Papers reviewed and determined not to be word norm studies. Use the flag icon to report errors or suggest re-inclusion.
18625 papers
The article is the first attempt to single out and describe categorical features of the fixed phrase scheme «Самое + V inf!». The status of the investigated syntactic model as a sentence of the phraseologized type was confirmed. The fixed phrase scheme expresses two enantiosemic meanings and is characterized by the existence of two structural variants. The first one is represented by the compulsory unchangeable component самое in combination with infinitive; herewith the initial meaning of the lexeme самое is absolutely de-actualized, both compulsory components do not have paradigmatic characteristics. The non-grammatical combination of the defining pronoun with the infinitive determines a high degree of phraseologization of the syntactic sentence model. The inner form of the lexeme самое is modern when the compulsory component самое stands with a noun, which leads to the maintaining of its morphological paradigm. The fixed phrase scheme does not possess syntactic paradigm. The idiomaticity of the first fixed phrase scheme variant is determined by the nonderivability of the meaning «expediency / inexpediency of smth» from the meanings of the separate lexical components; the nonderivability of the modus meaning of the component (semes «approval», «intensity», «expressivity»), the nonderivability of the syntactic meaning of the sentence model that is formally represented by the combination of the defining pronoun and the infinitive, contradictory to the grammatical norms, as well as of the stylistic seme «colloquialism». Besides, when this sentence is analyzed logically we can distinguish two propositions («now it is time» and «to learn»), which are represented by the simple sentence. Thus, the syntactic seme represented the meaning of structural and semantic sentence type is also idiomatic. While realizing the second, enantiosemic meaning of the fixed phrase scheme, the semes «negation», the modus component (semes «irony», «expressivity», etc.), syntactic seme and colloquial styleme are idiomatic. All these make the fixed phrase scheme the effective tool of communication in the conditions of the colloquial speech.
Abstract Achieving optimal health and well-being among African American older adults with chronic conditions requires addressing the needs of their caregivers. This study aimed to elucidate how African American females caring for older adults view health and the factors that influence health. We identified African American women ages 24 to 64 caring for an adult 60 years or older for group concept mapping (GCM), a mixed-methods approach. Participants (N=25) first completed idea generation by providing unlimited short, free-text responses to the focus prompt, “A healthy life for a caregiver includes: ___.” The 512 identified factors were reduced to 99 unique ideas. Participants then sorted the 99 ideas into clusters based on conceptual similarity and rated each idea on desirability and familiarity. Ratings were recorded on a 5-point Likert scale, ranging from very undesirable to extremely desirable and not at all familiar to extremely familiar. Data were analyzed and mapped via CS Global Max software. A cluster map with 12 outcome domains best fits the data. Identified clusters included: (1) Spirituality, (2) Maintaining relationships, (3) Good character, (4) Action to cope, (5) Preserving self, (6) Support, (7) Personal empowerment, (8) Resources, (9) Release (10) Striving for peace, (11) Wellness, (12) Self-care. Seven of the 99 ideas (representing 5 of 12 domains) were rated as desirable but unfamiliar (“go zone”). We identified elements necessary for health and wellbeing from the perspective of African American caregivers. Go-zone items represent opportunities to intervene to promote the health of African American women caring for older adults.
We study how to leverage off-the-shelf visual and linguistic data to cope with out-of-vocabulary answers in visual question answering task. Existing large-scale visual datasets with annotations such as image class labels, bounding boxes and region descriptions are good sources for learning rich and diverse visual concepts. However, it is not straightforward how the visual concepts can be captured and transferred to visual question answering models due to missing link between question dependent answering models and visual data without question. We tackle this problem in two steps: 1) learning a task conditional visual classifier, which is capable of solving diverse question-specific visual recognition tasks, based on unsupervised task discovery and 2) transferring the task conditional visual classifier to visual question answering models. Specifically, we employ linguistic knowledge sources such as structured lexical database (e.g. WordNet) and visual descriptions for unsupervised task discovery, and transfer a learned task conditional visual classifier as an answering unit in a visual question answering model. We empirically show that the proposed algorithm generalizes to out-of-vocabulary answers successfully using the knowledge transferred from the visual dataset.
