Human Language Technology Research Articles

New haptic systems for elicit emotions in audio-visual events for hearing impaired people.

In this research we have developed devices for hearing impaired people that can transmit the emotions evoked by music. In this way, we will all achieve the full inclusion of impairment people in all social activities. As this is an issue that goes beyond pure engineering research, it has been necessary to create a multidisciplinary group within the Spanish Centre for Captioning and Audio Description (CESyA), in direct collaboration with the Display and Photonic Applications Research Group (GDAF) and the Human Language and Accessibility Technologies Group (HULAT) of the Carlos III University (UC3M) and the Psychiatry and Applied Communication Departments of the Complutense University of Madrid. For this research, a pair of haptic gloves were developed, which transmit different characteristics extracted from music by means of vibrational sequences, intended to drive an emotional reinforcement. Another device, a “music box”, a vibration box, was also developed from the idea of using vibration as an instrument.Overall, the acceptance of the glove was positive, with an 80% of users accepting the negative video, repeatedly stating in the surveys that they "felt more when the glove vibrated". The music box was also well received. Volunteers felt that the system gave them a pleasant sensation and that they could feel the music. However, some participants associated the vibration with the sensation of an alarm clock, which was expectedly negative.Considering the obtained EEG results, we can conclude that the integration of simple multimodal stimuli, visual and vibrotactile, helps to enhance the activation of emotional processes linked with musical stimulation. It can also be concluded that multimodal visual and vibrotactile stimulation enhances both attentional and emotional processes. These results pave the way for future developments that will allow a hearing-impaired person to fully enjoy an audiovisual event, which will have a great impact on this group.

Translating Ulyssei@s: Work in progress between general and specialised translation in light of NMT

It is common knowledge that the publication of translation projects carried out by students is a very motivating factor for training. Since the second semester of 2015-2016, undergraduate translation students of the Faculty of Arts and Humanities of the University of Porto have participated in the development of the French and English versions of the digital encyclopaedia Ulyssei@s, created by the Institute for Comparative Literature Margarida Losa (ILCML) in the early 2010s. Besides motivating students, this translation project published on the Internet allows us to develop translation skills, foster teamwork, consolidate the three-phase methodology proposed by Gouadec (2002, 2007), and experiment with several essential tasks in language services: collaborative glossary-making, translation, revision of human translation and post-editing neural machine translation (NMT). Some of the challenges posed by the translation of Ulyssei@s are common to all types of translation, while others are specific to the translation of entries of this sui generis encyclopaedia, such as a variety of textual genres (biography, literary criticism, prose, poetry, theatre, essay) in hybrid texts and the need to find the original or the official translation of the works cited. This paper focuses on the data collected in the spring semester of 2022 and is based on the results of a survey submitted to students at the end of the project. It aims to share the experience of developing an inevitably unfinished project and adapting training and learning to the latest advances in Human Language Technology, since NMT has become a valuable tool for human translators, and Ulyssei@s is a digital encyclopaedia regularly enriched by ILCML collaborators with new entries.

Lexical Creativity in Community Translation Corpus from Ghana

Although corpus-based research is becoming more common in many languages and in translation studies in general, it tends to be scarce or even nonexistent in Akan and many other African languages that are often considered as minority languages. In fact, failure to use Akan in various fields has hindered its development to some extent, particularly in online electronic resources, human language technology and for optimising online presence. Thelimited availability of resources, such as reference corpora and specialised dictionaries, suggests that translators who work from and into Akan often have to rely solely on their ingenuity in the process of translation. The study, therefore, explored a bidirectional English–Akan parallel corpus, built earlier by the researcher in another study. Using the descriptive translation studies approach, Skopos theory and Baker’s (1992) translation strategies, the paperanalyses lexical creativity as a strategy used to resolve translation problems specific to the language pair: borrowing, compounding and reduplication. Akan may not be a typical minority language, but it is in a disadvantaged position compared to English, which serves as the official language. Consequently, many documents are primarily written in the official language, thereby creating a form of imbalance in terms of language development and translationchallenges for community translators and interpreters who work from and into Akan. The research sheds light on how the abovementioned corpus can empower community translation in Ghana. It also identifies lexical creativity as an effective strategy to stimulate the need for the development of such a community language and to curb the imbalance in relation to English regarding terminology, text-type development, reference resources and translation tools. As translation tools such as Linguee rely on parallel corpora, the bidirectional English–Akan parallel corpus may be uploaded online or enlarged to develop general and professional language tools, and can serve as a basis for other enquiries.

