Monolingual Task Research Articles

Filtering offensive language from multilingual social media contents: A deep learning approach

In the face of uncontrolled offensive content on social media, automated detection emerges as a critical need. This paper tackles this challenge by proposing a novel approach for identifying offensive language in multilingual, code-mixed, and script-mixed settings. The study presents a novel multilingual hybrid dataset constructed by merging diverse monolingual and bilingual resources. Further, we systematically evaluate the impact of input representations (Word2Vec, Global Vectors for Word Representation (or GloVe), Bidirectional Encoder Representations from Transformers (or BERT), and uniform initialization) and deep learning models (Convolutional Neural Network (or CNN), Bidirectional Long Short Term Memory (or Bi-LSTM), Bi-LSTM-Attention, and fine-tuned BERT) on detection accuracy. Our comprehensive experiments on a dataset of 42,560 social media comments from five languages (English, Hindi, German, Tamil, and Malayalam) reveal the superiority of fine-tuned BERT. Notably, it achieves a macro average F1-score of 0.79 for monolingual tasks and an impressive 0.86 for code-mixed and script-mixed tasks. These findings significantly advance offensive language detection methodologies and shed light on the complex dynamics of multilingual social media, paving the way for more inclusive and safer online communities.

MIRACL: A Multilingual Retrieval Dataset Covering 18 Diverse Languages

Abstract MIRACL is a multilingual dataset for ad hoc retrieval across 18 languages that collectively encompass over three billion native speakers around the world. This resource is designed to support monolingual retrieval tasks, where the queries and the corpora are in the same language. In total, we have gathered over 726k high-quality relevance judgments for 78k queries over Wikipedia in these languages, where all annotations have been performed by native speakers hired by our team. MIRACL covers languages that are both typologically close as well as distant from 10 language families and 13 sub-families, associated with varying amounts of publicly available resources. Extensive automatic heuristic verification and manual assessments were performed during the annotation process to control data quality. In total, MIRACL represents an investment of around five person-years of human annotator effort. Our goal is to spur research on improving retrieval across a continuum of languages, thus enhancing information access capabilities for diverse populations around the world, particularly those that have traditionally been underserved. MIRACL is available at http://miracl.ai/.

A Multitask Cross-Lingual Summary Method Based on ABO Mechanism

Recent cross-lingual summarization research has pursued the use of a unified end-to-end model which has demonstrated a certain level of improvement in performance and effectiveness, but this approach stitches together multiple tasks and makes the computation more complex. Less work has focused on alignment relationships across languages, which has led to persistent problems of summary misordering and loss of key information. For this reason, we first simplify the multitasking by converting the translation task into an equal proportion of cross-lingual summary tasks so that the model can perform only cross-lingual summary tasks when generating cross-lingual summaries. In addition, we splice monolingual and cross-lingual summary sequences as an input so that the model can fully learn the core content of the corpus. Then, we propose a reinforced regularization method based on the model to improve its robustness, and build a targeted ABO mechanism to enhance the semantic relationship alignment and key information retention of the cross-lingual summaries. Ablation experiments are conducted on three datasets of different orders of magnitude to demonstrate the effective enhancement of the model by the optimization approach; they outperform the mainstream approaches on the cross-lingual summarization task and the monolingual summarization task for the full dataset. Finally, we validate the model’s capabilities on a cross-lingual summary dataset of professional domains, and the results demonstrate its superior performance and ability to improve cross-lingual sequencing.

Нейромережеві підходи для задач письмового асистента

The article is devoted to the analysis of tasks for building a writing assistant, one of the most prominent fields of natural language processing and artificial intelligence in general. Specifically, we explore monolingual local sequence transduction tasks: grammatical and spelling errors correction, text simplification, paraphrase generation. To give a better understanding of the considered tasks, we show examples of expected rewrites. Then we take a deep look at such key aspects as existing publicly available datasets and their training splits, quality metrics for high quality evaluation, and modern solutions based primarily on neural networks. For each task, we analyze its main peculiarities and how they influence the state-of-the-art models. Eventually, we investigate the most eloquent shared features for the whole group of tasks in general and for approaches that provide solutions to them.