It is widely accepted that the human cognitive system organizes perceptual input into complex hierarchical descriptions which can be represented by tree structures. Tree structures have been used to describe linguistic, musical and visual perception. In this paper, we will investigate whether there exists an underlying model that governs perceptual organization in general. Our key idea is that the cognitive system strives for the simplest structure (the “simplicity principle”), but in doing so it is biased by the likelihood of previous experiences (the “likelihood principle”). We will present a model which combines these two principles by balancing the notion of most likely tree with the notion of shortest derivation. Experiments with linguistic and musical benchmarks (Penn Treebank and Essen Folksong Collection) show that such a combination outperforms models that are based on either simplicity or likelihood alone.
A lexical database of eastern Indonesia
It is essential in emotion science to have standardized stimuli to aid in replication as well as the elicitation of basic defensive and appetitive states. There are numerous standardized stimulus sets for eliciting emotion perception including scenes, faces, videos, and audio scripts (Coan & Allen, 2007). These stimuli are typically rated subjectively on pleasantness and activation/arousal scales. There are, however, few studies that provide normative pleasantness and arousal ratings from scripts useful in predicting emotional states for future studies on emotional imagery. The goals of the current study were to add to the literature 1) standardized pleasant, neutral, and unpleasant text-driven scripts rated on pleasantness and arousal, and 2) a more substantial set of stimuli that could be useful in electrophysiological studies requiring large amounts of experimental trials. Thirty-five undergraduate students were presented 133 text-driven scripts, which were classified as pleasant (i.e., excitement, erotica, relaxation), neutral, or unpleasant (i.e., contamination, embarrassment, threatening). While reading the script, participants actively imagined themselves in the presented scenario and then provided pleasantness and arousal ratings. As expected, from pleasant to neutral to unpleasant conditions there was a positive linear trend with pleasantness ratings (p<.001). Further, from pleasant to neutral conditions arousal ratings declined, whereas from neutral to unpleasant conditions arousal ratings rose. Thus, we found a significant quadratic trend for the arousal ratings (p<.001). Taken together, these preliminary data indicate that these novel text-driven scripts are eliciting the predicted emotional responses. Keywords: imagery, emotion, text-driven scripts, arousal, hedonic valence, norming study
Abstract Semantically ambiguous and emotional words occur frequently in language, and the different meanings of ambiguous words can sometimes have different emotional loads. For example, the Spanish word heroína (heroin/heroine) can refer to a drug or to a woman who performs a heroic act. Because both ambiguity and emotionality affect word processing, there is a need for normative databases that include data on the emotionality of the different meanings of such words. Thus far, no bases of this type are available in Spanish. With this in mind, the current study will present meaning-dependent affective (valence) ratings for 252 Spanish ambiguous words. The analyses performed show that (a) among ambiguous words, those words with meanings that have distinct affective valence are quite frequent, (b) ambiguous words rated as neutral in isolation can have meanings of opposite valence (i.e., negative-positive or positive-negative), and (c) the valence estimated for ambiguous words in isolation is better explained by the weighted average of the valence of their meanings by dominance. A database of this kind can be useful both for basic research (e.g., relationship between emotion and language and ambiguity processing) and for applied research (e.g., cognitive and emotional biases in emotional disorders and second language learning).
. Observers (N=56, 24 males, mean age=24.3±4.6) sniffed 40 randomly presented odors (27 food and 13 non-food smells) with a wide range of perceived hedonic valences, ranging from unpleasant and arousing (e.g., feces, fish), via pleasant and calming (e.g., clove, cinnamon) to pleasant and stimulating (e.g., peach, caramel). After sniffing, the participants reported their affective appraisal of each odor using the EmojiGrid. This Excels sheet lists the resulting mean valence and arousal ratings (with their SD values), together with corresponding mean ratings that were obtained with alternative rating tools in previous studies in the literature.
Abstract This paper presents the Slovene Training Corpus ssj500k 2.2, which has been annotated on the levels of tokenization, sentence segmentation, part-of-speech tagging, lemmatization, syntactic dependencies, named entities, verbal multi-word expressions, and semantic role labeling. It describes the individual layers of annotation and shows the scope of using the training corpus in the production of various lexicons, such as the lexicon of multi-word units and the valency lexicon of modern Slovene. It concludes by presenting our future work, i.e. the annotation of multi-word expressions based on the Slovene Lexical Database.
Meishan Zhang, Yue Zhang, Guohong Fu. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP). 2019.
This dataset is a PostgreSQL dump of a database generated using autocode_big from the data in the LexiRumah lexical database of eastern Indonesia and Timor-Leste. The dataset contains the forms from the parent dataset, together with automatically generated sound correspondence scorers, pairwise similarity scores, cognate classes, and alignments, all created using Lingpy's LexStat algorithm.
