COLING 2024 Papers

Backdoor attacks pose a critical security threat to natural language processing (NLP) models by establishing covert associations between trigger patterns and target labels without affecting normal accuracy. Existing attacks usually disregard fluency and semantic fidelity of poisoned text, rendering the malicious data easily detectable. However, text generation models can produce coherent and content-relevant text given prompts. Moreover, potential differences between human-written and AI-generated text may be captured by NLP models while being imperceptible to humans. More insidious threats could arise if attackers leverage latent features of AI-generated text as trigger patterns. We comprehensively investigate backdoor attacks on NLP models using AI-generated poisoned text obtained via continued writing or paraphrasing, exploring three attack scenarios: data, model and pre-training. For data poisoning, we fine-tune generators with attribute control to enhance the attack performance. For model poisoning, we leverage downstream tasks to derive specialized generators. For pre-training poisoning, we train multiple attribute-based generators and align their generated text with pre-defined vectors, enabling task-agnostic migration attacks. Experiments demonstrate that our method achieves effective attacks while maintaining fluency and semantic similarity across all scenarios. We hope this work can raise awareness of the security risks hidden in AI-generated text.

Medical dialogue generation (MDG) has gained increasing attention due to its substantial practical value. Previous works typically employ a sequence-to-sequence framework to generate medical responses by modeling dialogue context as sequential text with annotated medical entities. While these methods have been successful in generating fluent responses, they fail to provide process explanations of reasoning and require extensive entity annotation. To address these limitations, we propose the method Bootstrap Prompting for Explicit Reasoning in MDG (BP4ER), which explicitly model MDG’s multi-step reasoning process and iteratively enhance this reasoning process. We employ a least-to-most prompting strategy to guide a large language model (LLM) in explicit reasoning, breaking down MDG into simpler sub-questions. These sub-questions build on answers from previous ones. Additionally, we also introduce two distinct bootstrapping techniques for prompting, which autonomously correct errors and facilitate the LLM’s explicit reasoning. This approach eliminates the need for entity annotation and increases the transparency of the MDG process by explicitly generating the intermediate reasoning chain. Experimental results on the two publicly datasets show that BP4ER outperforms state-of-the-art methods across both objective and subjective evaluation.

We present a publicly available corpus of nineteenth-century newspaper text focused on crime in London, derived from the Gale British Library Newspapers corpus parts 1 and 2. The corpus comprises 600 newspaper excerpts and for each excerpt contains the original source image, the machine transcription of that image as found in the BLN and a gold standard manual transcription that we have created. We envisage the corpus will be helpful for the training and development of OCR and post-OCR correction methodologies for historical newspaper machine transcription—for which there is currently a dearth of publicly available resources. In this paper, we discuss the rationale behind gathering such a corpus, the methodology used to select, process, and align the data, and the corpus’ potential utility for historians and digital humanities researchers—particularly within the realms of neural machine translation-based post-OCR correction approaches, and other natural language processing tasks that are critically affected by erroneous OCR.

Event Extraction (EE) is a fundamental task in information extraction, aimed at identifying events and their associated arguments within textual data. It holds significant importance in various applications and serves as a catalyst for the development of related tasks. Despite the availability of numerous datasets and methods for event extraction in various languages, there has been a notable absence of a dedicated dataset for the Vietnamese language. To address this limitation, we propose BKEE, a novel event extraction dataset for Vietnamese. BKEE encompasses over 33 distinct event types and 28 different event argument roles, providing a labeled dataset for entity mentions, event mentions, and event arguments on 1066 documents. Additionally, we establish robust baselines for potential downstream tasks on this dataset, facilitating the analysis of challenges and future development prospects in the field of Vietnamese event extraction.

