Recent advances in language modeling have contributed to a growing emphasis on machine ethics, including researching and assessing moral judgments made by large language models (LLMs). This paper provides a critical survey of the existing datasets for exactly this assessment, with a special focus on the respective data sources. We address the current lack of theoretical grounding by providing an introduction to ethics and different frameworks from moral philosophy and moral psychology. Moreover, we identify four main data sources: webcrawled corpora, scholars, laypeople, and synthetic data generation. By discussing the strengths and weaknesses of these sources, we analyze their implications for the assessment of moral judgment. Importantly, systemizing the available datasets reveals an over-reliance on previous work, reinforcing existing shortcomings. Addressing the current limitations, we recommend adopting a consistent terminology and creating independently curated datasets based on interdisciplinary work. To ensure a clear delineation of normative approaches, we propose focusing on the assessment of moral consistency and certainty of LLMs as effective and well-defined indicators of their performance on moral judgment.
In a digital world where cybercrimes are on the rise, women and children in India face unique challenges that are often overlooked by existing solutions. The globally available platforms fail to address the nuances of Indian laws and the cultural stigma surrounding sensitive issues like mental health and legal aid. Our chatbot bridges this gap, offering a safe, confidential, and culturally relevant space for victims of cybercrime to seek help. Tailored specifically to the Indian context, the chatbot provides legal guidance based on local laws, empowering users to understand their rights and encouraging them to seek help. It also offers compassionate mental health counselling, addressing the emotional toll of cybercrime without the fear of judgment or taboo. By providing a seamless blend of legal aid and psychological support, this chatbot ensures that women and children affected by cybercrime have a reliable, accessible, and non-threatening resource to turn to.
This paper presents experiments on language identification for a Ukrainian-Russian code-switching dataset. Code-switching, a common phenomenon in multilingual societies, presents significant challenges for natural language processing. This study discusses various issues encountered during dataset creation, emphasizing the complexity of accurately annotating code-switching text. The study describes cases where identifying the language of individual tokens in sentences that switch between Ukrainian and Russian proves difficult even for human annotators. The relatedness of the languages and the use of Cyrillic in both orthographic systems complicate the task, leading to many cases where words are spelled identically despite clear phonetic differences between the languages that are not reflected in writing. The study explores different models and libraries for language identification on the token level. Experimental results suggest that BERT shows promising performance; however, other models, such as CRFs with n-grams, Char-level BiLSTM, and Word-level Neural Networks, are also promising for this task. This research contributes to the development of language processing technologies for multilingual contexts, with potential applications in sentiment analysis, information retrieval, and social media monitoring.
This paper presents an approach to computationally detecting face-threatening and paired actions in asynchronous online conversations. Action detection has been widely studied for synchronous chats. However, there are fewer models or datasets for asynchronous conversations, and they have not included some of the face-threatening actions central to online conversations involving misbehavior like trolling. We examine asynchronous crisis news related online conversations in Finnish, providing an annotation scheme for identifying central actions used in this conversational context. An important contribution is to include face-threatening actions in the scheme, and training computational classifiers for their detection with improved performance compared to prior work. We illustrate that face-threatening actions are important for analyzing conversations related to crisis news. We show that for computational action detection, it is essential to be able to represent how multiple actions may be performed within one comment, and how ambiguity in the expression of actions often leads to multiple possible label interpretations. Annotating actions using scores helps to reflect these characteristics. We also find that an ensemble of models trained on individual annotators’ annotations can best represent multiple potential interpretations of action labels. These are especially relevant for face-threatening actions.
The caused-motion construction (CMC, “She sneezed the foam off her cappuccino”) is one of the most well-studied constructions in Construction Grammar (CxG). It is a prime example for describing how constructions must carry meaning, as otherwise the fact that “sneeze” in this context takes two arguments and causes motion cannot be explained. We form the hypothesis that this remains challenging even for state-of-the-art Large Language Models (LLMs), for which we devise a test based on substituting the verb with a prototypical motion verb. To be able to perform this test at a statistically significant scale, in the absence of adequate CxG corpora, we develop a novel pipeline of NLP-assisted collection of linguistically annotated text. We show how dependency parsing and LLMs can be used to significantly reduce annotation cost and thus enable the annotation of rare phenomena at scale. We then evaluate OpenAI, Gemma3, Llama3, OLMo2, Mistral and Aya models for their understanding of the CMC using the newly collected corpus. We find that most models struggle with understanding the motion component that the CMC adds to a sentence.
Control vectors have recently gained popularity as a method for steering transformer-based generative language models. This paper contributes to this path of research by evaluating the robustness of these control vectors in multi- and cross-lingual question-answering settings mimicking the real-world deployment scenario, where models are expected to generate answers to challenging questions. We present a set of experiments to demonstrate that a control vector approach can be used to shift the output of a generative language model from one language to another, and to exercise stylistic control of the output across languages. Overall, we find that the control vector approach offers a relatively lightweight and effective path for developing methods to control the output of multilingual language models with multiple design choices affecting the real-world control performance.
