COLING 2024 Papers

Despite recent advancements showcasing the impressive capabilities of Large Language Models (LLMs) in conversational systems, we show that even state-of-the-art LLMs are morally inconsistent in their generations, questioning their reliability (and trustworthiness in general). Prior works in LLM evaluation focus on developing ground-truth data to measure accuracy on specific tasks. However, for moral scenarios that often lack universally agreed-upon answers, consistency in model responses becomes crucial for their reliability. To address this issue, we propose an information-theoretic measure called Semantic Graph Entropy (SaGE), grounded in the concept of “Rules of Thumb” (RoTs) to measure a model’s moral consistency. RoTs are abstract principles learned by a model and can help explain their decision-making strategies effectively. To this extent, we construct the Moral Consistency Corpus (MCC), containing 50K moral questions, responses to them by LLMs, and the RoTs that these models followed. Furthermore, to illustrate the generalizability of SaGE, we use it to investigate LLM consistency on two popular datasets – TruthfulQA and HellaSwag. Our results reveal that task accuracy and consistency are independent problems, and there is a dire need to investigate these issues further.

This paper introduces the upgrade of a training corpus for linguistic annotation of modern standard Slovene. The enhancement spans both the size of the corpus and the depth of annotation layers. The revised SUK 1.0 corpus, building on its predecessor ssj500k 2.3, has doubled in size, containing over a million tokens. This expansion integrates three preexisting open-access datasets, all of which have undergone automatic tagging and meticulous manual review across multiple annotation layers, each represented in varying proportions. These layers span tokenization, segmentation, lemmatization, MULTEXT-East morphology, Universal Dependencies, JOS-SYN syntax, semantic role labeling, named entity recognition, and the newly incorporated coreferences. The paper illustrates the annotation processes for each layer while also presenting the results of the new CLASSLA-Stanza annotation tool, trained on the SUK corpus data. As one of the fundamental language resources of modern Slovene, the SUK corpus calls for constant development, as outlined in the concluding section.

The rise of social media platforms has led to an increase in polarised online discussions, especially on political and socio-cultural topics such as elections and climate change. We propose a simple and entirely novel unsupervised method to better predict whether the authors of two posts agree or disagree, leveraging user stances about named entities obtained from their posts. We present STEntConv, a model which builds a graph of users and named entities weighted by stance and trains a Signed Graph Convolutional Network (SGCN) to detect disagreement between comment and reply posts. We run experiments and ablation studies and show that including this information improves disagreement detection performance on a dataset of Reddit posts for a range of controversial subreddit topics, without the need for platform-specific features or user history

Pre-trained language models (PLMs) are widely used for various tasks, but fine-tuning them requires sufficient data. Data augmentation approaches have been proposed as alternatives, but they vary in complexity, cost, and performance. To address these challenges, we propose STAGE (Simple Text Data Augmentation by Graph Exploration), a highly effective method for data augmentation. STAGE utilizes simple modification operations such as insertion, deletion, replacement, and swap. However, what distinguishes STAGE lies in the selection of optimal words for each modification. This is achieved by leveraging a word-relation graph called the co-graph. The co-graph takes into account both word frequency and co-occurrence, providing valuable information for operand selection. To assess the performance of STAGE, we conduct evaluations using seven representative datasets and three different PLMs. Our results demonstrate the effectiveness of STAGE across diverse data domains, varying data sizes, and different PLMs. Also, STAGE demonstrates superior performance when compared to previous methods that use simple modification operations or large language models like GPT3.

Test-time adaptation (TTA) aims to adapt the neural network to the distribution of the target domain using only unlabeled test data. Most previous TTA methods have achieved success under mild conditions, such as considering only a single or multiple independent static domains. However, in real-world settings, the test data is sampled in a correlated manner and the test environments undergo continual changes over time, which may cause previous TTA methods to fail in practical noise scenarios, i.e., independent noise distribution shifts, continual noise distribution shifts, and continual mixed distribution shifts. To address these issues, we elaborate a Stable Test-time Adaptation Framework, called STAF, to stabilize the adaptation process. Specifically, to boost model robustness to noise distribution shifts, we present a multi-stream perturbation consistency method, enabling weak-to-strong views to be consistent, guided by the weak view from the original sample. Meanwhile, we develop a reliable memory-based corrector which utilizes reliable snapshots between the anchor model and the adapt model to correct prediction bias. Furthermore, we propose a dynamic parameter restoration strategy to alleviate error accumulation and catastrophic forgetting that takes into account both the distribution shift and sample adaptation degree. Extensive experiments demonstrate the robustness and effectiveness of STAF, which pushes the boundaries of test-time adaptation to more realistic scenarios and paves the way for stable deployment of real-world applications.

