accounts of the history of Russian aspect, since it represents a major step in generalising the derivational aspect system. Earlier accounts concur that the po delimitatives spread fairly recently, too recently for the development to be connected to the loss of the aorist tense, which also had delimitative readings with atelic verbs. Using treebank data from the Tromsø Old Russian and Old Church Slavonic Treebank, enriched with tags for derivational morphology and semantics, I show that the po delimitatives were not marginal even in the earliest Slavic sources, either in terms of frequency or semantics, and that they first complemented and then competed with the delimitative aorists. It can thus be claimed that the exotic po delimitatives grew organically out of the old Indo-European inflectional aspect system.
In this paper, we investigate the usage of the convolutional neural network (CNN) to propose a novel end-to-end language processing structure to model textual data for this task. In particular, we propose a 3D CNN structure for the task, which is featured by spatial pyramid pooling (SPP). To our knowledge, it is the first time that 3D convolution and SPP structure are applied together in language processing issues. Compared with methods of 2D CNNs, the proposed method can effectively and efficiently capture the complicated internal relations in sentences. Furthermore, in previous work, the issue of sentence length variety is usually addressed by padding zero to make all sentences vectors to a fixed length, which causes too much redundant and useless noise. Inspired by the SPP structure for object detection in image processing, this issue can be well handled with the SPP, which divides the sentences into several length sections for respective pooling processing. Experiments are conducted for the task of sentence classification as well as relation classification. Experiments on Stanford Treebank, TREC, subj, and Yelp datasets demonstrate that our proposed method can outperform other state-of-the-art models, with respect to classification accuracy. Auxiliary attempts to leverage our method to SemEval-2010 Task 8 dataset further substantiate the model's capability of extracting features efficiently.
Recurrent neural networks (RNNs) are powerful models of sequential data. They have been successfully used in domains such as text and speech. However, RNNs are susceptible to overfitting; regularization is important. In this paper we develop Noisin, a new method for regularizing RNNs. Noisin injects random noise into the hidden states of the RNN and then maximizes the corresponding marginal likelihood of the data. We show how Noisin applies to any RNN and we study many different types of noise. Noisin is unbiased--it preserves the underlying RNN on average. We characterize how Noisin regularizes its RNN both theoretically and empirically. On language modeling benchmarks, Noisin improves over dropout by as much as 12.2% on the Penn Treebank and 9.4% on the Wikitext-2 dataset. We also compared the state-of-the-art language model of Yang et al. 2017, both with and without Noisin. On the Penn Treebank, the method with Noisin more quickly reaches state-of-the-art performance.
Recently, span-based constituency parsing has achieved competitive accuracies with extremely simple models by using bidirectional RNNs to model "spans". However, the minimal span parser of Stern et al. (2017a) which holds the current state of the art accuracy is a chart parser running in cubic time, O(n 3 ), which is too slow for longer sentences and for applications beyond sentence boundaries such as end-toend discourse parsing and joint sentence boundary detection and parsing. We propose a linear-time constituency parser with RNNs and dynamic programming using graph-structured stack and beam search, which runs in time O(nb 2 ) where b is the beam size. We further speed this up to O(nb log b) by integrating cube pruning. Compared with chart parsing baselines, this linear-time parser is substantially faster for long sentences on the Penn Treebank and orders of magnitude faster for discourse parsing, and achieves the highest F1 accuracy on the Penn Treebank among single model end-to-end systems.
The paper presents a new methodology aimed at acquiring typological evidence from "gold" treebanks for different languages. In particular, it investigates whether and to what extent algorithms developed for assessing the plausibility of automatically produced syntactic annotations could contribute to shed light on key issues of the linguistic typological literature. It reports the first and promising results of a case study focusing on word order patterns carried out on three different languages (English, Italian and Spanish).
Modern solutions for implicit discourse relation recognition largely build universal models to classify all of the different types of discourse relations. In contrast to such learning models, we build our model from first principles, analyzing the linguistic properties of the individual top-level Penn Discourse Treebank (PDTB) styled implicit discourse relations: Comparison, Contingency and Expansion. We find semantic characteristics of each relation type and two cohesion devices---topic continuity and attribution---work together to contribute such linguistic properties. We encode those properties as complex features and feed them into a NaiveBayes classifier, bettering baselines(including deep neural network ones) to achieve a new state-of-the-art performance level. Over a strong, feature-based baseline, our system outperforms one-versus-other binary classification by 4.83% for Comparison relation, 3.94% for Contingency and 2.22% for four-way classification.