STARK is a python-based command-line tool for extraction of dependency trees from parsed corpora, aimed at corpus-driven linguistic investigations of syntactic and lexical phenomena of various kinds. It takes a treebank in the CONLL-U format as input and returns a list of all relevant dependency trees with frequency information and other useful statistics, such as the strength of association between the nodes of a tree, or its significance in comparison to another treebank. \n \nFor installation, execution and the description of various user-defined parameter settings, see the official project page at: https://github.com/clarinsi/STARK \n \nIn comparison with v1, this version introduces several new features and improvements, such as the option to set parameters in the command line, compare treebanks or visualise results online.
The Dictionary of the Clothing Terminology of the Zilja Dialect in Canale Valley (Kanalska dolina – Val Canale – Kanaltal – Valcjanâl) is the result of dialectological research team-work, conducted between 2003–2014. It was designed as a pilot project for the documentation of endangered variety of Gailtal dialect in the linguistically mixed area of Canale Valley. Considering the importance of preservation of the words for culturally relevant terms and concepts, the topic was chosen by the members of a speech community themselves. The sound material was gathered with various fieldwork techniques, from workshops on the old clothing dialect terms, classic dialect elicitations and guided conversations to monitoring spontaneous speech. As an additional source we used spoken texts, prepared for Slovenian Broadcasting Radio Trst (Trieste) A and the cultural heritage project Zborzbirk. The dictionary contains 594 entries and is based on approximately 1,400 sound clips from around 16 hours of recordings. It is formatted as a trilingual (Slovene-German-Italian) concordance-based dictionary and of which the most common collocations are presented alongside clothing terminology. Each entry is equipped with sound clips for dialect lemma and examples of dialect use (http://hdl.handle.net/11356/1220). The encyclopaedic information is contained in illustrative photographs of traditional clothing items (http://hdl.handle.net/11356/1221) or added as occasional commentary subsection. The documentary section presents the relations between the dialect and standard lexicon and documents the potential geographical area of the word. The multimedia lexical database is organised as a source for grammatical description of the local dialect varieties, especially phonology and morphology. The morphological section of the dictionary thus contains all word forms, confirmed in the speech corpus, with special attention to individual phonetic and inflectional variation. For that purpose, the narrow phonetic transcription (Slovene dialectological transcription) was used for transcribing dialect material. The previous editions were published as printed book: Kenda-Jež, Karmen, Shranli smo jih v bančah: slovarski prispevek k poznavanju oblačilne kulture v Kanalski dolini = contributo lessicale alla conoscenza dellʼabbigliamento in Val Canale, Ukve: S.K.S. Planika Kanalska dolina; [s. l.]: Slori, ATS Od me-je; Ljubljana: Inštitut za slovenski jezik Frana Ramovša ZRC SAZU = Istituto per la lingua slovena “Fran Ramovš” CRS ASSA, ¹2007; Ljubljana: Založba ZRC, ZRC SAZU = Casa editrice CRS, CRS ASSA, ²2015.
WordNet is a large lexical database of English language, in which, nouns, verbs, adjectives, and adverbs are grouped into sets of cognitive synonyms (synsets). Each synset expresses a distinct concept. Synsets are interlinked by both semantic and lexical relations. WordNet is essentially used for word sense disambiguation, information retrieval, and text translation. In this paper, we propose several automatic methods to extract Information and Communication Technology (ICT)-related data from Princeton WordNet. We, then, add these extracted data to our Persian WordNet. The advantage of automated methods is reducing the interference of human factors and accelerating the development of our bilingual ICT WordNet. In our first proposed method, based on a small subset of ICT words, we use the definition of each synset to decide whether that synset is ICT. The second mechanism is to extract synsets which are in a semantic relation with ICT synsets. We also use two similarity criteria, namely LCS and S3M, to measure the similarity between a synset definition in WordNet and definition of any word in Microsoft dictionary. Our last method is to verify the coordinate of ICT synsets. Results show that our proposed mechanisms are able to extract ICT data from Princeton WordNet at a good level of accuracy.
