Transformer-based neural language models achieve state-of-the-art performance on various natural language processing tasks. However, an open question is the extent to which these models rely on word-order/syntactic or word co-occurrence/topic-based information when processing natural language. This work contributes to this debate by addressing the question of whether these models primarily use topic as a signal, by exploring the relationship between Transformer-based models’ (BERT and RoBERTa’s) performance on a range of probing tasks in English, from simple lexical tasks such as sentence length prediction to complex semantic tasks such as idiom token identification, and the sensitivity of these tasks to the topic information. To this end, we propose a novel probing method which we call topic-aware probing. Our initial results indicate that Transformer-based models encode both topic and non-topic information in their intermediate layers, but also that the facility of these models to distinguish idiomatic usage is primarily based on their ability to identify and encode topic. Furthermore, our analysis of these models’ performance on other standard probing tasks suggests that tasks that are relatively insensitive to the topic information are also tasks that are relatively difficult for these models.

Chinese spelling correction has achieved significant progress, but critical challenges remain, especially in handling visually and phonetically similar errors within complex syntactic structures. This paper introduces a novel approach combining a Long Short-Term Memory Network (LSTM)-enhanced Transformer for error detection and Bidirectional Encoder Representations from Transformers (BERT)-based correction with a dynamic adaptive weighting scheme. Transformer uses global attention mechanism to capture dependencies between any two positions in the input sequence. By processing each token in the sequence recursively, LSTM is able to more finely capture local context and sequential information within the sequence. Based on adaptive weighting coefficient, weights of multi-task learning are automatically adjusted to help the model better balance the learning process between the detection and correction network, enabling it to converge faster and achieve higher precision. Comprehensive evaluations demonstrate improved performance over existing baselines, particularly in addressing complex error patterns.

Audits of multilingual resources are reporting shockingly poor quality: “less than 50% … acceptable quality.” There is too much translationese in too many of our multilingual resources, e.g., Wikipedia, XNLI, FLORES, WordNet. We view translationese as a form of noise that makes it hard to generalize from a benchmark based on translation to a real task of interest that does not involve translation. Worse, too much of this translationese is in the “wrong” direction. Directionality matters. Professional translators translate from their weaker language into their stronger language. Unfortunately, many of our resources translate in the other direction, from a stronger (higher-resource) language into a weaker (lower-resource) language. In Wikipedia, for example, there is more translation out of English than into English. We recommend more investments in high-quality data, and less in translation, especially in the “wrong” direction.

In machine translation (MT), one of the challenging tasks is to translate the proper nouns and technical terms from the source language to the target language while preserving the phonetic equivalent of original term. Machine transliteration, an essential part of MT systems, plays a vital role in handling proper nouns and technical terms. In this paper, a hybrid attention-based encoder–decoder machine transliteration system is proposed for the low-resource English to the Assamese language. In this work, the proposed machine transliteration system is integrated with the previously published hybrid attention-based encoder–decoder neural MT model to improve the translation quality of English to the Assamese language. The proposed integrated MT system demonstrated good results across various performance metrics such as BLEU, sacreBLEU, METEOR, chrF, RIBES, and TER for English to Assamese translation. Additionally, human evaluation was also conducted to assess translation quality. The proposed integrated MT system was compared with two existing systems: the Bing translation service model and the Samanantar Indic translation model.

Recently, deep learning methods have achieved remarkable success in the Chinese word segmentation (CWS) task. Some of them enhance the CWS model by utilizing contextual features and external resources (e.g., sub-words, lexicon, and syntax). However, existing approaches fail to fully use the heterogeneous features and their structural information. Therefore, in this paper, we propose a heterogeneous information learning framework for CWS, named heterogeneous graph neural segmenter (HGNSeg), which exploits heterogeneous features with the graph convolutional networks and the pretrained language model. Experimental results on six benchmark datasets (e.g., SIGHAN 2005 and SIGHAN 2008) confirm that HGNSeg can effectively improve the performance of CWS. Importantly, HGNSeg also demonstrates an excellent ability to alleviate the out-of-vocabulary (OOV) issue in cross-domain scenarios.

