In this paper, we introduce a new beam search algorithm that improves the generalization of neural generators to unseen examples, especially in low-resource data-to-text settings. Our algorithm aims to reduce the number of omissions and hallucinations during the decoding process. For this purpose, it relies on two regression models to explicitly characterize factual errors. We explain how to create a new dataset to train these models given an original training set of less than a thousand data points. We apply our approach in the low-resource, legal setting using the French Plum2Text dataset, as well as in English using WebNLG. We observe in our experiment that this combination improves the faithfulness of pre-trained neural text generators using both human and automatic evaluation. Moreover, our approach offers a level of interpretability by predicting the number of omissions and hallucinations present in a given generation with respect to the input data. Finally, we visualize our algorithm’s exploration of the hypothesis space at different steps during the decoding process.

Multiple business scenarios require an automated generation of descriptive human-readable text from structured input data. This has resulted into substantial work on fact-to-text generation systems recently. Unfortunately, previous work on fact-to-text (F2T) generation has focused primarily on English mainly due to the high availability of relevant datasets. Only recently, the problem of cross-lingual fact-to-text (XF2T) was proposed for generation across multiple languages alongwith a dataset, XAlign for eight languages. However, there has been no rigorous work on the actual XF2T generation problem. We extend XAlign dataset with annotated data for four more languages: Punjabi, Malayalam, Assamese and Oriya. We conduct an extensive study using popular Transformer-based text generation models on our extended multi-lingual dataset, which we call XAlignV2. Further, we investigate the performance of different text generation strategies: multiple variations of pretraining, fact-aware embeddings and structure-aware input encoding. Our extensive experiments show that a multi-lingual mT5 model which uses fact-aware embeddings with structure-aware input encoding leads to best results (30.90 BLEU, 55.12 METEOR and 59.17 chrF++) across the twelve languages. We make our code, dataset and model publicly available, and hope that this will help advance further research in this critical area.

Studying data memorization in neural language models helps us understand the risks (e.g., to privacy or copyright) associated with models regurgitating training data and aids in the development of countermeasures. Many prior works—and some recently deployed defenses—focus on “verbatim memorization”, defined as a model generation that exactly matches a substring from the training set. We argue that verbatim memorization definitions are too restrictive and fail to capture more subtle forms of memorization. Specifically, we design and implement an efficient defense that _perfectly_ prevents all verbatim memorization. And yet, we demonstrate that this “perfect” filter does not prevent the leakage of training data. Indeed, it is easily circumvented by plausible and minimally modified “style-transfer” prompts—and in some cases even the non-modified original prompts—to extract memorized information. We conclude by discussing potential alternative definitions and why defining memorization is a difficult yet crucial open question for neural language models.

While GPT-3 has garnered significant attention for its capabilities in natural language generation, research on its use outside of English is still relatively limited. We focus on how GPT-3 can be fine-tuned for generating synthetic news articles in a low-resource language, namely Danish. The model’s performance is evaluated on the dimensions of human and machine detection in two separate experiments. When presented with either a real or GPT-3 generated news article, human participants achieve a 58.1% classification accuracy. Contrarily, a fine-tuned BERT classifier obtains a 92.7% accuracy on the same task. This discrepancy likely pertains to the fine-tuned GPT-3 model oversampling high-likelihood tokens in its text generation. Although this is undetectable to the human eye, it leaves a statistical discrepancy for machine classifiers to detect. We address how decisions in the experimental design favoured the machine classifiers over the human evaluators, and whether the produced synthetic articles are applicable in a real-world context.

In this work, we investigate Data Augmentation methods to improve the performance of state-of-the-art models for four different downstream tasks. Specifically, we propose Generative Adversarial Network using Language Models (GAN-LM) approach that combines a deep generative model with a pre-trained language model to produce diverse augmentations. We compare the GAN-LM to various conventional methods in non-contextual- and contextual-levels on four public datasets: ZESHEL for zero-shot entity linking, TREC for question classification, STS-B for sentence pairs semantic textual similarity (STS), and mSTS for multilingual sentence pairs STS. Additionally, we subsample these datasets to study the impact of such augmentations in low-resource settings where limited amounts of training data is available. Compared to the state-of-the-art methods in downstream tasks, we mostly achieve the best performance using GAN-LM approach. Finally, we investigate the way of combining the GAN-LM with other augmentation methods to complement our proposed approach. The developed code for reproducibility is included in the supplementary material.

Good figure captions help paper readers understand complex scientific figures. Unfortunately, even published papers often have poorly written captions. Automatic caption generation could aid paper writers by providing good starting captions that can be refined for better quality. Prior work often treated figure caption generation as a vision-to-language task. In this paper, we show that it can be more effectively tackled as a text summarization task in scientific documents. We fine-tuned PEGASUS, a pre-trained abstractive summarization model, to specifically summarize figure-referencing paragraphs (e.g., “Figure 3 shows...”) into figure captions. Experiments on large-scale arXiv figures show that our method outperforms prior vision methods in both automatic and human evaluations. We further conducted an in-depth investigation focused on two key challenges: (i) the common presence of low-quality author-written captions and (ii) the lack of clear standards for good captions. Our code and data are available at: https://github.com/Crowd-AI-Lab/Generating-Figure-Captions-as-a-Text-Summarization-Task.

In recent years, many NLP studies have focused solely on performance improvement. In this work, we focus on the linguistic and scientific aspects of NLP. We use the task of generating referring expressions in context (REG-in-context) as a case study and start our analysis from GREC, a comprehensive set of shared tasks in English that addressed this topic over a decade ago. We ask what the performance of models would be if we assessed them (1) on more realistic datasets, and (2) using more advanced methods. We test the models using different evaluation metrics and feature selection experiments. We conclude that GREC can no longer be regarded as offering a reliable assessment of models’ ability to mimic human reference production, because the results are highly impacted by the choice of corpus and evaluation metrics. Our results also suggest that pre-trained language models are less dependent on the choice of corpus than classic Machine Learning models, and therefore make more robust class predictions.

Knowledge Graph (KG)-to-Text generation aims at generating fluent natural-language text that accurately represents the information of a given knowledge graph. While significant progress has been made in this task by exploiting the power of pre-trained language models (PLMs) with appropriate graph structure-aware modules, existing models still fall short of generating faithful text, especially when the ground-truth natural-language text contains additional information that is not present in the graph. In this paper, we develop a KG-to-text generation model that can generate faithful natural-language text from a given graph, in the presence of noisy reference text. Our framework incorporates two core ideas: Firstly, we utilize contrastive learning to enhance the model’s ability to differentiate between faithful and hallucinated information in the text, thereby encouraging the decoder to generate text that aligns with the input graph. Secondly, we empower the decoder to control the level of hallucination in the generated text by employing a controllable text generation technique. We evaluate our model’s performance through the standard quantitative metrics as well as a ChatGPT-based quantitative and qualitative analysis. Our evaluation demonstrates the superior performance of our model over state-of-the-art KG-to-text models on faithfulness.

Research in Multi-document Summarization (MDS) mostly focuses on the English language and depends on large MDS datasets that are not available for other languages. Some of these approaches concatenate the source documents, resulting in overlong model inputs. Existing transformer architectures are unable to process such long inputs entirely, omitting documents in the summarization process. Other solutions address this issue by implementing multi-stage approaches that also require changes in the model architecture. In this paper, we introduce various sampling approaches based on information entropy that allow us to perform MDS in a single stage. These approaches also consider all source documents without using MDS training data nor changing the model’s architecture. Besides, we build a MDS test set of German news articles to assess the performance of our methods on abstractive multi-document summaries. Experimental results show that our entropy-based approaches outperform previous state-of-the-art on German MDS, while still remaining primarily abstractive. We release our code and MDS test set to encourage further research in German abstractive MDS.

An optimal delivery of arguments is key to persuasion in any debate, both for humans and for AI systems. This requires the use of clear and fluent claims relevant to the given debate. Prior work has studied the automatic assessment of argument quality extensively. Yet, no approach actually improves the quality so far. To fill this gap, this paper proposes the task of claim optimization: to rewrite argumentative claims in order to optimize their delivery. As multiple types of optimization are possible, we approach this task by first generating a diverse set of candidate claims using a large language model, such as BART, taking into account contextual information. Then, the best candidate is selected using various quality metrics. In automatic and human evaluation on an English-language corpus, our quality-based candidate selection outperforms several baselines, improving 60% of all claims (worsening 16% only). Follow-up analyses reveal that, beyond copy editing, our approach often specifies claims with details, whereas it adds less evidence than humans do. Moreover, its capabilities generalize well to other domains, such as instructional texts.

Large Language Models, and ChatGPT in particular, have recently grabbed the attention of the community and the media. Having reached high language proficiency, attention has been shifting toward its reasoning capabilities. In this paper, our main aim is to evaluate ChatGPT’s question generation in a task where language production should be driven by an implicit reasoning process. To this end, we employ the 20-Questions game, traditionally used within the Cognitive Science community to inspect the information seeking-strategy’s development. This task requires a series of interconnected skills: asking informative questions, stepwise updating the hypothesis space, and stopping asking questions when enough information has been collected. We build hierarchical hypothesis spaces, exploiting feature norms collected from humans vs. ChatGPT itself, and we inspect the efficiency and informativeness of ChatGPT’s strategy. Our results show that ChatGPT’s performance gets closer to an optimal agent only when prompted to explicitly list the updated space stepwise.