Conversational AI: Dialogue Systems, Conversational Agents, and Chatbots by Michael McTear

Conversational AI: Dialogue Systems, Conversational Agents, and Chatbots by Michael McTear

Creating electronic resources for African languages through digitisation: a technical report

The need for electronic resources for (under-resourced) African languages is an often stated one. These resources are needed for language research in general, and more specifically for the development of Human Language Technology (HLT) applications such as machine translation, speech recognition, electronic dictionaries, spelling and grammar checkers, and optical character recognition. These technologies rely on large quantities of high-quality electronic data. Digitisation is one of the strategies that can be used to collect such data. For the purpose of this paper, digitisation is understood as the conversion of analogue text, audio and video data into digital form, as well as the provision of born digital data that is currently not available in a format that enables downstream processing. There is a general perception that the African languages are under-resourced with regard to sufficient digitisation tools to function effectively in the modern digital world. Our paper is presented as a technical report, detailing the tools, procedures, best practices and standards that are utilised by the UP digitisation node to digitise text, audio and audio-visual material for the African languages. The digitisation effort is part of the South African Digital Languages Resources (SADiLaR) project (https://www.sadilar.org/index.php/en/), funded by the Department of Science and Innovation. Our report is based on a best practices document, developed through the course of our digitisation project and forms part of the deliverables as per contractual agreement between the UP digitisation node and the SADiLaR Hub. The workflow as explained in this document was designed with this specific project in mind; software and hardware utilised were also selected based on the constraints with regard to capacity and available technical skills in mind. We motivate our choice of Optical Character Recognition (OCR) software by referring to an earlier experiment in which we evaluated three commercially available OCR programmes. We did not attempt a full-scale evaluation of all available OCR software, but rather focused on selecting one that renders high quality outputs. We also reflect on one of the challenges specific to our project, i.e. copyright clearance. This is particularly relevant with regard to published material. In the absence of newspapers for specifically the African Languages (isiZulu being a notable exception), the biggest portion of textual material available for digitisation consists of printed material such as textbooks, novels, dramas, short stories and other literary genres. The digitisation process is driven by the availability of material for the different languages. Furthermore, obtaining copyright clearance from publishers is a prerequisite for digitisation and especially for the release of any digitised text data for further use and/or processing. Having information on a relatively small-scale digitisation workflow and best practices readily available will enable other interested parties to participate in the digitisation effort, thus contributing to the collection of electronic data for the African languages.

The Analysis of a GPT-based Sepedi Text Generation Model

Text generation is defined as a component of natural language processing that makes use of computational linguistics techniques to produce text that cannot be distinguished from human written text. This study aims to develop and analyse a Generative Pre-Trained Transformer 2 (GPT-2) language model to generate Sepedi phrases. The under-resourced Sepedi language is regarded as a disjunctive language. The Sepedi language orthographic representation presents challenges and has limited resources. The GPT-2 transformer requires large datasets, as well as state-of-the-art computational resources. The unstructured National Centre for Human Language Technology (NCHLT) Sepedi text dataset was used. The text generation model developed with the small dataset managed to get the lowest loss value of 2.36. The output text generated using this model produces a text that is syntactically correct with instances of grammatical errors. The model performed better than previously developed Sepedi text generation models by using transformer-based technique.

AmericasNLI: Machine translation and natural language inference systems for Indigenous languages of the Americas.

Little attention has been paid to the development of human language technology for truly low-resource languages-i.e., languages with limited amounts of digitally available text data, such as Indigenous languages. However, it has been shown that pretrained multilingual models are able to perform crosslingual transfer in a zero-shot setting even for low-resource languages which are unseen during pretraining. Yet, prior work evaluating performance on unseen languages has largely been limited to shallow token-level tasks. It remains unclear if zero-shot learning of deeper semantic tasks is possible for unseen languages. To explore this question, we present AmericasNLI, a natural language inference dataset covering 10 Indigenous languages of the Americas. We conduct experiments with pretrained models, exploring zero-shot learning in combination with model adaptation. Furthermore, as AmericasNLI is a multiway parallel dataset, we use it to benchmark the performance of different machine translation models for those languages. Finally, using a standard transformer model, we explore translation-based approaches for natural language inference. We find that the zero-shot performance of pretrained models without adaptation is poor for all languages in AmericasNLI, but model adaptation via continued pretraining results in improvements. All machine translation models are rather weak, but, surprisingly, translation-based approaches to natural language inference outperform all other models on that task.

