Text classification methods have been widely investigated as a way to detect content of low credibility: fake news, social media bots, propaganda, etc. Quite accurate models (likely based on deep neural networks) help in moderating public electronic platforms and often cause content creators to face rejection of their submissions or removal of already published texts. Having the incentive to evade further detection, content creators try to come up with a slightly modified version of the text (known as an attack with an adversarial example) that exploit the weaknesses of classifiers and result in a different output. Here we systematically test the robustness of common text classifiers against available attacking techniques and discover that, indeed, meaning-preserving changes in input text can mislead the models. The approaches we test focus on finding vulnerable spans in text and replacing individual characters or words, taking into account the similarity between the original and replacement content. We also introduce BODEGA: a benchmark for testing both victim models and attack methods on four misinformation detection tasks in an evaluation framework designed to simulate real use cases of content moderation. The attacked tasks include (1) fact checking and detection of (2) hyperpartisan news, (3) propaganda, and (4) rumours. Our experimental results show that modern large language models are often more vulnerable to attacks than previous, smaller solutions, e.g. attacks on GEMMA being up to 27% more successful than those on BERT. Finally, we manually analyse a subset adversarial examples and check what kinds of modifications are used in successful attacks.

Text readability assessment aims to automatically evaluate the degree of reading difficulty of a given text for a specific group of readers. Most of the previous studies considered it as a classification task and explored a wide range of linguistic features to express the readability of a text from different aspects, such as semantic-based and syntactic-based features. Intuitively, when the external form of a text becomes more complex, individuals will experience more reading difficulties. Based on this motivation, our research attempts to separate the textual external form from the text and investigate its efficiency in determining readability. Specifically, in this paper, we introduce a new concept, namely textual form complexity, to provide a novel insight into text readability. The main idea is that the readability of a text can be measured by the degree to which it is challenging for readers to overcome the distractions of external textual form and obtain the text’s core semantics. To this end, we propose a set of textual form features to express the complexity of the outer form of a text and characterize its readability. Findings show that the proposed external textual form features can be used as effective evaluation indexes to indicate the readability of text. It brings a new perspective to the existing research and provides a new complement to the existing rich features.

A vast amount of clinical data are still stored in unstructured text. Automatic extraction of medical information from these data poses several challenges: high costs of clinical expertise, restricted computational resources, strict privacy regulations, and limited interpretability of model predictions. Recent domain adaptation and prompting methods using lightweight masked language models showed promising results with minimal training data and allow for application of well-established interpretability methods. We are first to present a systematic evaluation of advanced domain-adaptation and prompting methods in a lower-resource medical domain task, performing multi-class section classification on German doctor’s letters. We evaluate a variety of models, model sizes (further-pre)training and task settings, and conduct extensive class-wise evaluations supported by Shapley values to validate the quality of small-scale training data and to ensure interpretability of model predictions. We show that in few-shot learning scenarios, a lightweight, domain-adapted pretrained language model, prompted with just 20 shots per section class, outperforms a traditional classification model, by increasing accuracy from $48.6\%$ to $79.1\%$. By using Shapley values for model selection and training data optimization, we could further increase accuracy up to $84.3\%$. Our analyses reveal that pretraining of masked language models on general-language data is important to support successful domain-transfer to medical language, so that further-pretraining of general-language models on domain-specific documents can outperform models pretrained on domain-specific data only. Our evaluations show that applying prompting based on general-language pretrained masked language models combined with further-pretraining on medical-domain data achieves significant improvements in accuracy beyond traditional models with minimal training data. Further performance improvements and interpretability of results can be achieved, using interpretability methods such as Shapley values. Our findings highlight the feasibility of deploying powerful machine learning methods in clinical settings and can serve as a process-oriented guideline for lower-resource languages and domains such as clinical information extraction projects.