Large language models underestimate the impact of negations on how much they change the meaning of a sentence. Therefore, learned evaluation metrics based on these models are insensitive to negations. In this paper, we propose NegBLEURT, a negation-aware version of the BLEURT evaluation metric. For that, we designed a rule-based sentence negation tool and used it to create the CANNOT negation evaluation dataset. Based on this dataset, we fine-tuned a sentence transformer and an evaluation metric to improve their negation sensitivity. Evaluating these models on existing benchmarks shows that our fine-tuned models outperform existing metrics on the negated sentences by far while preserving their base models’ performances on other perturbations.

Automatically summarizing radiology reports into a concise impression can reduce the manual burden of clinicians and improve the consistency of reporting. Previous work aimed to enhance content selection and factuality through guided abstractive summarization. However, two key issues persist. First, current methods heavily rely on domain-specific resources to extract the guidance signal, limiting their transferability to domains and languages where those resources are unavailable. Second, while automatic metrics like ROUGE show progress, we lack a good understanding of the errors and failure modes in this task. To bridge these gaps, we first propose a domain-agnostic guidance signal in form of variable-length extractive summaries. Our empirical results on two English benchmarks demonstrate that this guidance signal improves upon unguided summarization while being competitive with domain-specific methods. Additionally, we run an expert evaluation of four systems according to a taxonomy of 11 fine-grained errors. We find that the most pressing differences between automatic summaries and those of radiologists relate to content selection including omissions (up to 52%) and additions (up to 57%). We hypothesize that latent reporting factors and corpus-level inconsistencies may limit models to reliably learn content selection from the available data, presenting promising directions for future work.

Prior art investigating task-oriented dialog and automatic generation of such dialogs have focused on single-user dialogs between a single user and an agent. However, there is limited study on adapting such AI agents to multi-user conversations (involving multiple users and an agent). Multi-user conversations are richer than single-user conversations containing social banter and collaborative decision making. The most significant challenge impeding such studies is the lack of suitable multi-user task-oriented dialogs with annotations of user belief states and system actions. One potential solution is multi-user dialog generation from single-user data. Many single-user dialogs datasets already contain dialog state information (intents, slots), thus making them suitable candidates. In this work, we propose a novel approach for expanding single-user task-oriented dialogs (e.g. MultiWOZ) to multi-user dialogs in a zero-shot setting.

Recent studies have used human continuations of Implicit Causality (IC) prompts collected in linguistic experiments to evaluate discourse understanding in large language models (LLMs), focusing on the well-known IC coreference bias in the LLMs’ predictions of the next word following the prompt. In this study, we investigate how continuations of IC prompts can be used to evaluate the text generation capabilities of LLMs in a linguistically controlled setting. We conduct an experiment using two open-source GPT-based models, employing human evaluation to assess different aspects of continuation quality. Our findings show that LLMs struggle in particular with generating coherent continuations in this rather simple setting, indicating a lack of discourse knowledge beyond the well-known IC bias. Our results also suggest that a bias congruent continuation does not necessarily equate to a higher continuation quality. Furthermore, our study draws upon insights from the Uniform Information Density hypothesis, testing different prompt modifications and decoding procedures and showing that sampling-based methods are particularly sensitive to the information density of the prompts.

Neural data-to-text systems lack the control and factual accuracy required to generate useful and insightful summaries of multidimensional data. We propose a solution in the form of data views, where each view describes an entity and its attributes along specific dimensions. A sequence of views can then be used as a high-level schema for document planning, with the neural model handling the complexities of micro-planning and surface realization. We show that our view-based system retains factual accuracy while offering high-level control of output that can be tailored based on user preference or other norms within the domain.

In this paper, we present a system for augmenting virtual AI characters with long-term memory, enabling them to remember facts about themselves, their world, and past experiences. We propose a memory-creation pipeline that converts raw text into condensed memories and a memory-retrieval system that utilizes these memories to generate character responses. Using a fact-checking pipeline based on GPT-4, our evaluation demonstrates that the character responses are grounded in the retrieved memories and maintain factual accuracy. We discuss the implications of our system for creating engaging and consistent virtual characters and highlight areas for future research, including large language model (LLM) guardrailing and virtual character personality development.

In-context learning (ICL) for large language models has proven to be a powerful approach for many natural language processing tasks. However, determining the best method to select examples for ICL is nontrivial as the results can vary greatly depending on the quality, quantity, and order of examples used. In this paper, we conduct a case study on text simplification (TS) to investigate how to select the best and most robust examples for ICL. We propose Metric-Based in-context Learning (MBL) method that utilizes commonly used TS metrics such as SARI, compression ratio, and BERT-Precision for selection. Through an extensive set of experiments with various-sized GPT models on standard TS benchmarks such as TurkCorpus and ASSET, we show that examples selected by the top SARI scores perform the best on larger models such as GPT-175B, while the compression ratio generally performs better on smaller models such as GPT-13B and GPT-6.7B. Furthermore, we demonstrate that MBL is generally robust to example orderings and out-of-domain test sets, and outperforms strong baselines and state-of-the-art finetuned language models. Finally, we show that the behavior of large GPT models can be implicitly controlled by the chosen metric. Our research provides a new framework for selecting examples in ICL, and demonstrates its effectiveness in text simplification tasks, breaking new ground for more accurate and efficient NLG systems.

Language-capable robots must be able to efficiently and naturally communicate about objects in the environment. A key part of communication is Referring Form Selection (RFS): the process of selecting a form like it, that, or the N to use when referring to an object. Recent cognitive status-informed computational RFS models have been evaluated in terms of goodness-of-fit to human data. But it is as yet unclear whether these models actually select referring forms that are any more natural than baseline alternatives, regardless of goodness-of-fit. Through a human subject study designed to assess this question, we show that even though cognitive status-informed referring selection models achieve good fit to human data, they do not (yet) produce concrete benefits in terms of naturality. On the other hand, our results show that human utterances also had high variability in perceived naturality, demonstrating the challenges of evaluating RFS naturality.

In this work, we study dialogue scenarios that start from chit-chat but eventually switch to task-related services, and investigate how a unified dialogue model, which can engage in both chit-chat and task-oriented dialogues, takes the initiative during the dialogue mode transition from chit-chat to task-oriented in a coherent and cooperative manner. We firstly build a transition info extractor (TIE) that keeps track of the preceding chit-chat interaction and detects the potential user intention to switch to a task-oriented service. Meanwhile, in the unified model, a transition sentence generator (TSG) is extended through efficient Adapter tuning and transition prompt learning. When the TIE successfully finds task-related information from the preceding chit-chat, such as a transition domain (“train” in Figure fig: system-initiated transition from chit-chat to task-oriented.), then the TSG is activated automatically in the unified model to initiate this transition by generating a transition sentence under the guidance of transition information extracted by TIE. The experimental results show promising performance regarding the proactive transitions. We achieve an additional large improvement on TIE model by utilizing Conditional Random Fields (CRF). The TSG can flexibly generate transition sentences while maintaining the unified capabilities of normal chit-chat and task-oriented response generation.

Current captioning datasets focus on object-centric captions, describing the visible objects in the image, often ending up stating the obvious (for humans), e.g. “people eating food in a park”. Although these datasets are useful to evaluate the ability of Vision & Language models to recognize and describe visual content, they do not support controlled experiments involving model testing or fine-tuning, with more high-level captions, which humans find easy and natural to produce. For example, people often describe images based on the type of scene they depict (“people at a holiday resort”) and the actions they perform (“people having a picnic”). Such concepts are based on personal experience and contribute to forming common sense assumptions. We present the High-Level Dataset, a dataset extending 14997 images from the COCO dataset, aligned with a new set of 134,973 human-annotated (high-level) captions collected along three axes: scenes, actions and rationales. We further extend this dataset with confidence scores collected from an independent set of readers, as well as a set of narrative captions generated synthetically, by combining each of the three axes. We describe this dataset and analyse it extensively. We also present baseline results for the High-Level Captioning task.

In data2text generation, tabular data is transformed into a text that expresses information from that source domain. While some text types, such as instructions, demand objective and neutral language without any expressive and evaluative content, many other text types are expected to provide expressions for these kinds of subjective meanings. In controllable, pipelined neural NLG separate learning models, notably regression models, can be used to predict whether some feature deviates sufficiently strongly from an expected value, so that evaluative language would be appropriate for verbalizing this finding. In this paper, we present an empirical study on the comprehension of evaluative adverbs and adjectival modifiers in car reviews, a text type that is characterized by a mixture of factual information with evaluations expressing positive or negative surprise. We show to what extend regression-based decision boundaries for producing evaluative content in controllable data2text NLG match the reader’s expectations that are raised by those evaluative markers. Finally we show that regression values in combination with standard deviation of the technical input data constitute reasonable Boolean thresholds for both positive and negative surprise, which provide the basis for the development of more complex models that also include the scalar base of adverbs and modifiers.