Contemporary large-scale data collection efforts have prioritized the amount of data collected to improve large language models (LLM). This quantitative approach has resulted in concerns for the rights of data subjects represented in data collections. This concern is exacerbated by a lack of documentation and analysis tools, making it difficult to interrogate these collections. Mindful of these pitfalls, we present a methodology for documentation-first, human-centered data collection. We apply this approach in an effort to train a multilingual LLM. We identify a geographically diverse set of target language groups (Arabic varieties, Basque, Chinese varieties, Catalan, English, French, Indic languages, Indonesian, Niger-Congo languages, Portuguese, Spanish, and Vietnamese, as well as programming languages) for which to collect metadata on potential data sources. We structure this effort by developing an online catalogue in English as a tool for gathering metadata through public hackathons. We present our tool and analyses of the resulting resource metadata, including distributions over languages, regions, and resource types, and discuss our lessons learned.
Named entity recognition is an important application within Danish NLP, essential within both industry and research. However, Danish NER is inhibited by a lack coverage across domains and entity types. As a consequence, no current models are capable of fine-grained named entity recognition, nor have they been evaluated for potential generalizability issues across datasets and domains. To alleviate these limitations, this paper introduces: 1) DANSK: a named entity dataset providing for high-granularity tagging as well as within-domain evaluation of models across a diverse set of domains; 2) and three generalizable models with fine-grained annotation available in DaCy 2.6.0; and 3) an evaluation of current state-of-the-art models’ ability to generalize across domains. The evaluation of existing and new models revealed notable performance discrepancies across domains, which should be addressed within the field. Shortcomings of the annotation quality of the dataset and its impact on model training and evaluation are also discussed. Despite these limitations, we advocate for the use of the new dataset DANSK alongside further work ongeneralizability within Danish NER.
Multiple-choice questions (MCQs) provide a widely used means of assessing reading comprehension. The automatic generation of such MCQs is a challenging language-technological problem that also has interesting educational applications. This article presents several methods for automatically producing reading comprehension questions MCQs from Swedish text. Unlike previous approaches, we construct models to generate the whole MCQ in one go, rather than using a pipeline architecture. Furthermore, we propose a two-stage method for evaluating the quality of the generated MCQs, first evaluating on carefully designed single-sentence texts, and then on texts from the SFI national exams. An extensive evaluation of the MCQ-generating capabilities of 12 different models, using this two-stage scheme, reveals that GPT-based models surpass smaller models that have been fine-tuned using small-scale datasets on this specific problem.
We are at a curious point in time where our ability to build language models (LMs) has outpaced our ability to analyze them. We do not really know how to reliably determine their capabilities, biases, dangers, knowledge, and so on. The benchmarks we have are often overly specific, do not generalize well, and are susceptible to data leakage. Recently, I have noticed a trend of using self-report studies, such as various polls and questionnaires originally designed for humans, to analyze the properties of LMs. I think that this approach can easily lead to false results, which can be quite dangerous considering the current discussions on AI safety, governance, and regulation. To illustrate my point, I will delve deeper into several papers that employ self-report methodologies and I will try to highlight some of their weaknesses.
Finetuning is a useful method for adapting Transformer-based text encoders to new tasks but can be computationally expensive for structured prediction tasks that require tuning at the token level. Furthermore, finetuning is inherently inefficient in updating all base model parameters, which prevents parameter sharing across tasks. To address these issues, we propose a method for efficient task adaptation of frozen Transformer encoders based on the local contribution of their intermediate layers to token representations. Our adapter uses a novel attention mechanism to aggregate intermediate layers and tailor the resulting representations to a target task. Experiments on several structured prediction tasks demonstrate that our method outperforms previous approaches, retaining over 99% of the finetuning performance at a fraction of the training cost. Our proposed method offers an efficient solution for adapting frozen Transformer encoders to new tasks, improving performance and enabling parameter sharing across different tasks.
Counterspeech offers direct rebuttals to hateful speech by challenging perpetrators of hate and showing support to targets of abuse. It provides a promising alternative to more contentious measures, such as content moderation and deplatforming, by contributing a greater amount of positive online speech rather than attempting to mitigate harmful content through removal. Advances in the development of large language models mean that the process of producing counterspeech could be made more efficient by automating its generation, which would enable large-scale online campaigns. However, we currently lack a systematic understanding of several important factors relating to the efficacy of counterspeech for hate mitigation, such as which types of counterspeech are most effective, what are the optimal conditions for implementation, and which specific effects of hate it can best ameliorate. This paper aims to fill this gap by systematically reviewing counterspeech research in the social sciences and comparing methodologies and findings with computer science efforts in automatic counterspeech generation. By taking this multi-disciplinary view, we identify promising future directions in both fields.