Digital Transformation of Legal Services and Access to Justice: Challenges and Possibilities

Abstract The pandemic affected the access to justice situation in terms of the never rapid shift to digitalisation of legal services, and in this article, we evaluate whether artificial intelligence (AI) and its state-of-the-art technologies like machine learning and human language technologies have the potential to improve access to legal services. For this purpose, we not only examine and identify problematic areas, but also share the empirical data and insights of the practical application of AI technologies, especially human language technologies. In the first part of the article, we explore how the internet has created the foundations for a new paradigm of society including institution law. The second part of the article is devoted for analysis of challenges for access to justice in post pandemic world. In the third part, we elaborate on questions about technical feasibility, legal and moral acceptability of the digitalisation of legal services. Then follows the case analysis of the practical application of human language technologies in legal domain.

Multilingualism, translanguaging and transknowledging

AbstractKey findings, analysis and recommendations that have emerged from a research project, ‘Using Human Language Technology to enhance academic integrity, inclusivity, knowledge exchange, student diversity and retention’ at the University of South Australia conducted in 2019 are discussed in this article. The primary purpose of the project was to address some of the challenges and opportunities afforded by increasing student and teacher diversity at a predominantly English-medium Australian university through newly enhanced human language translation technology (HLT) also known as machine translation (MT). This technology is frequently used for the translation of human language, and it falls under the umbrella of Artificial Intelligence (AI) technologies. From the institution’s perspective, key aims of the project were to contribute to the university’s Digital Learning Strategy priorities and core values embedded in a structural transformation of the university. These include integrity, accountability, diversity, social justice, engagement and collaboration. The researchers’ objectives focussed on multilingual pedagogies using HLT to support knowledge exchange (transknowledging), and translanguaging for all students. These disrupt inequitable hierarchies, and position bi-/multilingual students as valuable resources for monolingual staff and students.

Special Issue on Machine Learning and Natural Language Processing

The task of processing natural language automatically has been on the radar of researchers since the dawn of computing, fostering the rise of fields such as computational linguistics and human–language technologies [...]

ScAN: Suicide Attempt and Ideation Events Dataset.

Suicide is an important public health concern and one of the leading causes of death worldwide. Suicidal behaviors, including suicide attempts (SA) and suicide ideations (SI), are leading risk factors for death by suicide. Information related to patients' previous and current SA and SI are frequently documented in the electronic health record (EHR) notes. Accurate detection of such documentation may help improve surveillance and predictions of patients' suicidal behaviors and alert medical professionals for suicide prevention efforts. In this study, we first built Suicide Attempt and Ideation Events (ScAN) dataset, a subset of the publicly available MIMIC III dataset spanning over 12k+ EHR notes with 19k+ annotated SA and SI events information. The annotations also contain attributes such as method of suicide attempt. We also provide a strong baseline model ScANER (Suicide Attempt and Ideation Events Retreiver), a multi-task RoBERTa-based model with a retrieval module to extract all the relevant suicidal behavioral evidences from EHR notes of an hospital-stay and, and a prediction module to identify the type of suicidal behavior (SA and SI) concluded during the patient's stay at the hospital. ScANER achieved a macro-weighted F1-score of 0.83 for identifying suicidal behavioral evidences and a macro F1-score of 0.78 and 0.60 for classification of SA and SI for the patient's hospital-stay, respectively. ScAN and ScANER are publicly available.