To enhance the quality of generated stories, recent story generation models have been investigating the utilization of higher-level attributes like plots or commonsense knowledge. The application of prompt-based learning with large language models (LLMs), exemplified by GPT-3, has exhibited remarkable performance in diverse natural language processing (NLP) tasks. This paper conducts a comprehensive investigation, utilizing both automatic and human evaluation, to compare the story generation capacity of LLMs with recent models across three datasets with variations in style, register, and length of stories. The results demonstrate that LLMs generate stories of significantly higher quality compared to other story generation models. Moreover, they exhibit a level of performance that competes with human authors, albeit with the preliminary observation that they tend to replicate real stories in situations involving world knowledge, resembling a form of plagiarism.

Large Language Models (LLMs) have shown a tremendous capacity for generating literary text. However, their effectiveness in generating children’s stories has yet to be thoroughly examined. In this study, we evaluate the trustworthiness of children’s stories generated by LLMs using various measures, and we compare and contrast our results with both old and new children’s stories to better assess their significance. Our findings suggest that LLMs still struggle to generate children’s stories at the level of quality and nuance found in actual stories.

Text Style Transfer (TST) is performable through approaches such as latent space disentanglement, cycle-consistency losses, prototype editing etc. The prototype editing approach, which is known to be quite successful in TST, involves two key phases a) Masking of source style-associated tokens and b) Reconstruction of this source-style masked sentence conditioned with the target style. We follow a similar transduction method, in which we transpose the more difficult direct source to target TST task to a simpler Style-Masked Language Model (SMLM) Task, wherein, similar to BERT (CITATION), the goal of our model is now to reconstruct the source sentence from its style-masked version. We arrive at the SMLM mechanism naturally by formulating prototype editing/ transduction methods in a probabilistic framework, where TST resolves into estimating a hypothetical parallel dataset from a partially observed parallel dataset, wherein each domain is assumed to have a common latent style-masked prior. To generate this style-masked prior, we use “Explainable Attention” as our choice of attribution for a more precise style-masking step and also introduce a cost-effective and accurate “Attribution-Surplus” method of determining the position of masks from any arbitrary attribution model in O(1) time. We empirically show that this non-generational approach well suites the “content preserving” criteria for a task like TST, even for a complex style like Discourse Manipulation. Our model, the Style MLM, outperforms strong TST baselines and is on par with state-of-the-art TST models, which use complex architectures and orders of more parameters.

Models for affective text generation have shown a remarkable progress, but they commonly rely only on basic emotion theories or valance/arousal values as conditions. This is appropriate when the goal is to create explicit emotion statements (“The kid is happy.”). Emotions are, however, commonly communicated implicitly. For instance, the emotional interpretation of an event (“Their dog died.”) does often not require an explicit emotion statement. In psychology, appraisal theories explain the link between a cognitive evaluation of an event and the potentially developed emotion. They put the assessment of the situation on the spot, for instance regarding the own control or the responsibility for what happens. We hypothesize and subsequently show that including appraisal variables as conditions in a generation framework comes with two advantages. (1) The generation model is informed in greater detail about what makes a specific emotion and what properties it has. This leads to text generation that better fulfills the condition. (2) The variables of appraisal allow a user to perform a more fine-grained control of the generated text, by stating properties of a situation instead of only providing the emotion category. Our Bart and T5-based experiments with 7 emotions (Anger, Disgust, Fear, Guilt, Joy, Sadness, Shame), and 7 appraisals (Attention, Responsibility, Control, Circumstance, Pleasantness, Effort, Certainty) show that (1) adding appraisals during training improves the accurateness of the generated texts by 10 pp in F1. Further, (2) the texts with appraisal variables are longer and contain more details. This exemplifies the greater control for users.

Text style transfer (TST) involves transforming a text into a desired style while approximately preserving its content. The biggest challenge in TST in the general lack of parallel data. Many existing approaches rely on complex models using substantial non-parallel data, with mixed results. In this paper, we leverage a pretrained BART language model with minimal parallel data and incorporate low-resource methods such as hyperparameter tuning, data augmentation, and self-training, which have not been explored in TST. We further include novel style-based rewards in the training loss. Through extensive experiments in sentiment transfer, a sub-task of TST, we demonstrate that our simple yet effective approaches achieve well-balanced results, surpassing non-parallel approaches and highlighting the usefulness of parallel data even in small amounts.

Neural language models are increasingly deployed into APIs and websites that allow a user to pass in a prompt and receive generated text. Many of these systems do not reveal generation parameters. In this paper, we present methods to reverse-engineer the decoding method used to generate text (i.e., top-_k_ or nucleus sampling). Our ability to discover which decoding strategy was used has implications for detecting generated text. Additionally, the process of discovering the decoding strategy can reveal biases caused by selecting decoding settings which severely truncate a model’s predicted distributions. We perform our attack on several families of open-source language models, as well as on production systems (e.g., ChatGPT).

One of the challenges in text generation is to control text generation as intended by the user. Previous studies proposed specifying the keywords that should be included in the generated text. However, this approach is insufficient to generate text that reflect the user’s intent. For example, placing an important keyword at the beginning of the text would help attract the reader’s attention; however, existing methods do not enable such flexible control. In this paper, we tackle a novel task of controlling not only keywords but also the position of each keyword in the text generation. To this end, we propose a task-independent method that uses special tokens to control the relative position of keywords. Experimental results on summarization and story generation tasks show that the proposed method can control keywords and their positions. The experimental results also demonstrate that controlling the keyword positions can generate summary texts that are closer to the user’s intent than baseline.

Hallucinations in text generation occur when the system produces text that is not grounded in the input. In this work, we tackle the problem of hallucinations in neural chart summarization. Our analysis shows that the target side of chart summarization training datasets often contains additional information, leading to hallucinations. We propose a natural language inference (NLI) based method to preprocess the training data and show through human evaluation that our method significantly reduces hallucinations. We also found that shortening long-distance dependencies in the input sequence and adding chart-related information like title and legends improves the overall performance.

The positive text reframing (PTR) task which generates a text giving a positive perspective with preserving the sense of the input text, has attracted considerable attention as one of the NLP applications. Due to the significant representation capability of the pre-trained language model (PLM), a beneficial baseline can be easily obtained by just fine-tuning the PLM. However, how to interpret a diversity of contexts to give a positive perspective is still an open problem. Especially, it is more serious when the size of the training data is limited. In this paper, we present a PTR framework, that learns representations where the meaning and style of text are structurally disentangled. The method utilizes pseudo-positive reframing datasets which are generated with two augmentation strategies. A simple but effective multi-task learning-based model is learned to fuse the generation capabilities from these datasets. Experimental results on Positive Psychology Frames (PPF) dataset, show that our approach outperforms the baselines, BART by five and T5 by six evaluation metrics. Our source codes and data are available online.

Clickbait posts are a widespread problem in the webspace. The generation of spoilers, i.e. short texts that neutralize clickbait by providing information that makes it uninteresting, is one of the proposed solutions to the problem. Current state-of-the-art methods are based on passage retrieval or question answering approaches and are limited to generating spoilers only in the form of a phrase or a passage. In this work, we propose an ensemble of fine-tuned large language models for clickbait spoiler generation. Our approach is not limited to phrase or passage spoilers, but is also able to generate multipart spoilers that refer to several non-consecutive parts of text. Experimental evaluation demonstrates that the proposed ensemble model outperforms the baselines in terms of BLEU, METEOR and BERTScore metrics.

The faithfulness of abstractive text summarization at the named entities level is the focus of this study. We propose to add a new criterion to the summary selection method based on the “risk” of generating entities that do not belong to the source document. This method is based on the assumption that Out-Of-Document entities are more likely to be hallucinations. This assumption was verified by a manual annotation of the entities occurring in a set of generated summaries on the CNN/DM corpus. This study showed that only 29% of the entities outside the source document were inferrable by the annotators, leading to 71% of hallucinations among OOD entities. We test our selection method on the CNN/DM corpus and show that it significantly reduces the hallucination risk on named entities while maintaining competitive results with respect to automatic evaluation metrics like ROUGE.

Gàidhlig (Scottish Gaelic; gd) is spoken by about 57k people in Scotland, but remains an under-resourced language with respect to natural language processing in general and natural language generation (NLG) in particular. To address this gap, we developed the first datasets for Scottish Gaelic NLG, collecting both conversational and summarisation data in a single setting. Our task setup involves dialogues between a pair of speakers discussing museum exhibits, grounding the conversation in images and texts. Then, both interlocutors summarise the dialogue resulting in a secondary dialogue summarisation dataset. This paper presents the dialogue and summarisation corpora, as well as the software used for data collection. The corpus consists of 43 conversations (13.7k words) and 61 summaries (2.0k words), and will be released along with the data collection interface.

In various scenarios, such as conference oral presentations, company managers’ talks, and politicians’ speeches, individuals often contemplate the potential questions that may arise from their presentations. This common practice prompts the research question addressed in this study: to what extent can models generate multiple questions based on a given presentation transcript? To investigate this, we conduct pilot explorations using earnings conference call transcripts, which serve as regular meetings between professional investors and company managers. We experiment with different task settings and methods and evaluate the results from various perspectives. Our findings highlight that incorporating key points retrieval techniques enhances the accuracy and diversity of the generated questions.