Most of the studies that used normative ratings to compare emotional intensity of words have used a monolingual sample and then compared the emotionality between languages. This study compared valence ratings, i.e. pleasantness ratings of the same positive and negative emotion-laden words, presented in L1 Croatian and L2 English. All participants (n=27) were Croatian young learners at the age of 11 (Mage= 11.1), who were learning English in an instructed context. The participants were presented with a booklet containing eight English words (four positive and four negative) and their Croatian equivalents. They had to rate how pleasant or unpleasant each word was using a 9-point scale. The results showed that L1 emotion-laden words were more extreme in valence. However, in some cases, L2 words were slightly higher rated than L1 words. These results contribute to the bilingualism research by highlighting that differences in emotion processing of different languages could be noticeable in a dimension such as valence.
The BHSA (Biblia Hebraica Stuttgartensia Amstelodamensis) is the BHS text plus the linguistic annotations of the Eep Talstra Centre for Bible and Computer.
 The BHSA is available as a data set in Text-Fabric format. Text-Fabric is a minimalistic model to represent text: it provides addresses for all textual objects, so that it is easy to add arbitrary information at all textual levels, precisely and firmly anchored. A Text-Fabric resource resembles an IKEA ware house. The parts are nicely separated and stacked, so that they can be retrieved easily, to be combined into meaningful output later on. A consequence is that different teams with divergent purposes still can add to the same body of work, with a minimum of interference or duplication of work. Text-Fabric has helped with various types of data construction work, of which the most visible is the website SHEBANQ. We focus on two recent data combination jobs, (A) treebanks from the BHSA data and (B) a detailed comparison of the morphology in the BHSA and in the Open Scriptures effort. As the OSM is not yet finished, the comparison is repeatable.
Abstract Introduction Drug cue reactivity (DCR) is widely used in experimental settings for both assessment and intervention. There is no validated database of pictorial cues available for methamphetamine and opioids. Methods 360 images in three-groups (methamphetamine, opioid and neutral (control)) matched for their content (objects, hands, faces and actions) were selected in an initial development phase. 28 participants with a history of both methamphetamine and opioid use (37.1 ± 8.11 years old, 12 female) with over six months of abstinence were asked to rate images for craving, valence, arousal, typicality and relatedness. Results All drug images were differentiated from neutral images. Drug related images received higher arousal and lower valence ratings compared to neutral images (craving (0-100) for neutral (11.5±21.9), opioid (87.7±18.5), and methamphetamine (88±18), arousal (1-9) for neutral (2.4±1.9), opioid (4.6±2.7), and methamphetamine (4.6±2.6), and valence (1-9) for neutral (4.8±1.3), opioid (4.4±1.9), and methamphetamine (4.4±1.8)). There is no difference between methamphetamine and opioid images in craving, arousal and valence. There is a significant positive relationship between the amount of time that participants spent on drug-related images and the craving they reported for the image. Every 10 points of craving were associated with an increased response time of 383millisecond. Three image sets were automatically selected for equivalent fMRI tasks (methamphetamine and opioids) from the database (tasks are available at github). Conclusion LIBR MOCD provides a resource of validated images/tasks for future DCR studies. Additionally, researchers can select several sets of unique but equivalent images based-on their psychological/physical characteristics for multiple assessments/interventions.
WordNet is a large lexical database originally conceived as a model of human semantic memory. While its design loosely follows ontological principles, it was developed by linguists and psychologists without the benefit of input from philosophers. WordNet's unexpected popularity as a tool for Natural Language Processing and Knowledge Engineering revealed the irregularities of the lexicon and the need to clearly distinguish the lexical from the conceptual level consistent with ontological research, as articulated in the work of Nicola Guarino and his colleagues. This paper reviews WordNet's development from a lexicon to a resource that incorporates some ontological principles.
Named entity recognition (NER) is widely used in natural language processing applications and downstream tasks. However, most NER tools target flat annotation from popular datasets, eschewing the semantic information available in nested entity mentions. We describe NNE-a fine-grained, nested named entity dataset over the full Wall Street Journal portion of the Penn Treebank (PTB). Our annotation comprises 279,795 mentions of 114 entity types with up to 6 layers of nesting. We hope the public release of this large dataset for English newswire will encourage development of new techniques for nested NER.
Parsers are available for only a handful of the world's languages, since they require lots of training data. How far can we get with just a small amount of training data? We systematically compare a set of simple strategies for improving low-resource parsers: data augmentation, which has not been tested before; cross-lingual training; and transliteration. Experimenting on three typologically diverse low-resource languages---North Sámi, Galician, and Kazah---We find that (1) when only the low-resource treebank is available, data augmentation is very helpful; (2) when a related high-resource treebank is available, cross-lingual training is helpful and complements data augmentation; and (3) when the high-resource treebank uses a different writing system, transliteration into a shared orthographic spaces is also very helpful.
