Balanced Corpus Research Articles

Формальная модель оценки образа персонажа художественного произведения (на материале романа Дж. Оруэлла «1984»)

The aim of the research is to develop and test a formal model for evaluating the image of a character in a fictional work based on the corpus approach. The scientific originality lies in the fact that for the first time such an analysis of the character of a fictional work is carried out on the basis of the corpus approach. In the course of the work, a balanced corpus of the original text of the novel “1984” by G. Orwell was formed, which includes the marking of hierarchical relationships between the tokens. A number of search queries to the corpus were carried out using the author’s software “Balanced Linguistic Corpus Generator and Corpus Manager”. With the help of this tool, a frequency list of proper names from the novel was obtained and lists of tokens characterizing its main characters were output. The data obtained is extracted during the operation of a special algorithm built into the corpus manager, which selects from the text significant character traits and related lexical units. Thus, a model for evaluating the image of a character is built as a formal description of a system of interconnected hierarchical text elements. The results of the study were analyzed and compared with the results of mental interpretation. Thus, the characters’ profiles were compiled in terms of a formal description of their actions and those objects to which their actions were directed.

Indonesian disaster named entity recognition from multi source information using bidirectional LSTM (BiLSTM)

Precise logistic support is essential after a disaster occurs. It must be timely, accurate, targeted, and based on existing needs. However, obtaining sufficient and accurate information related to logistic distribution locations remains a key problem. Therefore, implementing Named Entity Recognition (NER) can address this issue. In recent years, news coverage through Indonesian digital news media and social media accounts has emerged as a promising source for building a disaster data corpus. This study implemented NER to extract and identify named entities from text-based information, particularly from Indonesian digital news media. In addition to using regular entities from the NER standard, this study introduced new entities specialized for disaster-related information, including DISASTER, SCALE, SUPPLIES, CASUALTIES, and OUTSIDE. The new disaster corpus in the Indonesian language for the NER model was obtained with an imbalanced dataset composition. To overcome this problem, random oversampling was applied. This study also utilized the BiLSTM model to recognize each entity in new textual information, evaluating its performance when the proposed Indonesian disaster corpus was used as a training reference in the deep learning model. Several optimization algorithms applied in BiLSTM were evaluated. The results showed improved BiLSTM performance using Adam optimization and a balanced corpus. Performance indicators achieved were 93.4 %, 82.4 %, and 87.5 % for precision, recall, and F1-score, respectively. The BiLSTM network captured long-range dependencies in sequential data provided by NER. Oversampling ensured that the proposed NER model could precisely recognize all entities and reduce biased results. Thus, the BiLSTM method can better identify entities in the textual corpus of Indonesian disaster-related online news.

Reaching the ceiling? Empirical scaling behaviour for deep EEG pathology classification

Machine learning techniques, particularly deep convolutional neural networks (ConvNets), are increasingly being used to automate clinical EEG analysis, with the potential to reduce the clinical burden and improve patient care. However, further research is required before they can be used in clinical settings, particularly regarding the impact of the number of training samples and model parameters on their testing error. To address this, we present a comprehensive study of the empirical scaling behaviour of ConvNets for EEG pathology classification. We analysed the testing error with increasing the training samples and model size for four different ConvNet architectures. The focus of our experiments is width scaling, and we have increased the number of parameters to up to 1.8 million. Our evaluation was based on two publicly available datasets: the Temple University Hospital (TUH) Abnormal EEG Corpus and the TUH Abnormal Expansion Balanced EEG Corpus, which together contain 10,707 training samples. The results show that the testing error follows a saturating power-law with both model and dataset size. This pattern is consistent across different datasets and ConvNet architectures. Furthermore, empirically observed accuracies saturate at 85%–87%, which may be due to an imperfect inter-rater agreement on the clinical labels. The empirical scaling behaviour of the test performance with dataset and model size has significant implications for deep EEG pathology classification research and practice. Our findings highlight the potential of deep ConvNets for high-performance EEG pathology classification, and the identified scaling relationships provide valuable recommendations for the advancement of automated EEG diagnostics.

Study on the Domain Adaption of Korean Speech Act using Daily Conversation Dataset and Petition Corpus

In Korean, quantitative speech act studies have usually been conducted on single utterances with unspecified sources. In this study, we annotate sentences from the National Institute of Korean Language's Messenger Corpus and the National Petition Corpus, as well as example sentences from an academic paper on contemporary Korean vlogging, and check the discrepancy between human annotation and model prediction. In particular, for sentences with differences in locutionary and illocutionary forces, we analyze the causes of errors to see if stylistic features used in a particular domain affect the correct inference of speech act. Through this, we see the necessity to build and analyze a balanced corpus in various text domains, taking into account cases with different usage roles, e.g., messenger conversations belonging to private conversations and petition corpus/vlogging script that have an unspecified audience.