Rule-based text generators lack the coverage and fluency of their neural counterparts, but have two big advantages over them: (i) they are entirely controllable and do not hallucinate; and (ii) they can fully explain how an output was generated from an input. In this paper we leverage these two advantages to create large and reliable synthetic datasets with multiple human-intelligible intermediate representations. We present the Modular Data-to-Text (Mod-D2T) Dataset which incorporates ten intermediate-level representations between input triple sets and output text; the mappings from one level to the next can broadly be interpreted as the traditional modular tasks of an NLG pipeline. We describe the Mod-D2T dataset, evaluate its quality via manual validation and discuss its applications and limitations. Data, code and documentation are available at https://github.com/mille-s/Mod-D2T.

This demo paper presents a brief introduction of MiReportor, a computer-aided medical imaging report generator, which leverages a unified framework of medical image understanding and generation to predict readable descriptions for medical images, and assists radiologists in imaging reports writing.

Over the past decade, a variety of neural architectures for data-to-text generation (NLG) have been proposed. However, each system typically has its own approach to pre- and post-processing and other implementation details. Diversity in implementations is desirable, but it also confounds attempts to compare model performance: are the differences due to the proposed architectures or are they a byproduct of the libraries used or a result of pre- and post-processing decisions made? To improve reproducibility, we re-implement several pre-Transformer neural models for data-to-text NLG within a single framework to facilitate direct comparisons of the models themselves and better understand the contributions of other design choices. We release our library at https://github.com/NapierNLP/enunlg to serve as a baseline for ongoing work in this area including research on NLG for low-resource languages where transformers might not be optimal.

VisuaLLM is a Python library that enables interactive visualization of common tasks in natural language generation with pretrained language models (using HuggingFace’s model API), with tight integration of benchmark datasets and fine-grained generation control. The system runs as a local generation backend server and features a web-based frontend, allowing simple interface configuration by minimal Python code. The currently implemented views include data visualization, next-token prediction with probability distributions, and decoding parameter control, with simple extension to additional tasks.

Live commentaries are essential for enhancing spectators’ enjoyment and understanding during sports events or e-sports streams. We introduce a live audio commentator system designed specifically for a racing game, driven by the high demand in the e-sports field. While a player is playing a racing game, our system tracks real-time user play data including speed and steer rotations, and generates commentary to accompany the live stream. Human evaluation suggested that generated commentary enhances enjoyment and understanding of races compared to streams without commentary. Incorporating additional modules to improve diversity and detect irregular events, such as course-outs and collisions, further increases the preference for the output commentaries.

Most languages in the world do not have sufficient data available to develop neural-network-based natural language generation (NLG) systems. To alleviate this resource scarcity, we propose a novel challenge for the NLG community: low-resource language corpus development (LOWRECORP). We present an innovative framework to collect a single dataset with dual tasks to maximize the efficiency of data collection efforts and respect language consultant time. Specifically, we focus on a text-chat-based interface for two generation tasks – conversational response generation grounded in a source document and/or image and dialogue summarization (from the former task). The goal of this shared task is to collectively develop grounded datasets for local and low-resourced languages. To enable data collection, we make available web-based software that can be used to collect these grounded conversations and summaries. Submissions will be assessed for the size, complexity, and diversity of the corpora to ensure quality control of the datasets as well as any enhancements to the interface or novel approaches to grounding conversations.

We propose a shared task of human-like long story generation, LSG Challenge, that asks models to output a consistent human-like long story (a Harry Potter generic audience fanfic in English), given a prompt of about 1K tokens. We suggest a novel statistical metric of the text structuredness, GloVe Autocorrelations Power/ Exponential Law Mean Absolute Percentage Error Ratio (GAPELMAPER) and the use of previously-known UNION metric and a human evaluation protocol. We hope that LSG can open new avenues for researchers to investigate sampling approaches, prompting strategies, autoregressive and non-autoregressive text generation architectures and break the barrier to generate consistent long (40K+ word) texts.

Recent large pre-trained models have achieved strong performance in multimodal language generation, which requires a joint effort of vision and language modeling. However, most previous generation tasks are based on single image input and produce short text descriptions that are not grounded on the input images. In this work, we propose a shared task on visually grounded story generation. The input is an image sequence, and the output is a story that is conditioned on the input images. This task is particularly challenging because: 1) the protagonists in the generated stories need to be grounded in the images and 2) the output story should be a coherent long-form text. We aim to advance the study of vision-based story generation by accepting submissions that propose new methods as well as new evaluation measures.

We propose the VDG Challenge: a shared task that addresses and benchmarks the task of utterance generation in collaborative visual dialogue. The task features two challenging datasets, an evaluation protocol and a tentative schedule. Our shared task will allow researchers to unravel problems of modelling multi-modal interaction and fit of the existing approaches in the NLP and NLG communities.

In the context of language learning, feedback comment generation is the task of generating hints or explanatory notes for learner texts that help understand why a part of text is erroneous. This paper presents our approach to the Feedback Comment Generation Shared Task, collocated with the 16th International Natural Language Generation Conference (INLG 2023). The approach augments the generation of feedback comments by a self-supervised identification of feedback types in a multitasklearning setting. Within the shared task, other approaches performed more effective, yet the combined modeling of feedback type classification and feedback comment generation is superior to performing eedback generation only.

This paper presents a solution to the GenChal 2022 shared task dedicated to feedback comment generation for writing learning. In terms of this task given a text with an error and a span of the error, a system generates an explanatory note that helps the writer (language learner) to improve their writing skills. Our solution is based on fine-tuning the T5 model on the initial dataset augmented according to syntactical dependencies of the words located within indicated error span. The solution of our team ‘nigula’ obtained second place according to manual evaluation by the organizers.

We report on the results of the first ever shared task on feedback comment generation for language learners held as Generation Challenge (GenChal) in INLG 2022, which we call FCG GenChal. Feedback comment generation for language learners is a task where, given a text and a span, a system generates, for the span, an explanatory note that helps the writer (language learner) improve their writing skills. We show how well we can generate feedback comments with present techniques. We also shed light on the task properties and the difficulties in this task, with insights into the task including data development, evaluation, and comparisons of generation systems.

In this paper, we present strong baselines for the task of Feedback Comment Generation for Writing Learning. Given a sentence and an error span, the task is to generate a feedback comment explaining the error. Sentences and feedback comments are both in English. We experiment with LLMs and also create multiple pseudo datasets for the task, investigating how it affects the performance of our system. We present our results for the task along with extensive analysis of the generated comments with the aim of aiding future studies in feedback comment generation for English language learners.

In this paper, we propose a novel method, retrieval, masking, and generation, for feedback comment generation. Feedback comment generation is a task in which a system generates feedback comments such as hints or explanatory notes for language learners, given input text and position showing where to comment. In the conventional study, the retrieve-and-edit method for retrieving feedback comments in the data pool and editing the comments has been thought effective for this task. However, the performance of this method does not perform as well as other conventional methods because its model learns to edit tokens that do not need to be rewritten in the retrieved comments. To mitigate this problem, we propose a method for combining retrieval, masking, and generation based on the retrieve-and-edit method. Specifically, tokens of feedback comments retrieved from the data pool are masked, and this masked feedback comment is used as a template to generate feedback comments. The proposed method should prevent unnecessary conversion by using not retrieved feedback comments directly but masking them. Our experiments on feedback comment generation demonstrate that the proposed method outperforms conventional methods.

In this paper, we introduce our Tokyo Metropolitan University Feedback Comment Generation system submitted to the feedback comment generation task for INLG 2023 Generation Challenge. In this task, a source sentence and offset range of preposition uses are given as the input. Then, a system generates hints or explanatory notes about preposition uses as the output. To tackle this generation task, we finetuned pretrained sequence-to-sequence language models. The models using BART and T5 showed significant improvement in BLEU score, demonstrating the effectiveness of the pretrained sequence-to-sequence language models in this task. We found that using part-of-speech tag information as an auxiliary input improves the generation quality of feedback comments. Furthermore, we adopt a simple postprocessing method that can enhance the reliability of the generation. As a result, our system achieved the F1 score of 47.4 points in BLEU-based evaluation and 60.9 points in manual evaluation, which ranked second and third on the leaderboard.

This paper describes the Tokyo Tech and AIST system in the GenChal 2022 shared task, which is the first shared task of feedback comment generation. We adopted five methods: data cleaning, fine-tuning pre-trained models, correcting errors in learners’ sentences, appending a correcting operation, and filtering out irrelevant outputs. Our system achieved F1 = 43.4 on the test dataset.