A lot of lexical research conducted by experts use existing words in the English Thesaurus. However, the English Thesaurus only gives synonyms of the words searched for and does not provide similarities between words that can be called WordNet. So in this study, a WordNet can be made that can help research that uses databases for English. The similarity between words makes many researchers today are still looking for word relations by manual method, or still use the English Thesaurus. The making of WordNet is expected to be very useful for researchers who want a lexical database for their research, which is currently still relatively small. Therefore it is better to make an English WordNet which will later accommodate words that have the same meaning or Synonym Sets and this WordNet focuses on grouping those words. So that researchers can do lexical research more broadly and unlimitedly with the existence of words that are still unclear in their similarities. Calculation with clustering gets an F1 Score Result at 10.68%, Recall at 8.53% and Precision at 14.28%.
Perception of emotional valence and emotional memory performance vary across the menstrual cycle. However, the consequences of altered ovarian hormone levels due to the intake of hormonal contraceptives on these emotional and cognitive processes remain to be established. In the present study, which included 2169 healthy young females, we show that hormonal contraceptives (HC) users rated emotional pictures as more emotional than HC-non-users and outperformed non-users in terms of better memory recall of emotional pictures. The observed association between HC-status and memory performance was partially mediated by the perception of emotional picture valence, indicating that increased valence ratings of emotional pictures in HC-users led to their better emotional memory performance. These findings extend the knowledge about the relation of HC-intake with emotional valence perception and emotional memory performance. Further, the findings might stimulate further research investigating the interrelation of enhanced memory for emotional events and the increased risk for anxiety-related psychiatric disorders in women.
This study focuses on a comprehensive analysis and manual re-annotation of the Turkish IMST-UD Treebank, which was automatically converted from the IMST Treebank (Sulubacak et al., 2016b). In accordance with the Universal Dependencies' guidelines and the necessities of Turkish grammar, the existing treebank was revised. The current study presents the revisions that were made alongside the motivations behind the major changes. Moreover, it reports the parsing results of a transition-based dependency parser and a graph-based dependency parser obtained over the previous and updated versions of the treebank. In light of these results, we have observed that the re-annotation of the Turkish IMST-UD treebank improves performance with regards to dependency parsing.
Head-driven phrase structure grammar (HPSG) enjoys a uniform formalism representing rich contextual syntactic and even semantic meanings. This paper makes the first attempt to formulate a simplified HPSG by integrating constituent and dependency formal representations into head-driven phrase structure. Then two parsing algorithms are respectively proposed for two converted tree representations, division span and joint span. As HPSG encodes both constituent and dependency structure information, the proposed HPSG parsers may be regarded as a sort of joint decoder for both types of structures and thus are evaluated in terms of extracted or converted constituent and dependency parsing trees. Our parser achieves new state-of-the-art performance for both parsing tasks on Penn Treebank (PTB) and Chinese Penn Treebank, verifying the effectiveness of joint learning constituent and dependency structures. In details, we report 95.84 F1 of constituent parsing and 97.00% UAS of dependency parsing on PTB.
We introduce the first German treebank for Twitter microtext, annotated within the framework of Universal Dependencies. The new treebank includes over 12,000 tokens from over 500 tweets, independently annotated by two human coders. In the paper, we describe the data selection and annotation process and present baseline parsing results for the new testsuite.
This paper presents challenges and observations on creating a code-switching treebank based on ongoing annotation efforts of a Turkish-German spoken corpus following the Universal Dependencies annotation scheme. We present and discuss a number of issues that arise because of the need for consistent multilingual annotation within a single treebank, as well as the informal language which is where code-switching is observed most. Besides proposing solutions to these issues, our aim in this paper is to stimulate discussion and facilitate consistency over upcoming code-switching annotation projects.
This paper shows the extent to which treebanks of Ancient Greek play a central role in the ongoing Pedalion project at the University of Leuven. Building on diverse treebanks readily available today, the project aims to make progress in the automated parsing of classical and postclassical Greek texts. Rather than developing new technology as such, our project endeavours to make deliberate and methodical use of the technology that already exists, essentially by combining and adapting both technology and data. This contribution offers a 'roadmap' of our project, surveying (a) the existing work on which we can rely, (b) the strategies which we adopt to reach better results in the automated processing of Ancient Greek and (c) the deliverables that have already been realised or are forthcoming.