Emotions, which are responses to salient events, can be realized in text implicitly, for instance with mere references to facts (e.g., “That was the beginning of a long war”). Interpreting affective meanings thus relies on the readers’ background knowledge, but that is hardly modeled in computational emotion analysis. Much work in the field is focused on the word level and treats individual lexical units as the fundamental emotion cues in written communication. We shift our attention to word relations. We leverage Frame Semantics, a prominent theory for the description of predicate-argument structures, which matches the study of emotions: frames build on a “semantics of understanding” whose assumptions rely precisely on people’s world knowledge. Our overarching question is whether and to what extent the events that are represented by frames possess an emotion meaning. To carry out a large corpus-based correspondence analysis, we automatically annotate texts with emotions as well as with FrameNet frames and roles, and we analyze the correlations between them. Our main finding is that substantial groups of frames have an emotional import. With an extensive qualitative analysis, we show that they capture several properties of emotions that are purported by theories from psychology. These observations boost insights on the two strands of research that we bring together: emotion analysis can profit from the event-based perspective of frame semantics; in return, frame semantics gains a better grip of its position vis-à-vis emotions, an integral part of word meanings.
In this article we present the first dataset of multiple choice questions for assessing reading comprehension in Ukrainian. The dataset is based on the texts from the Ukrainian national tests for reading comprehension, and the MCQs themselves are created semi-automatically in three stages. The first stage was to use GPT-3 to generate the MCQs zero-shot, the second stage was to select MCQs of sufficient quality and revise the ones with minor errors, whereas the final stage was to expand the dataset with the MCQs written manually. The dataset is created by the Ukrainian language native speakers, one of whom is also a language teacher. The resulting corpus has slightly more than 900 MCQs, of which only 43 MCQs could be kept as they were generated by GPT-3.
This is a proposal for publishing resource papers as registered reports in the Northern European Journal of Language Technology. The idea is that authors write a data collection plan with a full data statement, to the extent that it can be written before data collection starts. Once the proposal is approved, publication of the final resource paper is guaranteed, as long as the data collection plan is followed (modulo reasonable changes due to unforeseen circumstances). This proposal changes the reviewing process from an antagonistic to a collaborative enterprise, and hopefully encourages NLP resources to develop and publish more high-quality datasets. The key advantage of this proposal is that it helps to promote responsible resource development (through constructive peer review) and to avoid research waste.
Emotions, which are responses to salient events, can be realized in text implicitly, for instance with mere references to facts (e.g., “That was the beginning of a long war”). Interpreting affective meanings thus relies on the readers’ background knowledge, but that is hardly modeled in computational emotion analysis. Much work in the field is focused on the word level and treats individual lexical units as the fundamental emotion cues in written communication. We shift our attention to word relations. We leverage Frame Semantics, a prominent theory for the description of predicate-argument structures, which matches the study of emotions: frames build on a “semantics of understanding” whose assumptions rely precisely on people’s world knowledge. Our overarching question is whether and to what extent the events that are represented by frames possess an emotion meaning. To carry out a large corpus-based correspondence analysis, we automatically annotate texts with emotions as well as with FrameNet frames and roles, and we analyze the correlations between them. Our main finding is that substantial groups of frames have an emotional import. With an extensive qualitative analysis, we show that they capture several properties of emotions that are purported by theories from psychology. These observations boost insights on the two strands of research that we bring together: emotion analysis can profit from the event-based perspective of frame semantics; in return, frame semantics gains a better grip of its position vis-a`-vis emotions, an integral part of word meanings.
Document clustering is frequently used in applications of natural language processing, e.g. to classify news articles or creating topic models. In this paper, we study document clustering with the common clustering pipeline that includes vectorization with BERT or Doc2Vec, dimension reduction with PCA or UMAP, and clustering with K-Means or HDBSCAN. We discuss the inter- actions of the different components in the pipeline, parameter settings, and how to determine an appropriate number of dimensions. The results suggest that BERT embeddings combined with UMAP dimension reduction to no less than 15 dimensions provides a good basis for clustering, regardless of the specific clustering algorithm used. Moreover, while UMAP performed better than PCA in our experiments, tuning the UMAP settings showed little impact on the overall performance. Hence, we recommend configuring UMAP so as to optimize its time efficiency. According to our topic model evaluation, the combination of BERT and UMAP, also used in BERTopic, performs best. A topic model based on this pipeline typically benefits from a large number of clusters.
People differ fundamentally in what motivates them to pursue a goal and how they approach it. For instance, some people seek growth and show eagerness, whereas others prefer security and are vigilant. The concept of regulatory focus is employed in psychology, to explain and predict this goal-directed behavior of humans underpinned by two unique motivational systems – the promotion and the prevention system. Traditionally, text analysis methods using closed-vocabularies are employed to assess the distinctive linguistic patterns associated with the two systems. From an NLP perspective, automatically detecting the regulatory focus of individuals from text provides valuable insights into the behavioral inclinations of the author, finding its applications in areas like marketing or health communication. However, the concept never made an impactful debut in computational linguistics research. To bridge this gap we introduce the novel task of regulatory focus classification from text and present two complementary German datasets – (1) experimentally generated event descriptions and (2) manually annotated short social media texts used for evaluating the generalizability of models on real-world data. First, we conduct a correlation analysis to verify if the linguistic footprints of regulatory focus reported in psychology studies are observable and to what extent in our datasets. For automatic classification, we compare closed-vocabulary-based analyses with a state-of-the-art BERT-based text classification model and observe that the latter outperforms lexicon-based approaches on experimental data and is notably better on out-of-domain Twitter data.