The purpose of feedback comment generation is to provide useful feedback comments for a wide range of errors in learners’ essays from a language learning perspective. Since it is difficult to obtain appropriate comments at a practical level with rule-based or retrieval- based methods, we explore neural-based gen- erative methods with pre-trained models. We further assume the effectiveness of consider- ing grammatical terms in generating feedback comments. Specifically, this paper proposes T5-based models using predicted grammati- cal terms, submitted to FCG GenChal, and presents their results. By using correct gram- matical terms, our model could improve the BLEU score by 19.0 points, compared with the baseline T5 without grammatical terms on the development dataset. Furthermore, by using predicted grammatical terms, our model could improve the manual evaluation score by 2.33 points, compared with the baseline T5 without grammatical terms on the test dataset.

We held our 5th annual AIWolf international contest to automatically play the Werewolf game “Mafia”, where players try finding liars via conversations, aiming at promoting developments in creating agents of more natural conversations in higher level, such as longer contexts, personal relationships, semantics, pragmatics, and logics, revealing the capabilities and limits of the generative AIs. In our Natural Language Division of the contest, we had six Japanese speaking agents from five teams, and three English speaking agents, to mutually run games. By using the game logs, We performed human subjective evaluations and detailed log analysis. We found that the entire system performance has largely improved over the previous year, due to the recent advantages of the LLMs. However, it is not perfect at all yet; the generated talks are sometimes inconsistent with the game actions, it is still doubtful that the agents could infer roles by logics rather than superficial utterance generations. It is not explicitly observed in this log but it would be still difficult to make an agent telling a lie, pretend as a villager but it has an opposite goal inside. Our future work includes to reveal the capability of the LLMs, whether they can make the duality of the “liar”, in other words, holding a “true” and a “false” circumstances of the agent at the same time, even holding what these circumstances look like from other agents.

This paper describes Zoom’s submission to the Second Shared Task on Automatic Minuting at INLG 2023. We participated in Task A: generating abstractive summaries of meetings. Our final submission was a transformer model utilizing data from a similar domain and data augmentation by large language models, as well as content-based segmentation. The model produces summaries covering meeting topics and next steps and performs comparably to a large language model at a fraction of the cost. We also find that re-summarizing the summaries with the same model allows for an alternative, shorter summary.

This paper describes the approach we followed for our submission to the Second Run of the Automatic Minuting Shared Task. Our methodology centers around employing BART-based models fine-tuned on diverse summarization corpora. The segmented meeting transcripts are fed into the models, generating summaries that are subsequently combined and formatted into the final meeting minutes.

This report describes the development of our system for automatic minuting created for the AutoMin 2023 Task A. As a baseline, we utilize a system based on the BART encoder-decoder model paired with a preprocessing pipeline similar to the one introduced by the winning solutions at AutoMin 2021. We then further explore the possibilities for iterative summarization by constructing an iterative minuting dataset from the provided data, finetuning on it and feeding the model previously generated minutes. We also experiment with adding more context by utilizing the Longformer encoder-decoder model and finetuning it on the SAMSum dataset. Our submitted solution is of the baseline approach, since we were unable to match its performance with our iterative variants. With the baseline, we achieve a ROUGE-1 score of 0.368 on the ELITR minuting corpus development set. We finally explore the performance of Vicuna 13B quantized language model for summarization.

This document reports the approach of our team Darbarer for the main task (Task A) of the AutoMin 2023 challenge. Our system is composed of four main modules. The first module relies on a text simplification model aiming at standardizing the utterances of the conversation and compressing the input in order to focus on informative content. The second module handles summarization by employing a straightforward segmentation strategy and a fine-tuned BART-based generative model. Then a titling module has been trained in order to propose a short description of each summarized block. Lastly, we apply a post-processing step aimed at enhancing readability through specific formatting rules. Our contributions lie in the first, third and last steps. Our system generates precise and concise minutes. We provide a detailed description of our modules, discuss the difficulty of evaluating their impact and propose an analysis of observed errors in our generated minutes.

This paper documents the approach of Team NTR for the Second Shared Task on Automatic Minuting (AutoMin) at INLG 2023. The goal of this work is to develop a module for automatic generation of meeting minutes based on a meeting transcript text produced by an Automated Speech Recognition (ASR) system (Task A). We consider minuting as a supervised machine learning task on pairs of texts: the transcript of the meeting and its minutes. We use a two-staged minuting pipeline that consists of segmentation and summarization. We experiment with semantic segmentation and multi-language approaches and Large Language Model Dolly, and achieve Rouge1-F of 0.2455 and BERT-Score of 0.8063 on the English part of ELITR test set and Rouge1-F of 0.2430 and BERT-Score of 0.8332 on the EuroParl dev set with the submitted Naive Segmentation + Dolly7b pipeline.

In this article, we report the findings of the second shared task on Automatic Minuting (AutoMin) held as a Generation Challenge at the 16th International Natural Language Generation (INLG) Conference 2023. The second Automatic Minuting shared task is a successor to the first AutoMin which took place in 2021. The primary objective of the AutoMin shared task is to garner participation of the speech and natural language processing and generation community to create automatic methods for generating minutes from multi-party meetings. Five teams from diverse backgrounds participated in the shared task this year. A lot has changed in the Generative AI landscape since the last AutoMin especially with the emergence and wide adoption of Large Language Models (LLMs) to different downstream tasks. Most of the contributions are based on some form of an LLM and we are also adding current outputs of GPT4 as a benchmark. Furthermore, we examine the applicability of GPT-4 for automatic scoring of minutes. Compared to the previous instance of AutoMin, we also add another domain, the minutes for EU Parliament sessions, and we experiment with a more fine-grained manual evaluation. More details on the event can be found at https://ufal.github.io/automin-2023/.

The spread of hate speech (HS) in the digital age poses significant challenges, with online platforms becoming breeding grounds for harmful content. While many natural language processing (NLP) studies have focused on identifying hate speech, few have explored the generation of counter narratives (CNs) as means to combat it. Previous studies have shown that computational models often generate CNs that are dull and generic, and therefore do not resonate with hate speech authors. In this paper, we explore the personalization capabilities of computational models for generating more targeted and engaging CNs. This paper investigates various strategies for incorporating author profiling information into GPT-2 and GPT-3.5 models to enhance the personalization of CNs to combat online hate speech. We investigate the effectiveness of incorporating author profiling aspects, more specifically the age and gender information of HS authors, in tailoring CNs specifically targeted at HS spreaders. We discuss the challenges, opportunities, and future directions for incorporating user profiling information into CN interventions.

Recent computational approaches for combating online hate speech involve the automatic generation of counter narratives by adapting Pretrained Transformer-based Language Models (PLMs) with human-curated data. This process, however, can produce in-domain overfitting, resulting in models generating acceptable narratives only for hatred similar to training data, with little portability to other targets or to real-world toxic language. This paper introduces novel attention regularization methodologies to improve the generalization capabilities of PLMs for counter narratives generation. Overfitting to training-specific terms is then discouraged, resulting in more diverse and richer narratives. We experiment with two attention-based regularization techniques on a benchmark English dataset. Regularized models produce better counter narratives than state-of-the-art approaches in most cases, both in terms of automatic metrics and human evaluation, especially when hateful targets are not present in the training data. This work paves the way for better and more flexible counter-speech generation models, a task for which datasets are highly challenging to produce.

Hate speech is a critical problem in our society and social media platforms are often an amplifier for this phenomenon. Recently the use of Counter Narratives (informative and non-aggressive responses) has been proposed as a viable solution to counter hateful content that goes beyond simple detection-removal strategies. In this paper we present a novel approach along this line of research, which utilizes the implied statement (bias) expressed in the hate speech to retrieve an appropriate counter narrative. To this end, we first trained and tested several LMs that, given a hateful post, generate the underlying bias and the target group. Then, for the counter narrative selection task, we experimented with several methodologies that either use or not use the implied bias during the process. Experiments show that using the target group information allows the system to better focus on relevant content and that implied statement for selecting counter narratives is better than the corresponding standard approach that does not use it. To our knowledge, this is the first attempt to build an automatic selection tool that uses hate speech implied bias to drive Counter Narrative selection.

Counterspeech on social media is rare. Consequently, it is difficult to collect naturally occurring examples, in particular for languages without annotated datasets. In this work, we study methods to increase the relevance of social media samples for counterspeech annotation when we lack annotated resources. We use the example of sourcing German data for counterspeech annotations from Twitter. We monitor tweets from German politicians and activists to collect replies. To select relevant replies we a) find replies that match German abusive keywords or b) label replies for counterspeech using a multilingual classifier fine-tuned on English data. For both approaches and a baseline setting, we annotate a random sample and use bootstrap sampling to estimate the amount of counterspeech. We find that neither the multilingual model nor the keyword approach achieve significantly higher counts of true counterspeech than the baseline. Thus, keyword lists or multi-lingual classifiers are likely not worth the added complexity beyond purposive data collection: Already without additional filtering, we gather a meaningful sample with 7,4% true counterspeech.

Counterspeech has been proposed as a solution to the proliferation of online hate. Research has shown that natural language processing (NLP) approaches could generate such counterspeech automatically, but there are competing ideas for how NLP models might be used for this task and a variety of evaluation metrics whose relationship to one another is unclear. We test three different approaches and collect ratings of the generated counterspeech for 1,740 tweet-participant pairs to systematically compare the counterspeech on three aspects: quality, effectiveness and user preferences. We examine which model performs best at which metric and which aspects of counterspeech predict user preferences. A free-form text generation approach using ChatGPT performs the most consistently well, though its generations are occasionally unspecific and repetitive. In our experiment, participants’ preferences for counterspeech are predicted by the quality of the counterspeech, not its perceived effectiveness. The results can help future research approach counterspeech evaluation more systematically.