Shi, Huang, and Lee (2017a) obtained state-of-the-art results for English and Chinese dependency parsing by combining dynamic-programming implementations of transition-based dependency parsers with a minimal set of bidirectional LSTM features. However, their results were limited to projective parsing. In this paper, we extend their approach to support non-projectivity by providing the first practical implementation of the MH 4 algorithm, an Opn 4 q mildly nonprojective dynamic-programming parser with very high coverage on non-projective treebanks. To make MH 4 compatible with minimal transition-based feature sets, we introduce a transition-based interpretation of it in which parser items are mapped to sequences of transitions. We thus obtain the first implementation of global decoding for non-projective transition-based parsing, and demonstrate empirically that it is more effective than its projective counterpart in parsing a number of highly non-projective languages.
We study the problem of analyzing tweets with Universal Dependencies. We extend the UD guidelines to cover special constructions in tweets that affect tokenization, part-of-speech tagging, and labeled dependencies. Using the extended guidelines, we create a new tweet treebank for English (Tweebank v2) that is four times larger than the (unlabeled) Tweebank v1 introduced by Kong et al. (2014). We characterize the disagreements between our annotators and show that it is challenging to deliver consistent annotation due to ambiguity in understanding and explaining tweets. Nonetheless, using the new treebank, we build a pipeline system to parse raw tweets into UD. To overcome annotation noise without sacrificing computational efficiency, we propose a new method to distill an ensemble of 20 transition-based parsers into a single one. Our parser achieves an improvement of 2.2 in LAS over the un-ensembled baseline and outperforms parsers that are state-of-the-art on other treebanks in both accuracy and speed.
Sentiment analysis of short texts such as single sentence has been a research hotspot of naturallanguage processing (NLP). There still exists the challenge of effectively handling the problem with limited contextual information and semantic features. Hence, in this paper, an attention-based bidirectional LSTM neural network (AB-BiLSTM) is proposed to solve the problem. The proposed model can attend the qualitative and informative parts and learn semantic features from both directions of a sentence to perform sentiment analysis of short texts. The proposed model mainly contributes to take advantage of the attention mechanism to capture the informative parts of a sentence without any syntactic features and lexicon features. The model is conducted on the Stanford Sentiment Treebank dataset and the Movie Review Data provided by Cornell University for single sentence sentiment analysis of binary classification. The experimental results indicate that the proposal in this study has superior performances over theexisting methods, without taking the attention mechanism into account.
Grammar induction is the task of learning syntactic structure without the expert-labeled treebanks (Charniak and Carroll, 1992; Klein and Manning, 2002). Recent work on latent tree learning offers a new family of approaches to this problem by inducing syntactic structure using the supervision from a downstream NLP task (Yogatama et al., 2017; Maillard et al., 2017; Choi et al., 2018). In a recent paper published at ICLR, Shen et al. (2018) introduce such a model and report near state-of-the-art results on the target task of language modeling, and the first strong latent tree learning result on constituency parsing. During the analysis of this model, we discover issues that make the original results hard to trust, including tuning and even training on what is effectively the test set. Here, we analyze the model under different configurations to understand what it learns and to identify the conditions under which it succeeds. We find that this model represents the first empirical success for neural network latent tree learning, and that neural language modeling warrants further study as a setting for grammar induction.
The paper analyses a corpus of metalinguistic texts (grammars, remarks) from the first grammar of French (Palsgrave 1972 [1530]) until today in order to retrace the metalinguistic perceptions and the prescriptive interventions in the domain of the French adverb. It is shown that adverbs ending in <italic>-ment</italic> and paraphrases were favored to the detriment of the old oral tradition of using short adverbs. In the second part, three methods are suggested for measuring the effect metalinguistic considerations and norms have on use. This innovative approach aims at bringing together the diachrony of grammaticography with the diachrony of use. In fact, a more objective approach is crucially needed in the domain of linguistic norm where speculation, polemic argumentation and ideology prevail.
Our study aims to explore how much information about areal patterns of colexification we can gain from lexical databases such as CLICS and ASJP. We adopt a bottom-up (rather than hypothesis-driven) approach, identifying areal patterns in three steps: (i) determine spatial autocorrelations in the data, (ii) identify clusters as candidates for convergence areas and (iii) test the clusters resulting from the second step controlling for genealogical relatedness. Moreover, we identify a (genealogical) diversity index for each cluster. This approach yields promising results, which we regard as a proof of concept, but we also point out some drawbacks of the use of major lexical databases.
English part-of-speech taggers regularly make egregious errors related to noun-verb ambiguity, despite having achieved 97%+ accuracy on the WSJ Penn Treebank since 2002. These mistakes have been difficult to quantify and make taggers less useful to downstream tasks such as translation and text-to-speech synthesis. This paper creates a new dataset of over 30,000 naturally-occurring non-trivial examples of noun-verb ambiguity. Taggers within 1% of each other when measured on the WSJ have accuracies ranging from 57% to 75% accuracy on this challenge set. Enhancing the strongest existing tagger with contextual word embeddings and targeted training data improves its accuracy to 89%, a 14% absolute (52% relative) improvement. Downstream, using just this enhanced tagger yields a 28% reduction in error over the prior best learned model for homograph disambiguation for textto-speech synthesis.