The South African National Lexicography Units — Two Decades Later

The lexicography practice in South Africa has distinctive features and to a great extent relates to the political dispensation current at a given period. During the previous political dispensation only two dictionary units existed in South Africa that were state funded, namely the Bureau of the Woordeboek van die Afrikaanse Taal (WAT) and the Dictionary for South African English (DSAE). These dictionary units, therefore have a long dictionary history. Some of the African languages, such as Xhosa, Zulu, Tswana and Northern Sotho also had dictionary units at the time but they were situated at and funded by tertiary institutions. The current national government requires dictionaries in all official languages for proper communication in languages the citizens understand best, i.e. their respective mother tongues. The Pan-South African Language Board (PanSALB), established in 1996, was a direct consequence of the country's new multilingual dispensation. This dispensation required equal rights for all the official languages and the legislation governing PanSALB was therefore amended to allow for equal justice to all dictionary projects for the official South African languages. This led to the establishment of national lexicography units for each of the official South African languages. A strategic planning process conducted at the Bureau of the Woordeboek van die Afrikaanse Taal not only influenced the lexicography practice at the Bureau, but also had a huge impact on the newly established national lexicography units (NLUs). Both the articles published in Lexikos, that originated in 1991, and the activities of the African Association for Lexicography (AFRILEX), established in 1995, have a huge influence on the activities of the NLUs. Although the NLUs are already 21 years old and should officially be of age, the question remains whether they really gained their independence. This article focusses on the heritage of the Bureau of the WAT as National Lexicography Unit for Afrikaans and the example that it has set to the other NLUs. Keywords: academic journal, association, communication, dictionary, feasibility study, human language technologies, language for special purposes, legislation, lexicography, management, national lexicography units, strategic planning, terminography, terminology

NLP technologies for analysing user generated Twitter data to identify the reputation of universities in the Valencian Community, Spain

The measurement of a university's reputation is currently based on rankings published nationally and internationally. These rankings are based on different criteria directly related to research and teaching, thereby creating a reputation. However, the proliferation of digital media has made it is possible to know the real opinion of stakeholders on a range of issues, including brand perception. Therefore, analysing datasets from social networks, blogs and the comment sections of websites has become an increasingly worthwhile task to getting a thorough understanding of an organisation's reputation. One way to approach this task is to use natural language processing (NLP) technologies as part of the process of conceiving the social reputation of institutional brands by interpreting a large amount of comments from social users. NLP technologies are useful for identifying and quantifying positive, neutral, negative and other information posted on social networks, i.e., Twitter. Through this research, we intend to monitor the situation of universities in the Valencian Community in terms of audience opinions, based on a justified selection of human language technologies that provide the necessary data from tweets, to create different classifications such as reputation, audience, among others.

NLP technologies for analysing user generated Twitter data to identify the reputation of universities in the Valencian Community, Spain

The measurement of a university's reputation is currently based on rankings published nationally and internationally. These rankings are based on different criteria directly related to research and teaching, thereby creating a reputation. However, the proliferation of digital media has made it is possible to know the real opinion of stakeholders on a range of issues, including brand perception. Therefore, analysing datasets from social networks, blogs and the comment sections of websites has become an increasingly worthwhile task to getting a thorough understanding of an organisation's reputation. One way to approach this task is to use natural language processing (NLP) technologies as part of the process of conceiving the social reputation of institutional brands by interpreting a large amount of comments from social users. NLP technologies are useful for identifying and quantifying positive, neutral, negative and other information posted on social networks, i.e., Twitter. Through this research, we intend to monitor the situation of universities in the Valencian Community in terms of audience opinions, based on a justified selection of human language technologies that provide the necessary data from tweets, to create different classifications such as reputation, audience, among others.

High quality recordings and transcriptions of speech via remote platforms

High quality recordings and transcriptions of speech are important to a wide variety of disciplines from linguistics and human language technology to biomedical screening, diagnosis and tracking. The wide availability of internet connections and powerful mobile devices offers low cost opportunities for collecting speech data at scale. But even in the era of open source software, numerous challenges remain. For example, video call applications such as Zoom have become widespread and allow for recording, but only provide lossy codecs with low frame rates, and often contain missing, repeated, or interpolated frames, as well as freezes, longer dropouts and other audio artefacts. LDC has recently developed a suite of tools to allow high quality internet-based audio recordings and transcription, with a premium on portability and flexibility, using secure cloud computing services for storage and back-end processing. We present here the current design and capabilities of our software, as well as availability in terms of open source code and app distribution. We also discuss future plans; planned capabilities include 2 + sided conversational recordings, connecting participants via the internet, a modern extension of Conversational Telephone Speech (CTS) collections.