Linguistic conventions that arise in dialogue reflect common ground and can increase communicative efficiency. Social robots that can understand these conventions and the process by which they arise have the potential to become efficient communication partners. Nevertheless, it is unclear how robots can engage in convention formation when presented with both familiar and new information. We introduce an adaptable game platform, SPOTTER, to study the dynamics of convention formation for visually grounded referring expressions in both human-human and human-robot interaction. Specifically, we seek to elicit convention forming for members of an inner circle of well-known individuals in the common ground, as opposed to individuals from an outer circle, who are unfamiliar. We release an initial corpus of 5000 utterances from two exploratory pilot experiments in Dutch. Different from previous work focussing on human-human interaction, we find that referring expressions for both familiar and unfamiliar individuals maintain their length throughout human-robot interaction. Stable conventions are formed, although these conventions can be impacted by distracting outer circle individuals. With our distinction between familiar and unfamiliar, we create a contrastive operationalization of common ground, which aids research into convention formation.

Singleton mentions, i.e. entities mentioned only once in a text, are important to how humans understand discourse from a theoretical perspective. However previous attempts to incorporate their detection in end-to-end neural coreference resolution for English have been hampered by the lack of singleton mention spans in the OntoNotes benchmark. This paper addresses this limitation by combining predicted mentions from existing nested NER systems and features derived from OntoNotes syntax trees. With this approach, we create a near approximation of the OntoNotes dataset with all singleton mentions, achieving ~94% recall on a sample of gold singletons. We then propose a two-step neural mention and coreference resolution system, named SPLICE, and compare its performance to the end-to-end approach in two scenarios: the OntoNotes test set and the out-of-domain (OOD) OntoGUM corpus. Results indicate that reconstructed singleton training yields results comparable to end-to-end systems for OntoNotes, while improving OOD stability (+1.1 avg. F1). We conduct error analysis for mention detection and delve into its impact on coreference clustering, revealing that precision improvements deliver more substantial benefits than increases in recall for resolving coreference chains.

The proliferation of news media outlets has increased the demand for intelligent systems capable of detecting redundant information in news articles in order to enhance user experience. However, the heterogeneous nature of news can lead to spurious findings in these systems: Simple heuristics such as whether a pair of news are both about politics can provide strong but deceptive downstream performance. Segmenting news similarity datasets into topics improves the training of these models by forcing them to learn how to distinguish salient characteristics under more narrow domains. However, this requires the existence of topic-specific datasets, which are currently lacking. In this article, we propose a novel dataset of similar news, SPICED, which includes seven topics: Crime & Law, Culture & Entertainment, Disasters & Accidents, Economy & Business, Politics & Conflicts, Science & Technology, and Sports. Futhermore, we present four different levels of complexity, specifically designed for news similarity detection task. We benchmarked the created datasets using MinHash, BERT, SBERT, and SimCSE models.

Detecting salient parts in text using natural language processing has been widely used to mitigate the effects of information overflow. Nevertheless, most of the datasets available for this task are derived mainly from academic publications. We introduce SPACE-IDEAS, a dataset for salient information detection from innovation ideas related to the Space domain. The text in SPACE-IDEAS varies greatly and includes informal, technical, academic and business-oriented writing styles. In addition to a manually annotated dataset we release an extended version that is annotated using a large generative language model. We train different sentence and sequential sentence classifiers, and show that the automatically annotated dataset can be leveraged using multitask learning to train better classifiers.

Sharing textual content in the form of public posts on online platforms remains a significant part of the social web. Research on stylometric profiling suggests that despite users’ discreetness, and even under the guise of anonymity, the content and style of such posts may still reveal detailed author information. Studying how this might be inferred and obscured is relevant not only to the domain of cybersecurity, but also to those studying bias of classifiers drawing features from web corpora. While the collection of gold standard data is expensive, prior work shows that distant labels (i.e., those gathered via heuristics) offer an effective alternative. Currently, however, pre-existing corpora are limited in scope (e.g., variety of attributes and size). We present the SOBR corpus: 235M Reddit posts for which we used subreddits, flairs, and self-reports as distant labels for author attributes (age, gender, nationality, personality, and political leaning). In addition to detailing the data collection pipeline and sampling strategy, we report corpus statistics and provide a discussion on the various tasks and research avenues to be pursued using this resource. Along with the raw corpus, we provide sampled splits of the data, and suggest baselines for stylometric profiling. We close our work with a detailed set of ethical considerations relevant to the proposed lines of research.