With the rise of the metaverse, immersive multimodal conversation has attracted more and more researchers’ attention. Multimodal contexts will become more important for human-computer interaction in the metaverse, especially in shopping domain. Unlike traditional conversation tasks, immersive multimodal conversation has challenges such as multimodal ambiguous candidate identification and multimodal coreference resolution, which makes it more difficult to dialog state tracking and response generation, as described in SIMMC 2.1 challenge, a part of DSTC11. In particular, as the number of objects in the scene increases, the difficulty will increase dramatically. We proposed a prompt-based multi-task learning Encoder-Decoder, in which different subtasks use different prompts to make the model tend to focus on the current subtask. We achieve the winner in ambiguous candidates indentification and runner-up in multimodal coreference resolution (MM-Coref), multimodal dialog state tracking (MM-DST) and assistant response generation. Our code and model are made publicly available at https://github.com/scutcyr/dstc11-simmc2.1-scut-bds-lab.

Recently, research using multimodal datasets containing image and text information has been conducted actively. One of them is the SIMMC2.1 dataset. It is a more complicated dataset than answering a conversation using only text because it should predict an answer after understanding the relationship between images and text. Therefore, there are limitations to answering a conversation only using text-based models such as BERT or GPT-2, so models with both image and language understanding abilities should be considered. We propose a new model that is effective for the ambiguous candidate identification task in DSTC11 SIMMC2.1 Tark. It consists of a simple pipeline model structure, which has two steps. The first step is to check whether there is ambiguity in the current user utterance, and the second step is to extract objects mentioned in the ambiguous utterance of the user. We suggest a new learning framework with a pre-trained image model and text model that is effective for the ambiguous candidate identification task. Experiments show that the proposed method can improve the model performance, and our model achieved 3rd place in sub-task 1 of the SIMMC2.1 track.

To fulfill complex user requirements in a situated conversational scenario, the agent needs to conduct step-by-step multi-modal logic reasoning, which includes locating objects, querying information and searching objects. However, existing methods omit this multi-step procedure and therefore constitutes the risk of shortcuts when making predictions. For example, they may directly copy the information from the dialogue history or simply use the textual description without perform visual reasoning. To address this issue and further boost the system performance, we apply the dual process theory to plug a reasoner into the original transformer based model for step-by-step reasoning. When system 2 completes multi-step reasoning, its output is regarded as final prediction. Our proposed method achieved the 1st rank on the summing scores across all four DSTC-11 SIMMC 2.1 sub-tasks.

The Situated Interactive MultiModal Conversations (SIMMC2.1) Challenge 2022 is hosted by the Eleventh Dialog System Technology Challenge (DSTC11). This is the third consecutive year multimodal dialog systems have been selected as an official track of the competition, promoted by the continued interest in the research community. The task of SIMMC is to create a shopping assistant agent that can communicate with customers in a virtual store. It requires processing store scenes and product catalogs along with the customer’s request. The task is decomposed into four steps and each becomes a subtask. In this work, we explore the common approaches to modeling multimodality and find the method with the most potential. We also identify a discrepancy in using pretrained language models for dialog tasks and devise a simple domain-adaptation method. Our model came in third place for object coreferencing, dialog state tracking, and response generation tasks.

Intent recognition is critical for task-oriented dialogue systems. However, for emerging domains and new services, it is difficult to accurately identify the key intent of a conversation due to time-consuming data annotation and comparatively poor model transferability. Therefore, the automatic induction of dialogue intention is very important for intelligent dialogue systems. This paper presents our solution to Track 2 of Intent Induction from Conversations for Task-Oriented Dialogue at the Eleventh Dialogue System Technology Challenge (DSTC11). The essence of intention clustering lies in distinguishing the representation of different dialogue utterances. The key to automatic intention induction is that, for any given set of new data, the sentence representation obtained by the model can be well distinguished from different labels. Therefore, we propose a multi-stage coarse-to-fine contrastive learning model training scheme including unsupervised contrastive learning pre-training, supervised contrastive learning pre-training, and fine-tuning with joint contrastive learning and clustering to obtain a better dialogue utterance representation model for the clustering task. In the released DSTC11 Track 2 evaluation results, our proposed system ranked first on both of the two subtasks of this Track.

We present our work on Track 2 in the Dialog System Technology Challenges 11 (DSTC11). DSTC11-Track2 aims to provide a benchmark for zero-shot, cross-domain, intent-set induction. In the absence of in-domain training dataset, robust utterance representation that can be used across domains is necessary to induce users’ intentions. To achieve this, we leveraged a multi-domain dialogue dataset to fine-tune the language model and proposed extracting Verb-Object pairs to remove the artifacts of unnecessary information. Furthermore, we devised the method that generates each cluster’s name for the explainability of clustered results. Our approach achieved 3rd place in the precision score and showed superior accuracy and normalized mutual information (NMI) score than the baseline model on various domain datasets.

Natural Language Understanding (NLU) is one of the most critical components of task-oriented dialogue, and it is often considered as an intent classification task. To achieve outstanding intent identification performance, system designers often need to hire a large number of domain experts to label the data, which is inefficient and costly. To address this problem, researchers’ attention has gradually shifted to automatic intent clustering methods, which employ low-resource unsupervised approaches to solve classification problems. The classical framework for clustering is deep clustering, which uses deep neural networks (DNNs) to jointly optimize non-clustering loss and clustering loss. However, for new conversational domains or services, utterances required to assign intents are scarce and the performance of DNNs is often dependent on large amounts of data. In addition, although re-clustering with k-means algorithm after training the network usually leads to better results, k-means methods often suffer from poor stability. To address these problems, we propose an effective two-stage progressive approach to refine the clustering. Firstly, we pre-train the network with contrastive loss using all conversations data and then optimize the clustering loss and contrastive loss simultaneously. Secondly, we propose adaptive progressive k-means to alleviate the randomness of vanilla k-means, achieving better performance and smaller deviation. Our method ranks second in DSTC11 Track2 Task 1, a benchmark for intent clustering of task-oriented dialogue, demonstrating the superiority and effectiveness of our method.

The focus of this work is to investigate unsupervised approaches to overcome quintessential challenges in designing task-oriented dialog schema: assigning intent labels to each dialog turn (intent clustering) and generating a set of intents based on the intent clustering methods (intent induction). We postulate there are two salient factors for automatic induction of intents: (1) clustering algorithm for intent labeling and (2) user utterance embedding space. We compare existing off-the-shelf clustering models and embeddings based on DSTC11 evaluation. Our extensive experiments demonstrate that the combined selection of utterance embedding and clustering method in the intent induction task should be carefully considered. We also present that pretrained MiniLM with Agglomerative clustering shows significant improvement in NMI, ARI, F1, accuracy and example coverage in intent induction tasks. The source codes are available at https://github.com/Jeiyoon/dstc11-track2.

In Task Oriented Dialogue (TOD) system, detecting and inducing new intents are two main challenges to apply the system in the real world. In this paper, we suggest the semantic multiview model to resolve these two challenges: (1) SBERT for General Embedding (GE), (2) Multi Domain Batch (MDB) for dialogue domain knowledge, and (3) Proxy Gradient Transfer (PGT) for cluster-specialized semantic. MDB feeds diverse dialogue datasets to the model at once to tackle the multi-domain problem by learning the multiple domain knowledge. We introduce a novel method PGT, which employs the Siamese network to fine-tune the model with a clustering method directly. Our model can learn how to cluster dialogue utterances by using PGT. Experimental results demonstrate that our multi-view model with MDB and PGT significantly improves the Open Intent Induction performance compared to baseline systems.

Although there have been remarkable advances in dialogue systems through the dialogue systems technology competition (DSTC), it remains one of the key challenges to building a robust task-oriented dialogue system with a speech interface. Most of the progress has been made for text-based dialogue systems since there are abundant datasets with written cor- pora while those with spoken dialogues are very scarce. However, as can be seen from voice assistant systems such as Siri and Alexa, it is of practical importance to transfer the success to spoken dialogues. In this paper, we describe our engineering effort in building a highly successful model that participated in the speech-aware dialogue systems technology challenge track in DSTC11. Our model consists of three major modules: (1) automatic speech recognition error correction to bridge the gap between the spoken and the text utterances, (2) text-based dialogue system (D3ST) for estimating the slots and values using slot descriptions, and (3) post-processing for recovering the error of the estimated slot value. Our experiments show that it is important to use an explicit automatic speech recognition error correction module, post-processing, and data augmentation to adapt a text-based dialogue state tracker for spoken dialogue corpora.