Boosting Bilingual Metalinguistic Awareness Under Dual Language Activation: Some Implications for Bilingual Education

AbstractMost studies on bilingual children's metalinguistic awareness assess metalinguistic awareness using monolingual tasks. This may not reflect how a bilingual's languages dynamically interact with each other in creating metalinguistic representations. We tested 33 Greek–Italian bilingual children (8–11 years) for metalinguistic awareness using acceptability‐rating tasks in which they had to judge and explain grammatical errors. The tasks were in monolingual and bilingual modes in order to show how far metalinguistic awareness in Italian benefited from the activation of Greek. The participants exhibited better metalinguistic awareness abilities in Italian in the bilingual acceptability‐rating task in which Greek was activated. The benefits of the bilingual mode were visible in the judgment and explanation of errors and were modulated by syntactic processing abilities in Italian, length of exposure to Italian, type of structure, and age. The results show that metalinguistic awareness can be shared across languages. We discuss the pedagogical implications of our findings.

Language-related Episodes Patterns Engendered in L2 Classrooms: Pedagogical Translation Tasks Versus Monolingual Writing Tasks

This classroom-based longitudinal study examines the language-related episodes (LREs) engendered by pedagogical translation tasks and monolingual tasks. Specifically, the study investigated the discourse features and discussion patterns of LREs yielded in discussions of the two different types of tasks in L2 classrooms. The present study was conducted in an authentic class environment and consisted of two experiments. One group participated in class discussions after being assigned L1-L2 translation tasks, while the other took part in class discussions after working on L2 writing tasks. The LREs elicited in the two group’s class discussions over ten sessions of class discussions were scrutinized. Inductive thematic analysis shows that LREs produced in the concerned experimental tasks included ‘concern-based LREs’ and six different types of ‘response-based LREs’. Striking difference was observed in LRE patterns produced by the translation group and the writing group. The findings revealed that pedagogical translation tasks engendered LREs with deeper level of engagement and more enduring discussions than L2 monolingual tasks. The concern-based and response-based LREs arisen in this study contributes to a new framework for LRE categorization.

Visual Pivoting for (Unsupervised) Entity Alignment

This work studies the use of visual semantic representations to align entities in heterogeneous knowledge graphs (KGs). Images are natural components of many existing KGs. By combining visual knowledge with other auxiliary information, we show that the proposed new approach, EVA, creates a holistic entity representation that provides strong signals for cross-graph entity alignment. Besides, previous entity alignment methods require human labelled seed alignment, restricting availability. EVA provides a completely unsupervised solution by leveraging the visual similarity of entities to create an initial seed dictionary (visual pivots). Experiments on benchmark data sets DBP15k and DWY15k show that EVA offers state-of-the-art performance on both monolingual and cross-lingual entity alignment tasks. Furthermore, we discover that images are particularly useful to align long-tail KG entities, which inherently lack the structural contexts necessary for capturing the correspondences. Code release: https://github.com/cambridgeltl/eva; project page: http://cogcomp.org/page/publication view/927.

Fake it Till You Make it: Self-Supervised Semantic Shifts for Monolingual Word Embedding Tasks

The use of language is subject to variation over time as well as across social groups and knowledge domains, leading to differences even in the monolingual scenario. Such variation in word usage is often called lexical semantic change (LSC). The goal of LSC is to characterize and quantify language variations with respect to word meaning, to measure how distinct two language sources are (that is, people or language models). Because there is hardly any data available for such a task, most solutions involve unsupervised methods to align two embeddings and predict semantic change with respect to a distance measure. To that end, we propose a self-supervised approach to model lexical semantic change based on the perturbation of word vectors in the input corpora. We show that our method can be used for the detection of semantic change with any alignment method. Furthermore, it can be used to choose the landmark words to use in alignment and can lead to substantial improvements over the existing techniques for alignment. We illustrate the utility of our techniques using experimental results on three different datasets, involving words with the same or different meanings. Our methods not only provide significant improvements but also can lead to novel findings for the LSC problem.