Recently, we have witnessed a significant performance boosting for dialogue response selection task achieved by Cross-Encoder based models. However, such models directly feed the concatenation of context and response into the pre-trained model for interactive inference, ignoring the comprehensively independent representation modeling of context and response. Moreover, randomly sampling negative responses from other dialogue contexts is simplistic, and the learned models have poor generalization capability in realistic scenarios. In this paper, we propose a response selection model called BERT-BC that combines the representation-based Bi-Encoder and interaction-based Cross-Encoder. Three contrastive learning methods are devised for the Bi-Encoder to align context and response to obtain the better semantic representation. Meanwhile, according to the alignment difficulty of context and response semantics, the harder samples are dynamically selected from the same batch with negligible cost and sent to Cross-Encoder to enhance the model’s interactive reasoning ability. Experimental results show that BERT-BC can achieve state-of-the-art performance on three benchmark datasets for multi-turn response selection.

The BEIR dataset is a large, heterogeneous benchmark for Information Retrieval (IR), garnering considerable attention within the research community. However, BEIR and analogous datasets are predominantly restricted to English language. Our objective is to establish extensive large-scale resources for IR in the Polish language, thereby advancing the research in this NLP area. In this work, inspired by mMARCO and Mr. TyDi datasets, we translated all accessible open IR datasets into Polish, and we introduced the BEIR-PL benchmark – a new benchmark which comprises 13 datasets, facilitating further development, training and evaluation of modern Polish language models for IR tasks. We executed an evaluation and comparison of numerous IR models on the newly introduced BEIR-PL benchmark. Furthermore, we publish pre-trained open IR models for Polish language, marking a pioneering development in this field. The BEIR-PL is included in MTEB Benchmark and also available with trained models at URL https://huggingface.co/clarin-knext.

Since the Internet is flooded with hate, it is one of the main tasks for NLP experts to master automated online content moderation. However, advancements in this field require improved access to publicly available accurate and non-synthetic datasets of social media content. For the Polish language, such resources are very limited. In this paper, we address this gap by presenting a new open dataset of offensive social media content for the Polish language. The dataset comprises content from Wykop.pl, a popular online service often referred to as the Polish Reddit, reported by users and banned in the internal moderation process. It contains a total of 691,662 posts and comments, evenly divided into two categories: harmful and neutral (non-harmful). The anonymized subset of the BAN-PL dataset consisting on 24,000 pieces (12,000 for each class), along with preprocessing scripts have been made publicly available. Furthermore the paper offers valuable insights into real-life content moderation processes and delves into an analysis of linguistic features and content characteristics of the dataset. Moreover, a comprehensive anonymization procedure has been meticulously described and applied. The prevalent biases encountered in similar datasets, including post-moderation and pre-selection biases, are also discussed.

Large language models (LLMs) have achieved dramatic proficiency over NLP tasks with normal length. Recently, multiple studies have committed to extending the context length and enhancing the long text modeling capabilities of LLMs. To comprehensively evaluate the long context ability of LLMs, we propose BAMBOO, a multi-task long context benchmark. BAMBOO has been designed with four principles: comprehensive capacity evaluation, avoidance of data contamination, accurate automatic evaluation, and different length levels. It consists of 10 datasets from 5 different long text understanding tasks, i.e., question answering, hallucination detection, text sorting, language modeling, and code completion, to cover various domains and core capacities of LLMs. We conduct experiments with five widely-used long-context models and further discuss five key questions for long text research. In the end, we discuss problems of current long-context models and point out future directions for enhancing long text modeling capacities. We release our data, prompts, and code at https://anonymous.4open.science/r/BAMBOO/.

The International Classification of Diseases (ICD) is an authoritative medical classification system of different diseases and conditions for clinical and management purposes. ICD indexing aims to assign a subset of ICD codes to a medical record. Since human coding is labour-intensive and error-prone, many studies employ machine learning techniques to automate the coding process. ICD coding is a challenging task, as it needs to assign multiple codes to each medical document from an extremely large hierarchically organized collection. In this paper, we propose a novel approach for ICD indexing that adopts three ideas: (1) we use a multi-level deep dilated residual convolution encoder to aggregate the information from the clinical notes and learn document representations across different lengths of the texts; (2) we formalize the task of ICD classification with auxiliary knowledge of the medical records, which incorporates not only the clinical texts but also different clinical code terminologies and drug prescriptions for better inferring the ICD codes; and (3) we introduce a graph convolutional network to leverage the co-occurrence patterns among ICD codes, aiming to enhance the quality of label representations. Experimental results show the proposed method achieves state-of-the-art performance on a number of measures.