In a real-world environment, Dialogue State Tracking (DST) should use speech recognition results to perform tasks. However, most existing DST research has been conducted in text-based environments. This study aims to build a model that efficiently performs Automatic Speech Recognition-based DST. To operate robustly against speech noise, we used CopyT5, which adopted a copy mechanism, and trained the model using augmented data including speech noise. Furthermore, CopyT5 performed post-training using the masked language modeling method with the MultiWOZ dataset in T5 in order to learn the dialogue context better. The copy mechanism also mitigated name entity errors that may occur during DST generation. Experiments confirmed that data augmentation, post-training, and the copy mechanism effectively improve DST performance.

Though Dialogue State Tracking (DST) is a core component of spoken dialogue systems, recent work on this task mostly deals with chat corpora, disregarding the discrepancies between spoken and written language. In this paper, we propose OLISIA, a cascade system which integrates an Automatic Speech Recognition (ASR) model and a DST model. We introduce several adaptations in the ASR and DST modules to improve integration and robustness to spoken conversations. With these adaptations, our system ranked first in DSTC11 Track 3, a benchmark to evaluate spoken DST. We conduct an in-depth analysis of the results and find that normalizing the ASR outputs and adapting the DST inputs through data augmentation, along with increasing the pre-trained models size all play an important role in reducing the performance discrepancy between written and spoken conversations.

Prior research on dialogue state tracking (DST) is mostly based on written dialogue corpora. For spoken dialogues, the DST model trained on the written text should use the results (or hypothesis) of automatic speech recognition (ASR) as input. But ASR hypothesis often includes errors, which leads to significant performance drop for spoken dialogue state tracking. We address the issue by developing the following ASR error correction modules. First, we train a model to convert ASR hypothesis to ground truth user utterance, which can fix frequent patterns of errors. The model takes ASR hypotheses of two ASR models as input and fine-tuned in two stages. The corrected hypothesis is fed into a large scale pre-trained encoder-decoder model (T5) for DST training and inference. Second, if an output slot value from the encoder-decoder model is a name, we compare it with names in a dictionary crawled from Web sites and, if feasible, replace with the crawled name of the shortest edit distance. Third, we fix errors of temporal expressions in ASR hypothesis by using hand-crafted rules. Experiment results on the DSTC 11 speech-aware dataset, which is built on the popular MultiWOZ task (version 2.1), show that our proposed method can effectively mitigate the performance drop when moving from written text to spoken conversations.

This paper describes the systems submitted by team6 for ChatEval, the DSTC 11 Track 4 competition. We present three different approaches to predicting turn-level qualities of chatbot responses based on large language models (LLMs). We report improvement over the baseline using dynamic few-shot examples from a vector store for the prompts for ChatGPT. We also analyze the performance of the other two approaches and report needed improvements for future work. We developed the three systems over just two weeks, showing the potential of LLMs for this task. An ablation study conducted after the challenge deadline shows that the new Llama 2 models are closing the performance gap between ChatGPT and open-source LLMs. However, we find that the Llama 2 models do not benefit from few-shot examples in the same way as ChatGPT.

Open-domain automatic dialogue evaluation plays an important role in dialogue systems. While recent efforts are being put into making learning-based evaluation metrics correlate better with human evaluation, robust metrics for parallel corpora and multiple domains remain unexplored. Parallel corpora refer to corpora that express the same idea in different ways (e.g., translation, paraphrasing and back-translation). In this paper, we propose Parallel Corpora Alignment Framework (PCAF), which improves the consistency and robustness of model evaluation on parallel corpora. Firstly, parallel corpora are aligned in semantic space through parallel-corpora-aligned contrastive learning. Then, parallel-corpora-aligned distillation on multi-dataset is applied to further improve model’s generalization ability across multiple data domains. Our approach ranks second on the final test data of DSTC11 track4 subtask1 (“Multilingual Automatic Evaluation Metrics”, turn-level) and third on the subtask2 (“Robust Automatic Evaluation Metrics”, turn-level), which proves the strong generalization ability and robustness of our proposed approach.

Despite significant research effort in the development of automatic dialogue evaluation metrics, little thought is given to evaluating dialogues other than in English. At the same time, ensuring metrics are invariant to semantically similar responses is also an overlooked topic. In order to achieve the desired properties of robustness and multilinguality for dialogue evaluation metrics, we propose a novel framework that takes advantage of the strengths of current evaluation models with the newly-established paradigm of prompting Large Language Models (LLMs). Empirical results show our framework achieves state of the art results in terms of mean Spearman correlation scores across several benchmarks and ranks first place on both the Robust and Multilingual tasks of the DSTC11 Track 4 “Automatic Evaluation Metrics for Open-Domain Dialogue Systems”, proving the evaluation capabilities of prompted LLMs.

Task-oriented dialogue systems that employ external knowledge to generate informative responses have become an important field of research. This paper outlines our contribution to Track 5 of the Eleventh Dialog System Technology Challenge (DSTC11), which focuses on constructing high-performing, subjective knowledge-enriched task-oriented dialogue systems. Specifically, we investigate the complementarity of various language models to tackle the diverse knowledge selection task that involves multiple external sources. Based on this investigation, we propose pre- and post-generation model ensemble approaches to mitigate potential biases inherent in using a single model for the knowledge selection task. Finally, we utilize the consensus decoding approach to combine fine-tuned ensemble models and improve the performance of the generation system. Our system ranked 1st in human evaluation, even outperforming human annotation.

As Task-Oriented Dialog (TOD) systems have advanced, structured DB systems, which aim to collect relevant knowledge for answering user’s questions, have also progressed. Despite these advancements, these methods face challenges when dealing with subjective questions from users. To overcome this, DSTC11 released a subjective-knowledge-based TOD (SK-TOD) dataset and benchmark. This paper introduces a framework that effectively solves SK-TOD tasks by leveraging a Large Language Model (LLM). We demonstrate the proficient use of LLM for each sub-task, including an adapters-based method and knowledge-grounded data augmentation. Our proposed methods, which utilize LLM as an efficient tool, outperform baseline performance and approaches that directly use LLM as a one-step sub-task solver, showing superior task-specific optimization.

This paper presents our approach to building a generalized model for Track 5 in DSTC11: “Task-oriented Conversational Modeling with Subjective Knowledge” which addresses the challenge of generating responses to users’ utterances based on a variety of factual and subjective knowledge. To tackle this challenge, we first augmented the training data by leveraging contextual word embedding and back translation, thereby increasing the quantity of available data. Then, we utilized a large-size language model to enhance the acceptability of the augmented data and fine-tuned the model using augmented data. Specifically, we applied the DeBERTa-v3-large model for knowledge detection and selection, and the BART-large model for response generation. Our best model achieved the seventh rank in the objective evaluation and the second rank in the final official human evaluation. These outcomes serve as solid evidence that data augmentation and using a large-size model were highly effective for developing a conversational model system that incorporates objective and subjective knowledge.

This paper describes our submission to the fifth track of the 11th Dialog System Technology Challenge (DSTC-11), which focuses on “Task-oriented Conversational Modeling with Subjective Knowledge”. We focus on response generation and leverage a ranking strategy to ensemble individual models of BART, Long-T5, and a fine-tuned large language model based on LLaMA. The strategy is supplemented by other techniques like low rank adaptation to maintain efficient utilization of these large models while still achieving optimal performance. The experiments show that the ensemble method outperforms individual models and the baseline method. Our model was ranked 1st place in ROUGE_1, 2nd place in ROUGE_L score and 4th place in human evaluation among a total of 14 participating teams.

This paper presents our approach to the DSTC11 Track 5 selection task, which focuses on retrieving appropriate natural language knowledge sources for task-oriented dialogue. We propose typologically diverse back-translation method with typo noise, which could generate various structured user inquries. Through our noised back translation, we augmented inquiries by combining three different typologies of language sources with five different typo noise injections. Our experiments demonstrate that typological variety and typo noise aids the model in generalizing to diverse user inquiries in dialogue. In the competition, where 14 teams participated, our approach achieved the 5th rank for exact matching metric.

This paper discusses our approaches for task-oriented conversational modelling using subjective knowledge, with a particular emphasis on response generation. Our methodology was shaped by an extensive data analysis that evaluated key factors such as response length, sentiment, and dialogue acts present in the provided dataset. We used few-shot learning to augment the data with newly generated subjective knowledge items and present three approaches for DSTC11: (1) task-specific model exploration, (2) incorporation of the most frequent question into all generated responses, and (3) a waterfall prompting technique using a combination of both GPT-3 and ChatGPT.

The goal of DSTC11 track 5 is to build task-oriented dialogue systems that can effectively utilize external knowledge sources such as FAQs and reviews. This year’s challenge differs from previous ones as it includes subjective knowledge snippets and requires multiple snippets for a single turn. We propose a pipeline system for the challenge focusing on entity tracking, knowledge selection and response generation. Specifically, we devise a novel heuristic to ensemble the outputs from the rule-based method and neural model for entity tracking and knowledge selection. We also leverage metadata information in the knowledge source to handle fine-grained user queries. Our approach achieved the first place in objective evaluation and the third place in human evaluation of DSTC11 track 5.