Огляд нейромережевих підходів для задач умовної генерації тексту

The article is devoted to the review of conditional test generation, one of the most promising fields of natural language processing and artificial intelligence. Specifically, we explore monolingual local sequence transduction tasks: paraphrase generation, grammatical and spelling errors correction, text simplification. To give a better understanding of the considered tasks, we show examples of good rewrites. Then we take a deep look at such key aspects as publicly available datasets with the splits (training, validation, and testing), quality metrics for proper evaluation, and modern solutions based primarily on modern neural networks. For each task, we analyze its main characteristics and how they influence the state-of-the-art models. Eventually, we investigate the most significant shared features for the whole group of tasks in general and for approaches that provide solutions for them.

어휘접근 과정에서 나타나는 한국인 L2 학습자의 이중언어 음운적 동시활성화 양상

This study explored the likelihood of cross-language co-activation among Korean L2 learners of English when performing a monolingual phoneme monitoring task. The stimulus onset asynchrony (SOA: -2,000ms, +200ms, +400ms) between picture and target phoneme stimuli were manipulated. Results demonstrated that participants responded faster and showed lower error rates when performing the task in L1 compared to L2. Moreover, interference effects were observed in all three SOA-manipulated experiments irrelevant to the language type, which shows the presence of phonological co-activation of the language not- in-use during monolingual task. Stronger L1 interference effect was shown during L2 task, which reflects the difference in the degree of bilingual phonological co-activation. This study suggests that when given a short period of time to access the semantic system which is shared by two languages, phonological information of both languages are co-activated even for non-proficient L2 learners whose languages have different scripts.

Exploring Discriminative Word-Level Domain Contexts for Multi-Domain Neural Machine Translation.

Owing to its practical significance, multi-domain Neural Machine Translation (NMT) has attracted much attention recently. Recent studies mainly focus on constructing a unified NMT model with mixed-domain training corpora to switch translation between different domains. In these models, the words in the same sentence are not well distinguished, while intuitively, they are related to the sentence domain to varying degrees and thus should exert different effects on the multi-domain NMT model. In this article, we are committed to distinguishing and exploiting different word-level domain contexts for multi-domain NMT. For this purpose, we adopt multi-task learning to jointly model NMT and monolingual attention-based domain classification tasks, improving the NMT model in two ways: 1) One domain classifier and one adversarial domain classifier are introduced to conduct domain classifications of input sentences. During this process, two generated gating vectors are used to produce domain-specific and domain-shared annotations for decoder; 2) We equip decoder with an attentional domain classifier. Then, the derived attentional weights are utilized to refine the model training via word-level cost weighting, so that the impacts of target words can be discriminated by their relevance to sentence domain. Experimental results on several multi-domain translations demonstrate the effectiveness of our model.

Analysing the impact of TAPs on temporal, technical and cognitive effort in monolingual post-editing

ABSTRACTThis study analyses the impact of TAPs on effort in monolingual post-editing tasks involving two machine-translated texts from English into Brazilian Portuguese. The analysis focuses on indicators of temporal effort (task execution time, text production time, total pause time, and pause count), technical effort (numbers of insertions, deletions, navigation and return keystrokes, copy/cut-and-paste keystrokes, and mouse operations, as well as the total number of keystrokes and mouse operations), and cognitive effort (average fixation duration, fixation count, total gaze time, average pupil size, and duration of the longest fixation). Results from 43 participants indicate that temporal, technical and cognitive aspects of effort are significantly influenced by verbalization, i.e., the TAP condition, on the basis of 13 indicators: all four indicators of temporal effort, four out of seven indicators of technical effort (numbers of insertions, deletions, and mouse operations, and the total number of keystrokes and mouse operations), and all five indicators of cognitive effort. The results also point out that the impact of TAPs on post-editing effort does not depend on participants’ translation experience.

The Effect of Cognates on Cognitive Control in Late Sequential Multilinguals: A Bilingual Advantage?