Multimodal aspect-oriented sentiment classification (MABSC) task has garnered significant attention, which aims to identify the sentiment polarities of aspects by combining both language and vision information. However, the limited multimodal data in this task has become a big gap for the vision-language multimodal fusion. While large-scale vision-language pretrained models have been adapted to multiple tasks, their use for MABSC task is still in a nascent stage. In this work, we present an attempt to use the instruction tuning paradigm to MABSC task and leverage the ability of large vision-language models to alleviate the limitation in the fusion of textual and image modalities. To tackle the problem of potential irrelevance between aspects and images, we propose a plug-and-play selector to autonomously choose the most appropriate instruction from the instruction pool, thereby reducing the impact of irrelevant image noise on the final sentiment classification results. We conduct extensive experiments in various scenarios and our model achieves state-of-the-art performance on benchmark datasets, as well as in few-shot settings.

Practical and ethical dataset collection remains a challenge blocking many empirical methods in natural language processing, resulting in a lack of benchmarks or data on which to test hypotheses. We propose a solution to some of these areas by presenting a pipeline to reduce the research burden of producing image and text datasets when datasets may not exist. Our approach, with accompanying software tools, involves (1) generating text with LLMs; (2) creating accompanying image vignettes with text–to–image transformers; and (3) low-cost human validation. Based on existing literature that has struggled with quantitative evaluation (due to difficulty of data collection), we present the creation of 3 relevant datasets, and conduct a user study that demonstrates this approach is able to aid researchers in obtaining previously-challenging datasets. We provide sample data generated with this technique, the source code used to produce it, and discuss applicability and limitations.

In this position paper we present a methodology to automatically annotate French text for Cued Speech (CS), a communication system developed for people with hearing loss to complement speech reading at the phonetic level. This visual communication mode uses handshapes in different placements near the face in combination with the mouth movements (called ‘cues’ or ‘keys’) to make the phonemes of spoken language look different from each other. CS is used to acquire skills in lip reading, in oral communication and for reading. Despite many studies demonstrating its benefits, there are few resources available for learning and practicing it, especially in French. We thus propose a methodology to phonemize written corpora so that each word is aligned with the corresponding CS key(s). This methodology is proposed as part of a wider project aimed at creating an augmented reality system displaying a virtual coding hand where the user will be able to choose a text upon its complexity for cueing.

This paper describes different approaches for developing, for the first time, an automatic speech recognition system for two of the main dialects of Occitan, namely Gascon and Languedocian, and the results obtained in them. The difficulty of the task lies in the fact that Occitan is a less-resourced language. Although a great effort has been made to collect or create corpora of each variant (transcribed speech recordings for the acoustic models and two text corpora for the language models), the sizes of the corpora obtained are far from those of successful systems reported in the literature, and thus we have tested different techniques to compensate for the lack of resources. We have developed classical systems using Kaldi, creating an acoustic model for each variant and also creating language models from the collected corpora and from machine translated texts. We have also tried fine-tuning a Whisper model with our speech corpora. We report word error rates of 20.86 for Gascon and 13.52 for Languedocian with the Kaldi systems and 16.37 for Gascon and 11.74 for Languedocian with Whisper.

Word error rate (WER) is a metric used to evaluate the quality of transcriptions produced by Automatic Speech Recognition (ASR) systems. In many applications, it is of interest to estimate WER given a pair of a speech utterance and a transcript. Previous work on WER estimation focused on building models that are trained with a specific ASR system in mind (referred to as ASR system-dependent). These are also domain-dependent and inflexible in real-world applications. In this paper, a hypothesis generation method for ASR System-Independent WER estimation (SIWE) is proposed. In contrast to prior work, the WER estimators are trained using data that simulates ASR system output. Hypotheses are generated using phonetically similar or linguistically more likely alternative words. In WER estimation experiments, the proposed method reaches a similar performance to ASR system-dependent WER estimators on in-domain data and achieves state-of-the-art performance on out-of-domain data. On the out-of-domain data, the SIWE model outperformed the baseline estimators in root mean square error and Pearson correlation coefficient by relative 17.58% and 18.21%, respectively, on Switchboard and CALLHOME. The performance was further improved when the WER of the training set was close to the WER of the evaluation dataset.