Machine reading comprehension (MRC) refers to the process of instructing machines to comprehend and respond to inquiries based on a provided text. There are two primary methodologies for achieving this: extracting answers directly from the text or predicting them. Extracting answers involves anticipating the specific segment of text containing the answer, pinpointed by its starting and ending indices within the paragraph. Despite the increasing interest in MRC, exploration within the framework of the Arabic language faces limitations due to various challenges. A significant impediment arises from the inadequacy of resources available for Arabic textual content, which impedes the development of effective models. Furthermore, the inherent intricacies of Arabic, manifesting in its diverse linguistic forms including classical, modern standard, and colloquial, present distinctive hurdles for tasks involving language comprehension. This paper proposes an enhanced version of the bidirectional attention flow (BIDAF) model for Arabic MRC, constructed upon the Arabic Span-Extraction-based Reading Comprehension Benchmark (ASER). ASER comprises 10,000 sets of questions, answers, and passages, partitioned into a training set constituting 90% of the data and a testing set making up the remaining 10%. By introducing a new input feature based on parts-of-speech (POS) word embeddings and replacing Bidirectional Long Short-Term Memory (bi-LSTM) with bidirectional gated recurrent unit, significant improvements were observed. Eight different POS word embeddings were generated using both Continuous Bag of Words (CBOW) and Skip-gram methods, with varying dimensionalities. Evaluation metrics, including exact match (EM) and F1-measure, were utilized to assess model performance, with emphasis on the latter for its accuracy. The proposed enhanced BIDAF model achieved a remarkable accuracy of 75.22% on the ASER dataset, demonstrating its efficacy in Arabic MRC tasks. Additionally, rigorous statistical evaluation using a two-tailed paired samples t-test further validated the findings, highlighting the significance of the proposed enhancements in advancing Arabic language processing capabilities.

This paper addresses the optimization of retrieval-augmented generation (RAG) processes by exploring various methodologies, including advanced RAG methods. The research, driven by the need to enhance RAG processes as highlighted by recent studies, involved a grid-search optimization of 23,625 iterations. We evaluated multiple RAG methods across different vectorstores, embedding models, and large language models, using cross-domain datasets and contextual compression filters. The findings emphasize the importance of balancing context quality with similarity-based ranking methods, as well as understanding tradeoffs between similarity scores, token usage, runtime, and hardware utilization. Additionally, contextual compression filters were found to be crucial for efficient hardware utilization and reduced token consumption, despite the evident impacts on similarity scores, which may be acceptable depending on specific use cases and RAG methods.

Emotion recognition in conversation (ERC) faces two major challenges: biased predictions and poor calibration. Classifiers often disproportionately favor certain emotion categories, such as neutral, due to the structural complexity of classifiers, the subjective nature of emotions, and imbalances in training datasets. This bias results in poorly calibrated predictions where the model’s predicted probabilities do not align with the true likelihood of outcomes. To tackle these problems, we introduce the application of conformal prediction (CP) into ERC tasks. CP is a distribution-free method that generates set-valued predictions to ensure marginal coverage in classification, thus improving the calibration of models. However, inherent biases in emotion recognition models prevent baseline CP from achieving a uniform conditional coverage across all classes. We propose a novel CP variant, class spectrum conformation, which significantly reduces coverage bias in CP methods. The methodologies introduced in this study enhance the reliability of prediction calibration and mitigate bias in complex natural language processing tasks.

It is increasingly common to use chatbots as an interface to services. One of the main components of a chatbot is the Natural Language Understanding (NLU) model, which is responsible for interpreting the text and extracting the intent and entities present in that text. It’s possible to focus only on one of these tasks of NLU, such as intent classification. To train an NLU intent classification model, it’s generally necessary to use a considerable amount of annotated data, where each sentence of the dataset receives a label indicating an intent. Performing manually labeling data is arduous and impracticable, depending on the data volume. Thus, an unsupervised machine learning technique, such as data clustering, could be applied to find and label patterns in the data. For this task, it is essential to have an effective vector embedding representation of texts that depicts the semantic information and helps the machine understand the context, intent, and other nuances of the entire text. This paper extensively evaluates different text embedding models for clustering and labeling. We also apply some operations to improve the dataset’s quality, such as removing sentences and establishing various strategies for distance thresholds (cosine similarity) for the clusters’ centroids. Then, we trained some intent classification Models with two different architectures, one built with the Rasa framework and the other with a neural network (NN) using the attendance text from the Coronavirus Platform Service of Ceará, Brazil. We also manually annotated a dataset to be used as validation data. We conducted a study on semiautomatic labeling, implemented through clustering and visual inspection, which introduced some labeling errors to the intent classification models. However, it would be unfeasible to annotate the entire dataset manually. Nevertheless, results of competitive accuracy were still achieved with the trained models.

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.

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.