Statistical Machine Translation (SMT) systems use various probabilistic and statistical Natural Language Processing (NLP) methods to automatically translate from one language to another language while retaining the originality of the context. This paper aims to discuss the development of bilingual SMT models for translating English into fifteen low-resource Indic languages (ILs) and vice versa. The process to build the SMT model is described and explained using a workflow diagram. Samanantar and OPUS corpus are utilized for training, and Flores200 corpus is used for fine-tuning and testing purposes. The paper also highlights various preprocessing methods used to deal with corpus noise. The Moses open-source SMT toolkit is being investigated for the system’s development. The impact of distance-based reordering and Morpho-syntactic Descriptor Bidirectional Finite-State Encoder (msd-bidirectional-fe) reordering on ILs is compared in the paper. This paper provides a comparison of SMT models with Neural Machine Translation (NMT) for ILs. All the experiments assess the translation quality using standard metrics such as BiLingual Evaluation Understudy, Rank-based Intuitive Bilingual Evaluation Score, Translation Edit Rate, and Metric for Evaluation of Translation with Explicit Ordering. From the result, it is observed that msd-bidirectional-fe reordering performs better than the distance-based reordering model for ILs. It is also noticed that even though the IL-English and English-IL systems are trained using the same corpus, the former performs better for all the evaluation metrics. The comparison between SMT and NMT shows that across various languages, SMT performs better in some cases, while NMT outperforms in others.

Multilingual question answering (MQA) is an effective access to multilingual data to provide accurate and precise answers, irrespective of language. Although a wide range of datasets is available for monolingual QA systems in natural language processing, benchmark datasets specifically designed for MQA are considerably limited. The absence of comprehensive and benchmark datasets hinders the development and evaluation of MQA systems. To overcome this issue, the proposed work attempts to develop the EHMQuAD dataset, an MQA dataset for low-resource languages such as Hindi and Marathi accompanying the English language. The EHMQuAD dataset is developed using a synthetic corpora generation approach, and an alignment is performed after translation to make the dataset more accurate. Further, the EHMMQA model is proposed to create an abstract framework that uses a deep neural network that accepts pairs of questions and context and returns an accurate answer based on those questions. The shared question and shared context representation have been designed separately to develop this system. The experiments of the proposed model are conducted on the MMQA, Translated SQuAD, XQuAD, MLQA, and EHMQuAD datasets, and EM and F1-score are used as performance measures. The proposed model (EHMMQA) is collated with state-of-the-art MQA baseline models for all possible monolingual and multilingual settings. The results signify that EHMMQA is a considerable step toward the MQA system utilizing Hindi and Marathi languages. Hence, it becomes a new state-of-the-art model for Hindi and Marathi languages.

Multilingual pre-trained models make it possible to develop natural language processing (NLP) applications for low-resource languages (LRLs) using the model of resource-rich languages (RRLs). However, the structural characteristics of the target languages can impact task-specific learning. In this paper, we investigate the influence of structural diversities of languages on the system’s overall performance. Specifically, we propose a customized approach to leverage task-specific data of low-resource language families via transfer learning from RRL. Our findings are based on question-answering tasks using the XLM-R, mBERT, and IndicBERT transformer models and Indic languages (Hindi, Bengali, and Telugu). On the XQuAD-Hindi dataset, the few-shot learning using Bengali improves the benchmark mBERT (F1/EM) score by +(10.86/7.87) and XLM-R score by +(3.84/4.42). Few-shot learning using Telugu has also improved the mBERT score by +(10.42/7.36) and +(3.04/2.72) for XLM-R. In addition, our model has demonstrated benchmark-compatible performance in a zero-shot setup with single-epoch task learning. This approach can be adapted for other NLP tasks for LRLs.

At the London Tech Week event in early June, Nvidia CEO Jensen Huang praised the UK as the ‘envy of the world’ when it comes to AI researchers, but he also criticised it as the largest AI ecosystem in the world without its own infrastructure. The criticism is somewhat self-serving: when the UK does get around to building out that infrastructure, it’s certain to consist largely of chips sold by Huang’s company. It’s also unsurprising: Huang has been pitching the idea of ‘sovereign AI’ since at least 2023, conscious that nation states are the next deep pockets to target after the hyperscalers and generously funded model builders. In a world where the only real contenders in the race for AI supremacy are the US and China, we look at how the pursuit of AI sovereignty is playing out across the rest of the planet.

Warning: This paper is based on hate speech detection and may contain examples of abusive/ offensive phrases.