This article introduces SM-FEEL-BG – the first Bulgarian-language package, containing 6 datasets with Social Media (SM) texts with emotion, feeling, and sentiment labels and 4 classifiers trained on them. All but one dataset from these are freely accessible for research purposes. The largest dataset contains 6000 Twitter, Telegram, and Facebook texts, manually annotated with 21 fine-grained emotion/feeling categories. The fine-grained labels are automatically merged into three coarse-grained sentiment categories, producing a dataset with two parallel sets of labels. Several classification experiments are run on different subsets of the fine-grained categories and their respective sentiment labels with a Bulgarian fine-tuned BERT. The highest Acc. reached was 0.61 for 16 emotions and 0.70 for 11 emotions (incl. 310 ChatGPT 4-generated texts). The sentiments Acc. of the 11 emotions dataset was also the highest (0.79). As Facebook posts cannot be shared, we ran experiments on the Twitter and Telegram subset of the 11 emotions dataset, obtaining 0.73 Acc. for emotions and 0.80 for sentiments. The article describes the annotation procedures, guidelines, experiments, and results. We believe that this package will be of significant benefit to researchers working on emotion detection and sentiment analysis in Bulgarian.

Identifying early markers of Alzheimer’s disease (AD) trajectory enables intervention in early disease stages when our currently-available interventions are most likely to be beneficial. Research has shown that alterations in speech, as well as linguistic and semantic deviations in spontaneous conversation detected using natural language processing, manifest early in AD prior to some other observed cognitive deficits. Recent studies show that cerebrospinal fluid (CSF) levels serve as useful early biomarkers for identifying early AD, but CSF biomarkers are challenging to collect. A simpler alternative that has seen very rapid development is based on the use of plasma biomarkers as a blood draw is minimally invasive. Associating verbal and nonverbal characteristics from speech data with CSF and plasma biomarkers may open the door to less invasive, more efficient methods for early AD detection. We present SLaCAD, a new dataset to facilitate this process. We describe our data collection procedures, analyze the resulting corpus, and present preliminary findings that relate measures extracted from the audio and transcribed text to clinical diagnoses, CSF levels, and plasma biomarkers. Our findings demonstrate the feasibility of this and indicate that the collected data can be used to improve assessments of early AD.

Many utterances convey meanings that go beyond the literal meaning of a sentence. One class of such meanings is scalar implicatures, a phenomenon by which a speaker conveys the negation of a more informative utterance by producing a less informative utterance. This paper introduces a Natural Language Inference (NLI) dataset designed to investigate the ability of language models to interpret utterances with scalar implicatures. Our dataset is comprised of text extracted from the C4 English text corpus and annotated with both crowd-sourced and expert annotations. We evaluate NLI models based on DeBERTa to investigate 1) whether NLI models can learn to predict pragmatic inferences involving gradable adjectives and 2) whether models generalize to utterances involving unseen adjectives. We find that fine-tuning NLI models on our dataset significantly improves their performance to derive scalar implicatures, both for in-domain and for out-of domain examples. At the same time, we find that the investigated models still perform considerably worse on examples with scalar implicatures than on other types of NLI examples, highlighting that pragmatic inferences still pose challenges for current models.

Natural language inference (NLI) is an important language understanding benchmark. Two deficiencies of this benchmark are: i) most existing NLI datasets exist for English and a few other well-resourced languages, and ii) most NLI datasets are formed with a narrow set of annotators’ instructions, allowing the prediction models to capture linguistic clues instead of measuring true reasoning capability. We address both issues and introduce SI-NLI, the first dataset for Slovene natural language inference. The dataset is constructed from scratch using knowledgeable annotators with carefully crafted guidelines aiming to avoid commonly encountered problems in existing NLI datasets. We also manually translate the SI-NLI to English to enable cross-lingual model training and evaluation. Using the newly created dataset and its translation, we train and evaluate a variety of large transformer language models in a monolingual and cross-lingual setting. The results indicate that larger models, in general, achieve better performance. The qualitative analysis shows that the SI-NLI dataset is diverse and that there remains plenty of room for improvement even for the largest models.