Computational text-level discourse analysis mostly happens within Rhetorical Structure Theory (RST), whose structures have classically been presented as constituency trees, and relies on data from the RST Discourse Treebank (RST-DT); as a result, the RST discourse parsing community has largely borrowed from the syntactic constituency parsing community. The standard evaluation procedure for RST discourse parsers is thus a simplified variant of PARSEVAL, and most RST discourse parsers use techniques that originated in syntactic constituency parsing. In this article, we isolate a number of conceptual and computational problems with the constituency hypothesis. We then examine the consequences, for the implementation and evaluation of RST discourse parsers, of adopting a dependency perspective on RST structures, a view advocated so far only by a few approaches to discourse parsing. While doing that, we show the importance of the notion of headedness of RST structures. We analyze RST discourse parsing as dependency parsing by adapting to RST a recent proposal in syntactic parsing that relies on head-ordered dependency trees, a representation isomorphic to headed constituency trees. We show how to convert the original trees from the RST corpus, RST-DT, and their binarized versions used by all existing RST parsers to head-ordered dependency trees. We also propose a way to convert existing simple dependency parser output to constituent trees. This allows us to evaluate and to compare approaches from both constituent-based and dependency-based perspectives in a unified framework, using constituency and dependency metrics. We thus propose an evaluation framework to compare extant approaches easily and uniformly, something the RST parsing community has lacked up to now. We can also compare parsers’ predictions to each other across frameworks. This allows us to characterize families of parsing strategies across the different frameworks, in particular with respect to the notion of headedness. Our experiments provide evidence for the conceptual similarities between dependency parsers and shift-reduce constituency parsers, and confirm that dependency parsing constitutes a viable approach to RST discourse parsing.
Abstract This paper presents a treebank-based study of the effect the text form (prose vs. verse) has on the course of two grammatical changes in Medieval French: the loss of null subjects and the loss of OV word order. By means of statistical analysis, we demonstrate that naive estimates of the spread of overt subjects and VO orders give the impression that there is a significant difference between the rates of development in prose vs. verse. By contrast, estimates based on an abstract grammar competition model which distinguishes between grammar-ambiguous surface forms (overt personal subjects, null subjects in coordination contexts) and grammar-unambiguous surface forms (overt expletive subjects, null subjects in non-coordination contexts) show prose-verse parallelism, prose having an earlier change onset, in line with traditional intuitions. At a more general level, these results suggest that the product of the interaction of a particular grammar with universal pragmatic laws is constant, which can be observed if the factors responsible for variation in grammatical choices are controlled for.
In this paper, we explore the linguistic factors that influence an author's choice of discourse connectives in the production of a coherent text. We focus on the competition between so-called primary connectives (grammaticalized and mostly one-word expressions such as therefore) and secondary connectives (not yet fully grammaticalized compositional discourse phrases such as for this reason). We attempt to describe the linguistic constraints on and preferences in connective selection. The analysis is based on manually annotated data from the Prague Discourse Treebank 2.0 (PDiT), which contains almost 50000 sentences from Czech newspaper texts. We demonstrate that discourse connectives are used in accordance with the economy principle in language, i.e. authors aim to achieve the maximal result with minimal effort. They most frequently choose short and semantically more generalized primary connectives. However, in cases where the discourse relations can be misunderstood, authors prefer more complex and specific structures.
Nous nous intéressons à la performance du langage et aux contraintes causées par la capacité cognitive. Miller (1956) a montré que la mémoire à court-terme est limitée à 7±2 éléments, mais aujourd’hui, cette limite est actualisée à 4 selon Cowan (2001). La première question que nous nous posons est: comment mesurer la complexité syntaxique à partir des données observables considérant ce contrainte de la mémoire à court-terme? Deux autres questions que nous voudrions résoudre sont: quelles sont les relations syntaxiques qui créent vraiment de la complexité? Lorsqu’il s’agit d’une structure syntaxique plus (ou moins) complexe, quels sont les phénomènes qu’on pourrait observer dans nos données?Pour répondre à la première question, je me concentre sur des différents mesures du flux de dépendances (Kahane et al., 2017; Yan & Kahane, 2018). Notre hypothèse est que la mesure du flux à partir de treebanks annotés en syntaxe de dépendance permettraient de vérifier les hypothèses sur la mémoire à court-terme. Les deux autres questions alimentent mon actuel travail de thèse sur les configurations du flux de dépendance, ainsi que les liens entre flux et structure prosodique.