Cyberbullying, online harassment, etc., via offensive comments are pervasive across different social media platforms like ™Twitter, ™Facebook, ™YouTube, etc. Hateful comments must be detected and eradicated to prevent harassment and violence on social media. In the Natural Language Processing (NLP) domain, the most prevalent task is comment classification, which is challenging, and language models based on transformers are at the forefront of this advancement. This paper intends to analyze the performance of language models based on transformers like BERT, ALBERT, RoBERTa, and DistilBERT on the Indian hate speech datasets over binary classification. Here, we utilize the existing datasets, i.e., HASOC (Hindi and Marathi) and HS-Bangla. So, we evaluate several multilingual language models like MuRIL-BERT, XLM-RoBERTa, etc., few monolingual language models like RoBERTa-Hindi, Maha-BERT (Marathi), Bangla-BERT (Bangla), Assamese-BERT (Assamese), etc., and perform cross-lingual experiment also. For further analyses, we perform multilingual, monolingual, and cross-lingual experiments on our Hate Speech Assamese (HS-Assamese) (Indo-Aryan language family) and Hate Speech Bodo (HS-Bodo) (Sino-Tibetan language family) dataset (HS dataset version 2) also and achieved a promising result. The motivation of the cross-lingual experiment is to encourage researchers to learn about the power of the transformer. Note that no pre-trained language models are currently available for Bodo or any other Sino-Tibetan languages.

This paper provides data resources for low-resource hate speech detection. Specifically, we introduce two different data resources: (i) the HateBR 2.0 corpus, which is composed of 7,000 comments extracted from Brazilian politicians’ accounts on Instagram and manually annotated a binary class (offensive versus non-offensive) and hate speech targets. It consists of an updated version of the HateBR corpus, in which highly similar and one-word comments were replaced; and (ii) the multilingual offensive lexicon (MOL), which consists of 1,000 explicit and implicit terms and expressions annotated with context information. The lexicon also comprises native-speaker translations and its cultural adaptations in English, Spanish, French, German, and Turkish. Both corpus and lexicon were annotated by three different experts and achieved high inter-annotator agreement. Lastly, we implemented baseline experiments on the proposed data resources. Results demonstrate the reliability of data outperforming baseline dataset results in Portuguese, besides presenting promising results for hate speech detection in different languages.

There has been a steep rise in user-generated content on the Web and social media platforms during the last few years. While the ease of content creation allows anyone to create content, at the same time it is difficult to monitor and control the spread of detrimental content. Recent research in natural language processing and machine learning has shown some hope for the purpose. Approaches and methods are now being developed for the automatic flagging of problematic textual content, namely hate speech, cyberbullying, or fake news, though mostly for English language texts. This paper presents an algorithmic approach based on deep learning models for the detection of violent incidents from tweets in the Spanish language (binary classification) and categorizes them further into five classes – accident, homicide, theft, kidnapping, and none (multi-label classification). The performance is evaluated on the recently shared benchmark dataset, and it is found that the proposed approach outperforms the various deep learning models, with a weighted average precision, recall, and F1-score of 0.82, 0.81, and 0.80, respectively, for the binary classification. Similarly, for the multi-label classification, the proposed model reports weighted average precision, recall, and F1-score of 0.54, 0.79, and 0.64, respectively, which is also superior to the existing results reported in the literature. The study, thus, presents meaningful contribution to detection of violent incidents in Spanish language social media posts.

The exponential growth of social media data in the era of Web 2.0 has necessitated advanced techniques for sentiment analysis. While sentiment analysis in monolingual datasets has received significant attention that in code-mixed datasets still need to be studied more. Code-mixed data often contain a mixture of monolingual content (might be in transliterated form), single-script but multilingual content, and multi-script multilingual content. This paper explores the issue from three important angles. What will be the best strategy to deal with the data for sentiment detection? Whether to train the classifier with the whole of the dataset or only with the pure code-mixed subset from the dataset? How much important is the language identification (LID) for the task? If LID is to be done, how, and when will it be used to yield the best performance? We explore the questions in the light of three datasets of Tamil–English, Kannada–English, and Malayalam–English YouTube social media comments. Our solution incorporated mBERT and an optional LID module. We report our results using a set of metrics like precision, recall, $F_1$ score, and accuracy. The solutions provide considerable performance gain and some interesting insights for sentiment analysis from code-mixed data.