Interruption detection is a new yet challenging task in the field of speech processing. This article presents a comprehensive study on automatic speech interruption detection, from the definition of this task, the assembly of a specialized corpus, and the development of an initial baseline system. We provide three main contributions: Firstly, we define the task, taking into account the nuanced nature of interruptions within spontaneous conversations. Secondly, we introduce a new corpus of conversational data, annotated for interruptions, to facilitate research in this domain. This corpus serves as a valuable resource for evaluating and advancing interruption detection techniques. Lastly, we present a first baseline system, which use speech processing methods to automatically identify interruptions in speech with promising results. In this article, we derivate from theoretical notions of interruption to build a simplification of this notion based on overlapped speech detection. Our findings can not only serve as a foundation for further research in the field but also provide a benchmark for assessing future advancements in automatic speech interruption detection.

With the widespread adoption of automatic transcription tools, acquiring speech transcriptions within seconds has become a reality. Nonetheless, many of these tools yield unpunctuated outputs, potentially incurring additional costs. This paper presents a novel approach to integrating punctuation into the transcriptions generated by such automatic tools, specifically focusing on Spanish-speaking contexts. Leveraging the RoBERTa-bne model pre-trained with data from the Spanish National Library, our training proposal is augmented with additional corpora to enhance performance on less common punctuation marks, such as question marks. Also, the proposed model has been trained through fine-tuning pre-trained models, involving adjustments for token classification and using SoftMax to identify the highest probability token. The proposed model obtains promising results when compared with other Spanish reference paper models. Ultimately, this model aims to facilitate punctuation on live transcriptions seamlessly and accurately. The proposed model will be applied to a real-case education project to improve the readability of the transcriptions.

Defining training, development and test set partitions for speech corpora is usually accomplished by hand. However, for the dataset under investigation, which contains a large number of speakers, eight different languages and code-switching between all the languages, this style of partitioning is not feasible. Therefore, we view the partitioning task as a resource allocation problem and propose to solve it automatically and optimally by the application of mixed-integer linear programming. Using this approach, we are able to partition a new 41.6-hour multilingual corpus of code-switched speech into training, development and testing partitions while maintaining a fixed number of speakers and a specific amount of code-switched speech in the development and test partitions. For this newly partitioned corpus, we present baseline speech recognition results using a state-of-the-art multilingual transformer model (Wav2Vec2-XLS-R) and show that the exclusion of very short utterances (<1s) results in substantially improved speech recognition performance.

We are concerned with mapping the discursive landscape of conspiracy narratives surrounding the COVID-19 pandemic. In the present study, we analyse a corpus of more than 1,000 German Telegram posts tagged with 14 fine-grained conspiracy narrative labels by three independent annotators. Since emerging narratives on social media are short-lived and notoriously hard to track, we experiment with different state-of-the-art approaches to few-shot and zero-shot text classification. We report performance in terms of ROC-AUC and in terms of optimal F1, and compare fine-tuned methods with off-the-shelf approaches and human performance.

Correctly inflecting determiners and adjectives so that they agree with the noun in nominal phrases (NPs) is a big challenge for learners of German. Given the increasing number of available learner corpora, a large-scale corpus-based study on the acquisition of this aspect of German morphosyntax would be desirable. In this paper, we present a pilot study in which we investigate how well nouns, their grammatical heads and the dependents that have to agree with the noun can be extracted automatically via dependency parsing. For six samples of the German learner corpus MERLIN (one per proficiency level), we found that in spite of many ungrammatical sentences in texts of low proficiency levels, human annotators find only few true ambiguities that would make the extraction of NPs and their heads infeasible. The automatic parsers, however, perform rather poorly on extracting the relevant elements for texts on CEFR levels A1-B1 (< 70%) but quite well from level B2 onwards ( 90%). We discuss the sources of errors and how performance could potentially be increased in the future.