Linguistic Resources for Bhojpuri, Magahi, and Maithili: Statistics about Them, Their Similarity Estimates, and Baselines for Three Applications

Corpus preparation for low-resource languages and for development of human language technology to analyze or computationally process them is a laborious task, primarily due to the unavailability of expert linguists who are native speakers of these languages and also due to the time and resources required. Bhojpuri, Magahi, and Maithili, languages of the Purvanchal region of India (in the north-eastern parts), are low-resource languages belonging to the Indo-Aryan (or Indic) family. They are closely related to Hindi, which is a relatively high-resource language, which is why we compare them with Hindi. We collected corpora for these three languages from various sources and cleaned them to the extent possible, without changing the data in them. The text belongs to different domains and genres. We calculated some basic statistical measures for these corpora at character, word, syllable, and morpheme levels. These corpora were also annotated with parts-of-speech (POS) and chunk tags. The basic statistical measures were both absolute and relative and were expected to indicate linguistic properties, such as morphological, lexical, phonological, and syntactic complexities (or richness). The results were compared with a standard Hindi corpus. For most of the measures, we tried to match the corpus size across the languages to avoid the effect of corpus size, but in some cases it turned out that using the full corpus was better, even if sizes were very different. Although the results are not very clear, we tried to draw some conclusions about the languages and the corpora. For POS tagging and chunking, the BIS tagset was used to manually annotate the data. The POS-tagged data sizes are 16,067, 14,669, and 12,310 sentences, respectively, for Bhojpuri, Magahi, and Maithili. The sizes for chunking are 9,695 and 1,954 sentences for Bhojpuri and Maithili, respectively. The inter-annotator agreement for these annotations, using Cohen’s Kappa, was 0.92, 0.64, and 0.74, respectively, for the three languages. These (annotated) corpora have been used for developing preliminary automated tools, which include POS tagger, Chunker, and Language Identifier. We have also developed the Bilingual dictionary (Purvanchal languages to Hindi) and a Synset (that can be integrated later in the Indo-WordNet) as additional resources. The main contribution of the work is the creation of basic resources for facilitating further language processing research for these languages, providing some quantitative measures about them and their similarities among themselves and with Hindi. For similarities, we use a somewhat novel measure of language similarity based on an n-gram-based language identification algorithm. An additional contribution is providing baselines for three basic NLP applications (POS tagging, chunking, and language identification) for these closely related languages.

Quantitative analysis of Sesotho sa Leboa part-of-speech taggers

This study reported in this article conducts a quantitative analysis of the Sesotho sa Leboa National Centre for Human Language Technology (NCHLT) part-of-speech annotated data set, and compares the quality of the NCHLT and CTexT part-of-speech taggers based on the data set. The two taggers were developed as part of the NCHLT Text project and are both based on Taljard et al.’s fine-grained tagset, aligned to the morphological structure of Sesotho sa Leboa. A gold standard data set of 7 153 tokens is utilised for comparison and evaluation of the overall performance of the two part-of-speech taggers and to perform fine-grained error analysis. We find that the NCHLT and CTexT taggers obtain 88.40% and 94.18% accuracy respectively and describe the linguistic nature of the most frequent errors observed in the two taggers.

NHSS: A speech and singing parallel database

We present a database of parallel recordings of speech and singing, collected and released by the Human Language Technology (HLT) laboratory at the National University of Singapore (NUS), that is called NUS-HLT Speak–Sing (NHSS) database. We release this database11https://hltnus.github.io/NHSSDatabase/. to the public to support research activities, that include, but not limited to comparative studies of acoustic attributes of speech and singing signals, cooperative synthesis of speech and singing voices, and speech-to-singing conversion. This database consists of recordings of sung vocals of English pop songs, the spoken counterpart of lyrics of the songs read by the singers in their natural reading manner, and manually prepared utterance-level and word-level annotations. The audio recordings in the NHSS database correspond to 100 songs sung and spoken by 10 singers, resulting in a total of 7 h of audio data. There are 5 male and 5 female singers, singing and reading the lyrics of 10 songs each. In this paper, we discuss the design methodology of the database, analyze the similarities and dissimilarities in characteristics of speech and singing voices, and provide some strategies to address relationships between these characteristics for converting one to another. We develop benchmark systems, which can be used as reference for speech-to-singing alignment, spectral mapping, and conversion using the NHSS database.