We tackle Paragraph-level Question Generation (abbr., PQG) in this paper. PQG is a task of automatically generating questions given paragraphs and answers. Identifying the relevant sentences to answers is crucial for reasoning the possible questions before generation. Accordingly, we propose a salience-guided approach to enhance PQG. Specifically, we construct an auxiliary task of identifying salient sentences that manifest relevance. Grounded on this auxiliary task and the main task of PQG, we strengthen the BART encoder during training within a multitask learning framework. In particular, we utilize the identified salient sentences as an explicit guidance to enable the salience-aware attention computation in the BART decoder. We experiment on the benchmark dataset FairytaleQA. The test results show that our approach yields substantial improvements compared to the BART baseline, achieving the Rouge-L, BLEU4, BERTScore, Q-BLUE-3 and F1-scores of about 56.56%, 19.78%, 61.19%, 54.33% and 43.55%, respectively. Both the source codes and models will be publicly available.

Ensuring universal access to written content, regardless of users’ language proficiency and cognitive abilities, is of paramount importance. Sentence simplification, which involves converting complex sentences into more accessible forms while preserving their meaning, plays a crucial role in enhancing text accessibility. This paper introduces SENTA, a system for sentence simplification in Slovene. The system consists of two components. First, a neural classifier identifies sentences that require simplification, and second, a large Slovene language model based on T5 architecture is fine-tuned to transform complex texts into a simpler form, achieving an excellent SARI score of 41. Both automatic and qualitative evaluations provide important insights into the problem, highlighting areas for future research in multilingual applications, and fluency maintenance. Finally, SENTA is integrated into a freely accessible, user-friendly user interface, offering a valuable service to less-fluent Slovene users.

Contrastive learning has recently achieved compelling performance in unsupervised sentence representation. As an essential element, data augmentation protocols, however, have not been well explored. The pioneering work SimCSE resorting to a simple dropout mechanism (viewed as continuous augmentation) surprisingly dominates discrete augmentations such as cropping, word deletion, and synonym replacement as reported. To understand the underlying rationales, we revisit existing approaches and attempt to hypothesize the desiderata of reasonable data augmentation methods: balance of semantic consistency and expression diversity. We then develop three simple yet effective discrete sentence augmentation schemes: punctuation insertion, modal verbs, and double negation. They act as minimal noises at lexical level to produce diverse forms of sentences. Furthermore, standard negation is capitalized on to generate negative samples for alleviating feature suppression involved in contrastive learning. We experimented extensively with semantic textual similarity on diverse datasets. The results support the superiority of the proposed methods consistently. Our key code is available at https://github.com/Zhudongsheng75/SDA

We introduce the Situated Corpus Of Understanding Transactions (SCOUT), a multi-modal collection of human-robot dialogue in the task domain of collaborative exploration. The corpus was constructed from multiple Wizard-of-Oz experiments where human participants gave verbal instructions to a remotely-located robot to move and gather information about its surroundings. SCOUT contains 89,056 utterances and 310,095 words from 278 dialogues averaging 320 utterances per dialogue. The dialogues are aligned with the multi-modal data streams available during the experiments: 5,785 images and 30 maps. The corpus has been annotated with Abstract Meaning Representation and Dialogue-AMR to identify the speaker’s intent and meaning within an utterance, and with Transactional Units and Relations to track relationships between utterances to reveal patterns of the Dialogue Structure. We describe how the corpus and its annotations have been used to develop autonomous human-robot systems and enable research in open questions of how humans speak to robots. We release this corpus to accelerate progress in autonomous, situated, human-robot dialogue, especially in the context of navigation tasks where details about the environment need to be discovered.

Large language models achieve impressive results by inferring conditional probability distributions in the context of user input to generate responses. However, they still have the following limitations in practical applications: 1) User queries are often colloquial and do not conform to the conditional probability distribution of LLM. 2) Unsupervised generation and recall of in-context examples(compared to random sampling) remains an open problem. To alleviate the above problems, we propose a novel Self-calibration, Self-recall and Self-aggregation prompt pipeline (S 3Prompt). Specifically, we first design a question calibration prompt to align colloquial queries with LLM context. Secondly, we construct a candidate recall prompt that allows LLM to generate potential background information, which is different from traditional retrieval-based QA. Finally, we design an information aggregation prompt to generate the final answer by aggregating the recalled information. Notably, we find that the self-generated information by LLM has a smaller gap when fused with LLM. We conducted comprehensive experiments on various datasets, including numerical reasoning, common sense reasoning, logical reasoning, and reading comprehension. The results showed that the performance of LLM can be significantly improved by using question calibration, candidate recall, and information aggregation, without requiring annotated datasets and model parameter updates.