A constituency treebank is a key component for deep syntactic parsing of natural language sentences. For Indonesian, this task is unfortunately hindered by the fact that the only one constituency treebank publicly available is rather small with just over 1000 sentences, and not only that, it employs a format incompatible with readily available constituency treebank processing tools. In this work, we present a conversion of the existing Indonesian constituency treebank to the widely accepted Penn Treebank format. Specifically, the conversion adjusts the bracketing format for compound words as well as the POS tagset according to the Penn Treebank format. In addition, we revised the word segmentation and POS tagging of a number of tokens. Finally, we performed an evaluation on the treebank quality by employing the Shift-Reduce parser from Stanford CoreNLP to create a parser model. A 10-fold cross-validated experiment on the parser model yields an F1-score of 70.90%.
In this paper we present a dependency treebank of travel domain sentences in Modern Standard Arabic. The text comes from a translation of the English equivalent sentences in the Basic Traveling Expressions Corpus. The treebank dependency representation is in the style of the Columbia Arabic Treebank. The paper motivates the effort and discusses the construction process and guidelines. We also present parsing results and discuss the effect of domain and genre difference on parsing.
Study Context: The question of whether relationships between valence and arousal might differ among older and younger adults has not yet been totally clarified. Previous studies focused on only age-related mean-differences, but in the current study mean differences and variance in emotional ratings for the International Affective Picture Systems (IAPS) were both examined in Japanese older and younger adults. METHODS: Participants were 31 older adults (69 ± 5.17 years) and 31 younger adults (19 ± 0.77 years). Each picture was projected on the screen for about 5 s in random order and participants subsequently rated its valence from "unhappy" to "happy" and its arousal from "calm" to "exciting," using 9-point scales. RESULTS: Pearson's correlation analysis showed that positive and negative valence tended to be negatively correlated with arousal in both age groups. The 95% Confidence Intervals for positive arousal in older adults included those of younger adults. Arousal ratings for negative pictures were higher than those for neutral pictures, and those for neutral pictures were higher than those for positive pictures in older adults. There were no significant differences between arousal ratings for neutral pictures and positive pictures in younger adults. CONCLUSION: Older adults tended to rate the pictures as more arousing, with higher arousal ratings for negative pictures than for positive pictures, and the variance of positive arousal in older adults was the highest variance for all conditions. The results of this study suggest that older adults may be sensitive to harmful negative experiences in order to make them less aversive.
Transformers have a potential of learning longer-term dependency, but are limited by a fixed-length context in the setting of language modeling. We propose a novel neural architecture Transformer-XL that enables learning dependency beyond a fixed length without disrupting temporal coherence. It consists of a segment-level recurrence mechanism and a novel positional encoding scheme. Our method not only enables capturing longer-term dependency, but also resolves the context fragmentation problem. As a result, Transformer-XL learns dependency that is 80% longer than RNNs and 450% longer than vanilla Transformers, achieves better performance on both short and long sequences, and is up to 1,800+ times faster than vanilla Transformers during evaluation. Notably, we improve the state-ofthe-art results of bpc/perplexity to 0.99 on en-wiki8, 1.08 on text8, 18.3 on WikiText-103, 21.8 on One Billion Word, and 54.5 on Penn Treebank (without finetuning). When trained only on WikiText-103, Transformer-XL manages to generate reasonably coherent, novel text articles with thousands of tokens. Our code, pretrained models, and hyperparameters are available in both Tensorflow and PyTorch 1.
Resources for syntactic parsing for Indonesian are very limited, as there are only two dependency treebanks publicly available and both are small in size.Not only that, we found out that the word segmentation method used by both treebanks needs improvement.Therefore, in this work we proposed a revision for one of these treebanks, Indonesian Parallel Universal Dependencies treebank.Besides improving word segmentation, we also improved POS tagging and syntactic annotations.Because in Indonesian grammar there are some special structures, we also proposed how to adjust UDv2 annotation guidelines with those Indonesian grammar rules.To evaluate the quality of the new treebank, we built Indonesian dependency parser model using Parsito (UDPipe) parser.Using ten-fold cross-validation, the model that built using the revised treebank had UAS of 83.33% and LAS of 79.39%, over the original treebank with UAS of 73.32% and LAS of 65.98%.