In spite of the superior performance deep neural networks have proven in thousands of applications in the past few years, addressing the over-sensitivity of these models to noise and/or intentional slight perturbations is still an active area of research. In the computer vision domain, perturbations can be directly applied to the input images. The task in the natural language processing domain is quite harder due to the discrete nature of natural languages. There has been a considerable amount of effort put to address this problem in high-resource languages like English. However, there is still an apparent lack of such studies in the Arabic language, and we aim to be the first to conduct such a study in this work. In this study, we start by training seven different models on a sentiment analysis task. Then, we propose a method to attack our models by means of the worst synonym replacement where the synonyms are automatically selected via the gradients of the input representations. After proving the effectiveness of the proposed adversarial attack, we aim to design a framework that enables the development of models robust to attacks. Three different frameworks are proposed in this work and a thorough comparison between the performance of these frameworks is presented. The three scenarios revolve around training the proposed models either on adversarial samples only or also including clean samples beside the adversarial ones, and whether or not to include weight perturbation during training.

Individuals with complex communication needs often rely on augmentative and alternative communication (AAC) systems to have conversations and communicate their wants. Such systems allow message authoring by arranging pictograms in sequence. However, the difficulty of finding the desired item to complete a sentence can increase as the user’s vocabulary increases. This paper proposes using BERTimbau, a Brazilian Portuguese version of Bidirectional Encoder Representations from Transformers (BERT), for pictogram prediction in AAC systems. To fine-tune BERTimbau, we constructed an AAC corpus for Brazilian Portuguese to use as a training corpus. We tested different approaches to representing a pictogram for prediction: as a word (using pictogram captions), as a concept (using a dictionary definition), and as a set of synonyms (using related terms). We also evaluated the usage of images for pictogram prediction. The results demonstrate that using embeddings computed from the pictograms’ caption, synonyms, or definitions have a similar performance. Using synonyms leads to lower perplexity, but using captions leads to the highest accuracies. This paper provides insight into how to represent a pictogram for prediction using a BERT-like model and the potential of using images for pictogram prediction.

Intent detection and slot filling are two necessary tasks for natural language understanding. Deep neural models have already shown great ability facing sequence labeling and sentence classification tasks, but they require a large amount of training data to achieve accurate results. However, in many low-resource languages, creating accurate training data is problematic. Consequently, in most of the language processing tasks, low-resource languages have significantly lower accuracy than rich-resource languages. Hence, training models in low-resource languages with data from a richer-resource language can be advantageous. To solve this problem, in this paper, we used pretrained language models, namely multilingual BERT (mBERT) and XLM-RoBERTa, in different cross-lingual and monolingual scenarios. To evaluate our proposed model, we translated a small part of the Airline Travel Information System (ATIS) dataset into Persian. Furthermore, we repeated the experiments on the MASSIVE dataset to increase our results’ reliability. Experimental results on both datasets show that the cross-lingual scenarios significantly outperform monolinguals ones.

Decompounding is an essential preprocessing step in text-processing tasks such as machine translation, speech recognition, and information retrieval (IR). Here, the IR issues are explored from five viewpoints. (A) Does word decompounding impact the Indian language IR? If yes, to what extent? (B) Can corpus-based decompounding models be used in the Indian language IR? If yes, how? (C) Can machine learning and deep learning-based decompounding models be applied in the Indian language IR? If yes, how? (D) Among the different decompounding models (corpus-based, hybrid machine learning-based, and deep learning-based), which provides the best effectiveness in the IR domain? (E) Among the different IR models, which provides the best effectiveness from the IR perspective? This study proposes different corpus-based, hybrid machine learning-based, and deep learning-based decompounding models in Indian languages (Marathi, Hindi, and Sanskrit). Moreover, we evaluate the effectiveness of each activity from an IR perspective only. It is observed that the different decompounding models improve IR effectiveness. The deep learning-based decompounding models outperform the corpus-based and hybrid machine learning-based models in Indian language IR. Among the different deep learning-based models, the Bi-LSTM-A model performs best and improves mean average precision (MAP) by 28.02% in Marathi. Similarly, the Bi-RNN-A model improves MAP by 18.18% and 6.1% in Hindi and Sanskrit, respectively. Among the retrieval models, the In_expC2 model outperforms others in Marathi and Hindi, and the BB2 model outperforms others in Sanskrit.