The present study investigated the influence of Dutch-German cognates resp. orthographic neighbors on controlled language processing (i.e., response inhibition). Two monolingual Stroop tasks (Dutch and German) were performed by Dutch-speaking participants who could and could not speak German, and by French-speaking participants who could speak German. The question is whether or not cognate language processing affects cognitive control, resulting in a possible bilingual advantage. In the German Stroop task, we found additional advantages in congruent, as well as incongruent, trials for the two Dutch-speaking groups, which postulates the existence of a cognate resp. orthographic neighbor facilitation effect, even when participants only know one of the two cognate languages. The findings are discussed in relation to two possible factors that can modulate the effect of bilingualism on cognitive control: cognateness and orthographic neighborhood. The results suggest the existence of a notification mechanism in the bilingual brain. This mechanism would notify the bilingual brain when dealing with cognates and orthographic neighbors.

Multilingual learners in language assessment: assessment design for linguistically diverse communities

ABSTRACTThe assessments designed for and analyzed in this study used a task-based language design template rooted in theories of language reflecting heteroglossic language practices and funds of knowledge learning theories, which were understood as transforming classroom teaching, learning, and assessment through continua of biliteracy lenses. Using a participatory action research model, we created assessment instruments for pre-service English teachers in Oaxaca, Mexico, integrating language practices from communities and classrooms into assessments. Participants completed two reading and writing tasks. Task 1 was intentionally designed to engage learners’ English and Spanish languages resources. Task 2 was restricted to English-only. Our analyses indicated (1) that pre-service English teachers performed better on the multilingual task than the monolingual English task at a level of statistical significance and (2) that integrating multilingual resources within assessment design can allow test-takers to demonstrate more complex or high-order thinking skills in the language they are learning. We are offering some empirical evidence of an assessment approach that is consistent with the broadly supported principle of making use of all students’ linguistic resources for the purpose of teaching and learning.

The impact of pragmatic markers and hedging on sentence comprehension: a case study ofcommeandgenre

ABSTRACTCurrent research on conceptual and semantic representations is mainly based on prototypical word classes, such as nouns and verbs. Hence, most models of language processing and language representation rely on experimental investigations on these word classes. Until today, only a few psycholinguistic studies centre on the processing of pragmatic markers and hedges and their effect on speech comprehension. The present article aims to give experimental evidence for the processing of semantic meaning patterns and pragmatic functions of pragmatic markers. The focus will be on the question, if pragmatic markers and hedges play a role in sentence processing. This main problem will be illustrated and discussed by means of experimental data. In a monolingual sentence verification task with lexical decision, the meaning patterns and functions of the partially equivalent pragmatic French markerscommeandgenreare investigated in Canadian and European French. The results of the sentence word verification task provide evidence for an impact of pragmatic functions and semantic meaning patterns of pragmatic markers on sentence processing.

Exploring early language detection in balanced bilingual children: The impact of language-specificity on cross-linguistic nonword recognition

Aims and objectives: Recent findings on the mechanisms of lexical access suggest that bilinguals are sensitive to the orthographic structure of their languages. Several studies have demonstrated that if presented with language-specific sub-lexical information, bilingual adults use this information to speed up word recognition, which provides evidence for language-selective lexical access. In the present study, we investigated the presence of such an early language detection mechanism in children. Methodology: Forty-six balanced bilingual third-graders performed two seemingly monolingual lexical decision tasks, one in English and one in German, including nonwords with different degrees of word-likeness in each language. Data and analysis: Accuracy scores and reaction times were analyzed for nonwords using mixed-effects models with the statistical software R. Findings: Results show no impact of language-specific sub-lexical information on children’s performance in either task. We argue that bilingual lexical access is initially language-nonselective, and that sensitivity to language-specific orthographic structures first emerges over time. In contrast to bilingual adults, language detection in bilingual children is exclusively based on lexical information. Originality: The present study provides first data on the detection mechanism for language membership at the early stages of bilingual reading development. We are the first to demonstrate an important difference in the architecture of the bilingual lexicon between children and adults. Implications: Findings contribute to knowledge on the development of lexical access in bilinguals and pose limitations to the generalizability of the Bilingual Interactive Activation Plus (BIA+) extended model.