Hierarchical text classification (HTC) is a natural language processing task which aims to categorise a text document into a set of classes from a hierarchical class structure. Recent approaches to solve HTC tasks focus on leveraging pre-trained language models (PLMs) and the hierarchical class structure by allowing these components to interact in various ways. Specifically, the Hierarchy-aware Prompt Tuning (HPT) method has proven to be effective in applying the prompt tuning paradigm to Bidirectional Encoder Representations from Transformers (BERT) models for HTC tasks. Prompt tuning aims to reduce the gap between the pre-training and fine-tuning phases by transforming the downstream task into the pre-training task of the PLM. Discriminative PLMs, which use a replaced token detection (RTD) pre-training task, have also shown to perform better on flat text classification tasks when using prompt tuning instead of vanilla fine-tuning. In this paper, we propose the Hierarchy-aware Prompt Tuning for Discriminative PLMs (HPTD) approach which injects the HTC task into the RTD task used to pre-train discriminative PLMs. Furthermore, we make several improvements to the prompt tuning approach of discriminative PLMs that enable HTC tasks to scale to much larger hierarchical class structures. Through comprehensive experiments, we show that our method is robust and outperforms current state-of-the-art approaches on two out of three HTC benchmark datasets.

Hate speech against individuals and groups with certain demographics is a major issue in social media. Supervised models for hate speech detection mostly utilize labeled data collections to understand textual semantics. However, hate speech detection is a complex task that involves several aspects, including topic and writing style. The complexity of hate speech can be represented by an ensemble of models learned from different aspects of data. Moreover, ensemble members or base models can be modified to give attention to particular aspects of hate speech. In this study, we extract different aspects of hate speech to construct ensembles, thereby improving the performance of hate speech detection by ensemble learning. We conduct detailed experiments on five datasets in multiple languages to generalize our observations. The experimental results, supported by statistical significance tests, show that the performance of hate speech detection can be improved by capturing multiple aspects of hate speech. Our ensemble construction approach outperforms the baselines in terms of the F1 score of the Hate class in 80% of the cases, and the Offensive class in 75% of the cases. We also compare our approach with state-of-the-art ensemble methods from shared tasks and find that our highest-performing method can improve the performance of the Hate class in two out of three datasets. We further discuss our approach and experimental results in terms of ensemble parameters and writing style among ensemble members.

The proliferation of social networks has caused an increase in the amount of textual content generated by users. The voluminous nature of such content poses a challenge to users, necessitating the development of technological solutions for automatic summarisation. This paper presents a two-stage framework for generating abstractive summaries from a collection of Twitter texts. In the first stage of the framework, event detection is carried out through clustering, followed by event summarisation in the second stage. Our approach involves generating contextualised vector representations of tweets and applying various clustering techniques to the vectors. The quality of the resulting clusters is evaluated, and the best clusters are selected for the summarisation task based on this evaluation. In contrast to previous studies, we experimented with various clustering techniques as a preprocessing step to obtain better event representations. For the summarisation task, we utilised pre-trained models of three state-of-the-art deep neural network architectures and evaluated their performance on abstractive summarisation of the event clusters. Summaries are generated from clusters that contain (a) unranked tweets, (b) all ranked tweets, and (c) the top 10 ranked tweets. Of these three sets of clusters, we obtained the best ROUGE scores from the top 10 ranked tweets. From the summaries generated from the clusters containing the top ten tweets, we obtained ROUGE-1 F score of 48%, ROUGE-2 F score of 37%, ROUGE-L F score of 44%, and ROUGE-SU F score of 33% which suggests that if relevant tweets are at the top of a cluster, and then better summaries are generated.

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.

Syntactic parsing is one of the areas in Natural Language Processing. The development of large-scale multilingual language models has enabled cross-lingual parsing approaches, which allows us to develop parsers for languages that do not have treebanks available. However, these approaches rely on the assumption that languages share orthographic representations and lexical entries. In this article, we investigate methods for developing a dependency parser for Xibe, a low-resource language that is written in a unique script. We first investigate lexicalized monolingual dependency parsing experiments to examine the effectiveness of word, part-of-speech, and character embeddings as well as pre-trained language models. Results show that character embeddings can significantly improve performance, while pre-trained language models decrease performance since they do not recognize the Xibe script. We also train delexicalized monolingual models, which yield competitive results to the best lexicalized model. Since the monolingual models are trained on a very small training set, we also investigate lexicalized and delexicalized cross-lingual models. We use six closely related languages as source language, which cover a wide range of scripts. In this setting, the delexicalized models achieve higher performance than lexicalized models. A final experiment shows that we can increase performance of the cross-lingual model by combining source languages and selecting the most similar sentences to Xibe as training set. However, all cross-lingual parsing results are still considerably lower than the monolingual model. We attribute the low performance of cross-lingual methods to syntactic and annotation differences as well as to the impoverished input of Universal Dependency Part-of-Speech tags that the delexicalized model has access to.