Does simplification hold true for machine translations? A corpus-based analysis of lexical diversity in text varieties across genres

Extensive studies have described the linguistic features of human translations and verified the existence of the simplification translation universal. However, little has been known about the linguistic features of machine translations, although machine translation, as a unique modality of translation, has become an integral part of translation practice. This study is intended to test whether the simplification translation universal observed in human translations also holds true for machine translations. If so, are simplification features in machine translations different significantly from those in human translations? And does genre significantly affect simplification features? To this end, we built a balanced comparable corpus containing three text varieties, i.e., machine translations, human translations and target-language originals across three genres namely contemporary novels, government documents and academic abstracts. Based on the corpus, we conducted a systematic comparison of lexical diversity, as a proxy for simplification, of different text varieties. The results show that simplification is corroborated overall in both machine and human translations when compared with target-language originals, and machine translations are more simplified than human translations. Additionally, genre is found to exert a significant influence on the lexical diversity of different text varieties. This study is expected to expand the scope of corpus-based translation studies on the one hand and to offer insights into the improvement of machine translation systems on the other hand.

Генерация файлов формата Moodle XML на основе данных сбалансированного лингвистического корпуса

The aim of the proposed research is to develop an algorithm for generating educational (linguodidactic) content for popular learning management systems based on corpus data. The scientific novelty lies in the fact that for the first time, a linguodidactic component is added to the original author’s concept of the method for generating a balanced linguistic corpus, implemented on the general principles of maximum automation and universality of application. During the research, the structures of LMS Moodle backup files were studied, a software module was developed for converting data from the author’s balanced linguistic corpus with morphological markup into XML files, which were imported and tested for integrity and correctness in terms of solving the posed linguodidactic task. As a result, it was proved that the used linguistic corpus has a high linguodidactic potential, the created algorithm shows good results, in this case, for generating test or training tasks in German, and can be expanded to material from other languages, the support of which is provided by the software tool – the corpus generator.

Local grammar approaches to speech acts in Chinese: A case study of exemplification

This study develops a local grammar of exemplification in Chinese and also briefly discusses local grammar analyses of apologies in five selected languages, aiming to explore the feasibility of employing the local grammar approach to account for speech acts in various languages. Using data taken from DiSCUSS, a balanced and representative corpus of spoken Chinese, the analyses identified 13 local grammar patterns of exemplification, which shows the applicability of local grammars in describing speech act realisations in languages other than English. Our discussion further suggests that the resulting local grammar descriptions can not only reveal language-specific features of speech act realisations, but also facilitates contrastive (e.g., cross-linguistic and cross-cultural) speech act studies. The theoretical implications for describing speech act realisations and pedagogical applications for teaching speech acts in L2 contexts are also discussed. It is concluded that local grammar can be a useful and viable approach that can significantly take forward speech act studies.

B-NER: A Novel Bangla Named Entity Recognition Dataset With Largest Entities and Its Baseline Evaluation

Within the Natural Language Processing (NLP) framework, Named Entity Recognition (NER) is regarded as the basis for extracting key information to understand texts in any language. As Bangla is a highly inflectional, morphologically rich, and resource-scarce language, building a balanced NER corpus with large and diverse entities is a demanding task. However, previously developed Bangla NER systems are limited to recognizing only three familiar entities: person, location, and organization. To address this significant limitation, we introduce a novel Bangla NER dataset B-NER, which was created using 22,144 manually annotated Bangla sentences collected from Bangla newspapers and Bangla Wikipedia. This dataset includes a total of 9,895 unique words which were manually categorized into eight different entity types, such as a person, organization, event, artifact, time indicator, natural phenomenon, geopolitical entity, and geographical location. Inter-annotator agreement experiments were conducted to validate the quality of annotations performed by three annotators, resulting in a Kappa score of 0.82. In this paper, we provide an outline of the annotation guideline illustrated with examples, discuss the B-NER dataset properties, and present benchmark evaluations of the dataset. To establish that B-NER is more comprehensive and balanced in comparison to other publicly accessible datasets, we conducted cross-dataset modeling and validation, i.e. trained NER model on one dataset while tested on another, and found that the model trained on B-NER performed the best in that settings. Furthermore, we performed exhaustive benchmark evaluations based on Bidirectional LSTM with fastText embeddings and sentence transformer models. Among these models, fine-tuned IndicBERT achieved noticeable results with a Macro-F1 of 86%. This dataset and baseline results will be publicly available under a CC-BY 4.0 license in the CoNLL-2002 format to facilitate further research on Bangla NER.