Assessing complexity and difficulty levels of machine-translated texts

This paper addresses a proposal for assessing complexity and difficulty levels of machine-translated texts in Portuguese to be further post-edited without the support of the source text (monolingual post-editing) in an experimental setting. By using two objective standard parameters, namely readability indexes and word frequency, and by proposing post-editors’ perception of difficulty to comprehend and to post-edit machine-translated texts as a new parameter, we sought to select texts with similar textual complexity or difficulty levels. This selection was necessary to carry out an experiment with four monolingual post-editing tasks in Portuguese involving machine-translated texts from three different source languages (English, Spanish, and Chinese). The application of readability indexes in conjunction with word frequency based on a corpus to analyze machine-translated texts into Portuguese to be used in experiments showed to be consistent and adequate. This method can also be applied to select texts to be used in Portuguese language classrooms and to select Portuguese texts to be included in Portuguese language textbooks. The findings can also be applied to the translation classroom, in which teachers can use the same methodology to select texts to be translated or post-edited or encourage students to analyze the texts themselves before performing a task, so students can become aware of the potential effort to be invested on a task or the real effort invested on the task after performing it. Finally, post-editors’ perception proved to be a sound parameter to validate text selection.

Cross-language and cross-modal activation in hearing bimodal bilinguals

Abstract This study investigates cross-language and cross-modal activation in bimodal bilinguals. Two groups of hearing bimodal bilinguals, natives (Experiment 1) and late learners (Experiment 2), for whom spoken Spanish is their dominant language and Spanish Sign Language (LSE) their non-dominant language, performed a monolingual semantic decision task with word pairs heard in Spanish. Half of the word pairs had phonologically related signed translations in LSE. The results showed that bimodal bilinguals were faster at judging semantically related words when the equivalent signed translations were phonologically related while they were slower judging semantically unrelated word pairs when the LSE translations were phonologically related. In contrast, monolingual controls with no knowledge of LSE did not show any of these effects. The results indicate cross-language and cross-modal activation of the non-dominant language in hearing bimodal bilinguals, irrespective of the age of acquisition of the signed language.

A New HMM-Based Voice Conversion Methodology Evaluated on Monolingual and Cross-Lingual Conversion Tasks

The work presented here proposes a new voice conversion (VC) approach based on hidden Markov models (HMMs) for spectral conversion and excitation estimation. This paper is divided in two main parts: First, an initial HMM-based VC system is presented and compared to a state-of-the-art ML-GMM VC system in a monolingual conversion scenario with parallel training data; The second part shows the necessary modifications to use the HMM VC system in a cross-lingual conversion scenario and compares it with a cross-lingual VC system based on artificial neural networks (ANNs). The results of the tests show improved performance of the proposed HMM VC system compared with both the ML-GMM and ANN-based VC alternatives, while at the same time keeping most of the flexibility afforded by the ANN approach with respect to training data requirements.

Lexical access in L2

Previous research on phonological priming in a Lexical Decision Task (LDT) has demonstrated that second language (L2) learners do not show inhibition typical for native (L1) speakers that results from lexical competition, but rather a reversed effect – facilitation (Gor, Cook, & Jackson, 2010). The present study investigates the source of the reversed priming effect and addresses two possible causes: a deficit in lexical representations and a processing constraint. Twenty-three advanced learners of Russian participated in two experiments. The monolingual Russian LDT task with priming addressed the processing constraint by manipulating the interstimulus interval (ISI, 350 ms and 500 ms). The translation task evaluated the robustness of lexical representations at both the phonolexical level (whole-word phonological representation) and the level of form-to-meaning mapping, thereby addressing the lexical deficit. L2 learners did not benefit from an increased ISI, indicating lack of support for the processing constraint. However, the study, found evidence for the representational deficit: when L2 familiarity with the words is controlled and L2 representations are robust, L2 learners demonstrate native-like processing accompanied by inhibition; however, when the words have fragmented (or fuzzy) representations, L2 lexical access is unfaithful and is accompanied by reduced lexical competition leading to facilitation effects.