Sharing ideas through communication with peers is the primary mode of human interaction. Consequently, extensive research has been conducted in the area of conversational AI, leading to an increase in the availability and diversity of conversational tasks, datasets, and methods. However, with numerous tasks being explored simultaneously, the current landscape of conversational AI has become fragmented. Consequently, initiating a well-thought-out model for a dialogue agent can pose significant challenges for a practitioner. Toward highlighting the critical ingredients needed for a practitioner to design a dialogue agent from scratch, the current study provides a comprehensive overview of the primary characteristics of a dialogue agent, the supporting tasks, their corresponding open-domain datasets, and the methods used to benchmark these datasets. We observe that different methods have been used to tackle distinct dialogue tasks. However, building separate models for each task is costly and does not leverage the correlation among the several tasks of a dialogue agent. As a result, recent trends suggest a shift toward building unified foundation models. To this end, we propose Unit, a Unified dialogue dataset constructed from conversations of varying datasets for different dialogue tasks capturing the nuances for each of them. We then train a Unified dialogue foundation model, GPT-2$^{\textrm{U}}$ and present a concise comparative performance of GPT-2$^{\textrm{U}}$ against existing large language models. We also examine the evaluation strategies used to measure the performance of dialogue agents and highlight the scope for future research in the area of conversational AI with a thorough discussion of popular models such as ChatGPT.

Linguistic synesthesia as a productive figurative language usage has received little attention in the field of Natural Language Processing (NLP). Although linguistic synesthesia is similar to metaphor concerning involving conceptual mappings and showing great usefulness in the NLP tasks such as sentiment analysis and stance detection, the well-studied methods of metaphor detection cannot be applied to the detection of linguistic synesthesia directly. This study incorporates comprehensive linguistic features (i.e., character and radical information, word segmentation information, and part-of-speech tagging) into a neural model to detect linguistic synesthetic usages in a sentence automatically. In particular, we employ a span-based boundary detection model to extract sensory words. In addition, a joint model is proposed to detect the original and synesthetic modalities of the sensory words collectively. Based on the experiments, our model is shown to achieve state-of-the-art results on the dataset for linguistic synesthesia detection. The results prove that leveraging culturally enriched linguistic features and joint learning are effective in linguistic synesthesia detection. Furthermore, as the proposed model leverages non-language-specific linguistic features, the model would be applied to the detection of linguistic synesthesia in other languages.

With recent advances in natural language generation, risks associated with the rapid proliferation and misuse of generative language models for malicious purposes steadily increase. Artificial text detection (ATD) has emerged to develop resources and computational methods to mitigate these risks, such as generating fake news and scientific article reviews. This paper introduces corpus of artificial texts (CoAT), a large-scale corpus of human-written and generated texts for the Russian language. CoAT spans six domains and comprises outputs from 13 text generation models (TGMs), which differ in the number of parameters, architectural choices, pre-training objectives, and downstream applications. We detail the data collection methodology, conduct a linguistic analysis of the corpus, and present a detailed analysis of the ATD experiments with widely used artificial text detectors. The results demonstrate that the detectors perform well on the seen TGMs, but fail to generalise to unseen TGMs and domains. We also find it challenging to identify the author of the given text, and human annotators significantly underperform the detectors. We release CoAT, the codebase, two ATD leaderboards, and other materials used in the paper.

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.

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.

Part of speech (POS) tagging, though considered to be preliminary to any Natural Language Processing (NLP) task, is crucial to account for, especially in low resource language like Khasi that lacks any form of formal corpus. POS tagging is context sensitive. Therefore, the task is challenging. In this paper, we attempt to investigate a deep learning approach to the POS tagging problem in Khasi. A deep learning model called Robustly Optimized BERT Pretraining Approach (RoBERTa) is pretrained for language modelling task. We then create RoBERTa for POS (RoPOS) tagging, a model that performs POS tagging by fine-tuning the pretrained RoBERTa and leveraging its embeddings for downstream POS tagging. The existing tagset that has been designed, customarily, for the Khasi language is employed for this work, and the corresponding tagged dataset is taken as our base corpus. Further, we propose additional tags to this existing tagset to meet the requirements of the language and have increased the size of the existing Khasi POS corpus. Other machine learning and deep learning models have also been tried and tested for the same task, and a comparative analysis is made on the various models employed. Two different setups have been used for the RoPOS model, and the best testing accuracy achieved is 92 per cent. Comparative analysis of RoPOS with the other models indicates that RoPOS outperforms the others when used for inferencing on texts that are outside the domain of the POS tagged training dataset.