An extended clinical EEG dataset with 15,300 automatically labelled recordings for pathology decoding

Automated clinical EEG analysis using machine learning (ML) methods is a growing EEG research area. Previous studies on binary EEG pathology decoding have mainly used the Temple University Hospital (TUH) Abnormal EEG Corpus (TUAB) which contains approximately 3,000 manually labelled EEG recordings. To evaluate and eventually even improve the generalisation performance of machine learning methods for EEG pathology, decoding larger, publicly available datasets is required. A number of studies addressed the automatic labelling of large open-source datasets as an approach to create new datasets for EEG pathology decoding, but little is known about the extent to which training on larger, automatically labelled dataset affects decoding performances of established deep neural networks. In this study, we automatically created additional pathology labels for the Temple University Hospital (TUH) EEG Corpus (TUEG) based on the medical reports using a rule-based text classifier. We generated a dataset of 15,300 newly labelled recordings, which we call the TUH Abnormal Expansion EEG Corpus (TUABEX), and which is five times larger than the TUAB. Since the TUABEX contains more pathological (75%) than non-pathological (25%) recordings, we then selected a balanced subset of 8,879 recordings, the TUH Abnormal Expansion Balanced EEG Corpus (TUABEXB). To investigate how training on a larger, automatically labelled dataset affects the decoding performance of deep neural networks, we applied four established deep convolutional neural networks (ConvNets) to the task of pathological versus non-pathological classification and compared the performance of each architecture after training on different datasets. The results show that training on the automatically labelled TUABEXB dataset rather than training on the manually labelled TUAB dataset increases accuracies on TUABEXB and even for TUAB itself for some architectures. We argue that automatically labelling of large open-source datasets can be used to efficiently utilise the massive amount of EEG data stored in clinical archives. We make the proposed TUABEXB available open source and thus offer a new dataset for EEG machine learning research.

Neural POS tagging of shahmukhi by using contextualized word representations

Part of Speech (POS) tagging has a preliminary role in building natural language processing applications. This paper presents the development and evaluation of the first POS tagged corpus along with a Bi-directional long-short memory (BiLSTM) network based POS tagger for Shahmukhi (Western Punjabi) at this scale. A balanced corpus of 0.13 million words has been annotated which contains text from 14 different text domains. A Shahmukhi POS tagset has been devised by studying the applicability of the CLE Urdu POS tagset and tagging guidelines have also been designed for annotation. A multi-step corpus evaluation process has been employed for tagged corpus including grammar-based and n-gram based consistency evaluations. The average inter-annotator agreement for all domains is 95.35% along with an average Kappa coefficient of 0.94. The performance of the BiLSTM POS tagger has been compared with the well-known language independent TreeTagger and the Stanford POS tagger. The accuracy of the tagger has been further improved by employing transfer learning by training context-free (Word2Vec) and contextualized (ELMo) word representations on a corpus of 14.9 Shahmukhi words which has been collected from World Wide Web. The tagger performed with an f-score of 96.11 and the accuracy of 96.12%. For a morphologically-rich and low-resourced language, these POS tagging results are quite promising.

Open Chinese Internet Sarcasm Corpus Construction: An Approach

Sarcasm is a commonly-used language phenomenon particularly on the Internet, which is often to convey criticism or negative emotions. A proper sarcasm corpus to help sarcasm study and detection can contribute to linguistic research and assist sentiment analysis, but an open Chinese corpus is found extremely lacking. In this paper, we referenced existing methods and data and constructed a balanced open Chinese Internet sarcasm corpus in a new approach to improve efficiency and data quality. The balanced open corpus contains multi-source and labeled 2,000 texts selected from bigger corresponding origin datasets. In our corpus, sarcasm and non-sarcasm, longer and shorter texts are both in 1:1 ratio.

Challenging stylometry: The authorship of the baroque play La Segunda Celestina

Abstract The aim of this study was to verify the possibility of Sor Juana Inés de la Cruz authoring the anonymous part of the baroque play La Segunda Celestina, commissioned to Agustín de Salazar, and left unfinished after his death. This is a first systematic stylometric study on this problem and a baroque hispanoamerican text. In our study, we faced building a balanced corpus from few available resources, and took extensive evaluation measures to deal with unclear stylometric signals. We use a variety of established attribution and verification methods, and introduce a novel evaluation procedure of examining historic texts with scarce corpora. The results support Sor Juana’s authorship, and unravel new connections between her and other authors of the time, showing, still undermined, powerful impact of her works on the epoch. The solutions adopted in solving methodological problems of such a complex task show how stylometry can overcome similar challenges.