A Deep Learning Model of Multiple Knowledge Sources Integration for Community Question Answering

The community Question Answering (cQA) problem requires the task that given a question it aims at selecting the most related question-answer tuples (a question and its answers) from the stored question-answer tuples data set. The core mission of this task is to measure the similarity (or relationship) between an input question and questions from the given question-answer data set. Under our observation, there are either various information sources as well as di erent measurement models which can provide complementary knowledge for indicating the relationship between questions and question-answer tuples. In this paper we address the problem of modeling and combining multiple knowledge sources for determining and ranking the most related question-answer tuples given an input question for cQA problem. Our proposed model will generate di erent features based on di erent representations of the data as well as on di erent methods and then integrate this information into the BERT model for similarity measurement in cQA problem. We evaluate our proposed model on the SemEval 2016 data set and achieve the state-of-the-art result.
 Keywords
 Community question answering, Multi knowledge sources, Deep learning, The BERT model
 References
 [1] C. Alberto, D. Bonadiman, G. D. S. Martino, Answer and Question Selection for Question Answering on Arabic and English Fora, in Proceedings of SemEval-2016, 2016, pp. 896-903.
 [2] Filice, D. Croce, A. Moschitti, R. Basili, Learning Semantic Relations between Questions and Answers, in Proceedings of SemEval-2016, 2016, pp. 1116-1123.
 [3] Wang, Z. Ming, T. S. Chua, A Syntactic Tree Matching Approach to Finding Similar Questions in Community-based qa Services, in SIGIR, 2009, pp. 187-194.
 [4] Pengfei, Q. Xipeng, C. Jifan, H. Xuanjing, Deep Fusion lstms for Text Semantic Matching, in Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics, Vol. 1, 2016, pp. 1034-1043,
 https://doi.org/ 10.18653/v1/P16-1098.
 [5] Jonas, T. Aditya, Siamese Recurrent Architectures for Learning Sentence Similarity, in Proceedings of the Thirtieth AAAI Conference on Artificial Intelligence (AAAI-16), 2016, pp. 2786-2792.
 [6] Jacob, C. M. Wei, L. Kenton, T. Kristina, Bert: Pre-training of Deep Bidirectional Transformers for Language Understanding, in Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, 2019, pp. 4171-4186.
 [7] Wissam, B. Fady, H. Hazem, Arabert: Transformer-based Model for Arabic Language Understanding, in Proceedings of the 4th Workshop on Open-Source Arabic Corpora and Processing Tools, with a Shared Task on O ensive Language Detection, 2020, pp. 9-15.
 [8] Lukovnikov, A. Fischer, J. Lehmann, Pretrained Transformers for Simple Question Answering Over Knowledge Graphs, ArXiv, abs/2001.11985, 2019.
 [9] V. Aken, B. Winter, A. Loser, F. Gers, How Does BERT Answer Questions?: A Layer-Wise Analysis of Transformer Representations, in Proceedings of the 28th ACM International Conference on Information and Knowledge Management, 2019.
 [10] Chan, Y. Fan, A Recurrent BERT-based Model for Question Generation, in Proceedings of the Second Workshop on Machine Reading for Question Answering, 2019, pp. 154-162.
 [11] Lan, M. Chen, S. Goodman, K. Gimpel, P. Sharma, R. Soricut, ALBERT: A Lite BERT for Self-supervised Learning of Language Representations, ArXiv, abs/1909.11942, 2020.
 [12] Ngai, Y. Park, J. Chen, M. Parsapoor, Transformer-Based Models for Question Answering on COVID19, ArXiv, abs/2101.11432, 2021.
 [13] Yu, L. Wu, Y. Deng, R. Mahindru, Q. Zeng, S. Guven, M. Jiang, A Technical Question Answering System with Transfer Learning, in Proceedings of the 2020 EMNLP (Systems Demonstrations), 2020, pp. 92-99.
 [14] S. McCarley, R. Chakravarti, A. Sil, Structured Pruning of a BERT-based Question Answering Model, arXiv: Computation and Language, 2019.
 [15] Almiman, N. Osman, M. Torki, Deep Neural Network Approach for Arabic Community Question Answering, Alexandria Engineering Journal, Vol. 59, 2020, pp. 4427-4434.
 [16] Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, I. Polosukhin, Attention is all you Need, in Advances in Neural Information Processing Systems, 2017, pp. 5998-6008.
 [17] T. Nguyen, A. C. Le, H. N. Nguyen, A Model of Convolutional Neural Network Combined with External Knowledge to Measure the Question Similarity for Community Question Answering Systems, International Journal of Machine Learning and Computing, Vol. 11, No. 3, 2021, pp. 194-201, https://doi.org/ 10.18178/ijmlc.2021.11.3.1035.

Manfred Stede and Jodi Schneider: Argumentation mining. Synthesis lectures on human language technologies, edited by Graeme Hirst

Manfred Stede and Jodi Schneider: Argumentation mining. Synthesis lectures on human language technologies, edited by Graeme Hirst