AN EFFECTIVE HYBRID STOCHASTIC GRADIENT DESCENT ARABIC SENTIMENT ANALYSIS WITH PARTIAL-ORDER MICROWORDS AND PIECEWISE DIFFERENTIATION

Social media networking sites, such as Instagram, Facebook, and Twitter, have become an inextricable part of our everyday lives. These social media networks are useful for sharing news, images, and other information. The study of social media sentiment has recently received a lot of attention, especially in Arabic sentiment analysis. Social media sites are distinguished by unusual language that differs from the traditional format of the language. As a result, there is a necessity for efficient ways for analyzing the massive amount of new word variants that appear regularly in the digital world and online world. This study proposes a piecewise Stochastic Gradient Descent (SGD)-based model for sentiment classification. The TF-IDF-based term weighting scheme is employed for textual feature representation. For enhancing the model performance, stemming and partially ordered microword representation of tweets with varying look ahead distances is employed. Also, various n-gram models are considered for textual feature representation, which also improves the model performance. The proposed model is simulated and evaluated with the help of publicly available tweet corpus9 which is a balanced tweet corpus. The effectiveness of the proposed model is estimated using various performance evaluation metrics. According to the experimental observations, the proposed method accurately categorizes the testing and validation tweets set with an accuracy of up to 92.23%.

Экспериментальное моделирование базы данных сбалансированного лингвистического корпуса

The research aims to build a functioning experimental model of a relational database for operating with a balanced linguistic corpus of a fiction work. Scientific novelty lies in the fact that for the first time within the framework of a humanities study, a database of a linguistic corpus is being modeled with a thorough description and taking into account technical details and based on the provisions of the author’s concept of professionally oriented programming. The work involved three stages: forming a technical task (the structure of two tables of a relational database was developed, the SQLite format was selected, additional columns of the tables were provided for the subsequent expansion of the content of research), writing the source code for creating and filling the database (the Python programming language and the spaCy natural language processing module were used) and testing it based on the material of the texts of three F. Kafka’s novels “The Castle”, “Amerika” and “The Trial” (three functioning databases were created). The research findings have shown that modern natural language processing software tools allow one to create automatically full-fledged databases for processing SQL queries, which can be further expanded manually or automatically.

Design and Development of a Text-to-Speech Synthesizer for Afan Oromo

Speech is one of the natural ways of communication between humans, later extended as a means for human–computer interaction. It helps visually impaired people to read electronic texts and is used in information retrieval and language education. This paper proposed the development of a text-to-speech synthesizer for Afan Oromo (Oromo Language), using unit selection speech synthesizer approaches. Although several works have been conducted in the area of text-to-speech synthesis for technologically favored languages for many years, every language has its own unique features. So, speech synthesizer systems developed for one language cannot be used for another language, because the structures of one language are not presumably representative of others. It is clear that each program is based on the system corresponding to the phonetic rules of a certain language. Besides, the existing text-to-speech synthesizer for Afan Oromo was reviewed in this study and the result of developed prototype results are showing promising, however, still, their performance needs a lot of improvement in terms of intelligibility and naturalness using novel approaches and quality of corpus. Therefore, this research was initiated to develop the possibility of developing a prototype text-to-speech synthesizer to improve the performance of the text-to-speech synthesizer. In this study, Afan Oromo corpus was collected from genuine sources and prepared speech datasets both text and audio in collaboration with Afan Oromo experts. The performance of the synthesizer was tested by proper users for its intelligibility and naturalness using Mean Opinion Scale (MOS). The obtained result of naturalness of the prototype is 4.44 (very good) out of 5, which indicated that the result obtained is encouraging and better performance than the existing TTS of Afan Oromo in terms of intelligibility and naturalness. But the result scored in terms of intelligibility still needs further work. The main challenge is Afan Oromo has many dialects, so preparing a balanced text corpus from each dialect is very tough. Moreover, enhancement of the work is predicted to bring a reasonable level of intelligibility to the system.

Corpus-Based Investigation of S-V Concord Patterns of Nouns with Latin Plural Endings

English subject-verb concord can pose problems to learners such as omission of third person singular inflection and overextension due to linguistic typology, phonological and perceptual factors, syntax, and the number of nouns acting as the subject of the sentence, among others. A group of nouns that invites confusion is what Huddleston and Pullum (2002) call nouns with Latin plural endings whose singular forms are uncommon compared to their plural counterparts. This investigation aims to provide teachers with guidance on the general and genre-based agreement patterns of these nouns using the balanced Corpus of Contemporary American English (COCA). This study found that based on their agreement patterns these nouns can be classified into three groups: 1) nouns whose tendencies are line with prescriptive rules; 2) noun whose tendencies are ambivalent; and 3) nouns which might have already been “reclassified” by speakers. By genres, there was a strong preference to use “plural” verbs with these nouns in the written and the academic components of the corpus while “singular” verbs are preferred in the spoken components. The implications of the study for teaching and research were also provided.

Partitive genitive constructions and agreement variations in Latvian

Abstract The article takes a closer look at the partitive genitive constructions in Latvian, their structure, semantics, and syntactic functions. With the help of a corpus analysis, an attempt has been made to find out what determines gender and number agreement variations between the partitive genitive constructions in subject position and the predicate, if it comprises a declinable participle. Attention was paid to such features as subject animacy, voice of the predicate, word order (SV / VS), quantifier lexeme and grammatical number of the genitive. 320 examples with partitive genitive constructions were selected from The Balanced Corpus of Modern Latvian LVK2018, including such quantifier lexemes as daļa ‘part’, skaits ‘number’, vairums ‘quantity’, vairākums ‘majority’, puse ‘half’, daudzums ‘amount / quantity’, lērums ‘bagful’, simts ‘hundred’, tūkstotis ‘thousand’, daudz ‘much/many’, maz ‘little / few’. The best results are given by the combination of values ‘animate’, ‘active’ and ‘SV’, in which 93.24% of the examples are partitive genitive agreement. Data analyzed so far suggest that quantifiers involving the semantic element of partitivity (e.g., daļa ‘part’) favor partitive genitive agreement, whereas quantifiers lacking this semantic element favor quantifier agreement. This is an exploratory quantitative study on agreement tendencies in Latvian.

Arabic Fake News Detection: A Fact Checking Based Deep Learning Approach

Fake news stories can polarize society, particularly during political events. They undermine confidence in the media in general. Current NLP systems are still lacking the ability to properly interpret and classify Arabic fake news. Given the high stakes involved, determining truth in social media has recently become an emerging research that is attracting tremendous attention. Our literature review indicates that applying the state-of-the-art approaches on news content address some challenges in detecting fake news’ characteristics, which needs auxiliary information to make a clear determination. Moreover, the ‘Social-context-based’ and ‘propagation-based’ approaches can be either an alternative or complementary strategy to content-based approaches. The main goal of our research is to develop a model capable of automatically detecting truth given an Arabic news or claim. In particular, we propose a deep neural network approach that can classify fake and real news claims by exploiting ‘Convolutional Neuron Networks’. Our approach attempts to solve the problem from the fact checking perspective, where the fact-checking task involves predicting whether a given news text claim is factually authentic or fake. We opt to use an Arabic balanced corpus to build our model because it unifies stance detection, stance rationale, relevant document retrieval and fact-checking. The model is trained on different well selected attributes. An extensive evaluation has been conducted to demonstrate the ability of the fact-checking task in detecting the Arabic fake news. Our model outperforms the performance of the state-of-the-art approaches when applied to the same Arabic dataset with the highest accuracy of 91%.

Developing of a Balanced Corpus for a Fiction Work (on the Material of F. Kafka’s Novels)

Developing of a Balanced Corpus for a Fiction Work (on the Material of F. Kafka’s Novels)

Bangladeshi Bangla speech corpus for automatic speech recognition research

This article reports the development of language resource for Bangladeshi Bangla spoken language (BBSL). Bangladeshi Bangla has inadequate large speech corpora for Large Vocabulary Continuous Speech Recognition (LVCSR) system. The accuracy of the automatic speech recognition (ASR) system rests on the quality of the speech corpus. This work discusses the common issues and activities related to the development of a large speech corpus named ▪ (SUBAK.KO). This corpus is designed to support ASR research in Bangladeshi Bangla. It has been labeled sentence-wise. We have trained this corpus with one of the well-known current End-to-End ASR algorithms, Recurrent Neural Networks (RNNs) with Connectionist Temporal Classification (CTC). To know the strengths and weaknesses, the CER (Character Error Rate) and the WER (Word Error Rate) of the trained RNNCTC model have been observed. Another open-source large Bangla ASR corpus has been trained using the same ASR algorithm. Both trained models have been compared to assess the quality of these corpora. It has been found that SUBAK.KO is a more balanced corpus and considered more regional accented speech variability for a LVCSR system compared to that open-source large Bangla ASR corpus.