We participate in the 11th Dialog System Technology Challenges (DSTC) track-5 called Task-oriented Conversational Modeling with Subjective Knowledge. Introducing subjective knowledge into task-oriented dialogue (TOD) can help the DS to understand variables of subjective user needs and to suit more dialogue scenarios. Track-5 includes several sub-tasks: 1) knowledge-seeking turn detection; 2) knowledge entity tracking; 3) knowledge entry selection; and 4) use of the selected knowledge entries for response generation. Besides the challenges of each sub-tasks own, there are two challenges across different sub-tasks. The first is that there are multiple valid knowledge entries for each knowledge-seeking turn, the accuracy of the knowledge entry selection is important for the quality of response generation. The second challenge is how to address the unseen dialogue/entities/entries in the validation and the test set. In this paper, we propose a difference-aware ensemble method to address these sub-tasks and the two challenges mentioned above. Our method helps to obtain more robust results and performs well on unseen instances. Among all the submissions for the test set, our method ranks 1st on the knowledge-seeking turn detection task and achieves 3rd on the overall automatic evaluation score. Our code and data will be released on GitHub.

Most research on task oriented dialog modeling is based on written text input. However, users interact with practical dialog systems often using speech as input. Typically, systems convert speech into text using an Automatic Speech Recognition (ASR) system, introducing errors. Furthermore, these systems do not address the differences in written and spoken language. The research on this topic is stymied by the lack of a public corpus. Motivated by these considerations, our goal in hosting the speech-aware dialog state tracking challenge was to create a public corpus or task which can be used to investigate the performance gap between the written and spoken forms of input, develop models that could alleviate this gap, and establish whether Text-to-Speech-based (TTS) systems is a reasonable surrogate to the more-labor intensive human data collection. We created three spoken versions of the popular written-domain MultiWoz task – (a) TTS-Verbatim: written user inputs were converted into speech waveforms using a TTS system, (b) Human-Verbatim: humans spoke the user inputs verbatim, and (c) Human-paraphrased: humans paraphrased the user inputs. Additionally, we provided different forms of ASR output to encourage wider participation from teams that may not have access to state-of-the-art ASR systems. These included ASR transcripts, word time stamps, and latent representations of the audio (audio encoder outputs). In this paper, we describe the corpus, report results from participating teams, provide preliminary analyses of their results, and summarize the current state-of-the-art in this domain.

With ever increasing interest in task-oriented dialog systems, the recent work on Situated and Interactive Multimodal Conversations (SIMMC 2.0) aims to develop personal assistants that interact with users, grounded in an immersive and co-observed setting of photo-realistic scenes. The dataset contains 11k task-oriented dialogs set in an interactive shopping scenario, spanning more than 117k utterances. In order to push research towards this next generation virtual assistants, the SIMMC 2.1 challenge was conducted at the Eleventh Dialog System Technology Challenge (DSTC) which had entries from across the world competing to achieve the state-of-the-art performance in the SIMMC 2.1 task. In this report, we present and compare 13 SIMMC 2.1 model entries from 5 trams across the world to understand the current progress made across the last three years (starting with SIMMC 1.0 and 2.0 challenges) for multimodal task-oriented dialog systems. We hope that our analysis throws light on components that showed promise in addition to identifying the gaps for future research towards this grand goal of an immersive multimodal conversational agent.

With increasing demand for and adoption of virtual assistants, recent work has investigated ways to accelerate bot schema design through the automatic induction of intents or the induction of slots and dialogue states. However, a lack of dedicated benchmarks and standardized evaluation has made progress difficult to track and comparisons between systems difficult to make. This challenge track, held as part of the Eleventh Dialog Systems Technology Challenge, introduces a benchmark that aims to evaluate methods for the automatic induction of customer intents in a realistic setting of customer service interactions between human agents and customers. We propose two subtasks for progressively tackling the automatic induction of intents and corresponding evaluation methodologies. We then present three datasets suitable for evaluating the tasks and propose simple baselines. Finally, we summarize the submissions and results of the challenge track, for which we received submissions from 34 teams.

The advent and fast development of neural networks have revolutionized the research on dialogue systems and subsequently have triggered various challenges regarding their automatic evaluation. Automatic evaluation of open-domain dialogue systems as an open challenge has been the center of the attention of many researchers. Despite the consistent efforts to improve automatic metrics’ correlations with human evaluation, there have been very few attempts to assess their robustness over multiple domains and dimensions. Also, their focus is mainly on the English language. All of these challenges prompt the development of automatic evaluation metrics that are reliable in various domains, dimensions, and languages. This track in the 11th Dialogue System Technology Challenge (DSTC11) is part of the ongoing effort to promote robust and multilingual automatic evaluation metrics. This article describes the datasets and baselines provided to participants and discusses the submission and result details of the two proposed subtasks.

Conventional Task-oriented Dialogue (TOD) Systems rely on domain-specific APIs/DBs or external factual knowledge to create responses. In DSTC11 track 5, we aims to provide a new challenging task to accommodate subjective user requests (e.g.,”Is the WIFI reliable?” or “Does the restaurant have a good atmosphere?” into TOD. We release a benchmark dataset, which contains subjective knowledge-seeking dialogue contexts and manually annotated responses that are grounded in subjective knowledge sources. The challenge track received a total of 48 entries from 14 participating teams.

While the fluency and coherence of Large Language Models (LLMs) in text generation have seen significant improvements, their competency in generating appropriate expressions of uncertainty remains limited.Using a multilingual closed-book QA task and GPT-3.5, we explore how well LLMs are calibrated and express certainty across a diverse set of languages, including low-resource settings. Our results reveal strong performance in high-resource languages but a marked decline in performance in lower-resource languages. Across all, we observe an exaggerated expression of confidence in the model, which does not align with the correctness or likelihood of its responses. Our findings highlight the need for further research into accurate calibration of LLMs especially in a multilingual setting.

The task of Visual Question Generation (VQG) is to generate human-like questions relevant to the given image. As VQG is an emerging research field, existing works tend to focus only on resource-rich language such as English due to the availability of datasets. In this paper, we propose the first Bengali Visual Question Generation task and develop a novel transformer-based encoder-decoder architecture that generates questions in Bengali when given an image. We propose multiple variants of models - (i) image-only: baseline model of generating questions from images without additional information, (ii) image-category and image-answer-category: guided VQG where we condition the model to generate questions based on the answer and the category of expected question. These models are trained and evaluated on the translated VQAv2.0 dataset. Our quantitative and qualitative results establish the first state of the art models for VQG task in Bengali and demonstrate that our models are capable of generating grammatically correct and relevant questions. Our quantitative results show that our image-cat model achieves a BLUE-1 score of 33.12 and BLEU-3 score of 7.56 which is the highest of the other two variants. We also perform a human evaluation to assess the quality of the generation tasks. Human evaluation suggests that image-cat model is capable of generating goal-driven and attribute-specific questions and also stays relevant to the corresponding image.

In Referring Expression Generation, model inputs are often composed of different representations, including the visual properties of the intended referent, its relative position and size, and the visual context. Yet, the extent to which this information influences the generation process of black-box neural models is largely unclear. We investigate the relative weighting of target, location, and context information in the attention components of a Transformer-based generation model. Our results show a general target bias, which, however, depends on the content of the generated expressions, pointing to interesting directions for future research.

Language-and-vision models have shown good performance in tasks such as image-caption matching and caption generation. However, it is challenging for such models to generate pragmatically correct captions, which adequately reflect what is happening in one image or several images. It is crucial to evaluate this behaviour to understand underlying reasons behind it. Here we explore to what extent contextual language-and-vision models are sensitive to different discourse, both textual and visual. In particular, we employ one of the multi-modal transformers (ViLBERT) and test if it can match descriptions and images, differentiating them from distractors of different degree of similarity that are sampled from different visual and textual contexts. We place our evaluation in the multi-sentence and multi-image setup, where images and sentences are expected to form a single narrative structure. We show that the model can distinguish different situations but it is not sensitive to differences within one narrative structure. We also show that performance depends on the task itself, for example, what modality remains unchanged in non-matching pairs or how similar non-matching pairs are to original pairs.

In any system that uses structured knowledge graph (KG) data as its underlying knowledge representation, KG-to-text generation is a useful tool for turning parts of the graph data into text that can be understood by humans. Recent work has shown that models that make use of pretraining on large amounts of text data can perform well on the KG-to-text task, even with relatively little training data on the specific graph-to-text task. In this paper, we build on this concept by using large language models to perform zero-shot generation based on nothing but the model’s understanding of the triple structure from what it can read. We show that ChatGPT achieves near state-of-the-art performance on some measures of the WebNLG 2020 challenge, but falls behind on others. Additionally, we compare factual, counter-factual and fictional statements, and show that there is a significant connection between what the LLM already knows about the data it is parsing and the quality of the output text.

The WebNLG task consists of mapping a knowledge graph to a text verbalising the con- tent of that graph. The 2017 WebNLG edi- tion required participating systems to gener- ate English text from a set of DBpedia triples, while the 2020 WebNLG+ challenge addition- ally included generation into Russian and se- mantic parsing of English and Russian texts. In contrast, WebNLG 2023 focuses on four under-resourced languages which are severely under-represented in research on text genera- tion, namely Breton, Irish, Maltese and Welsh. In addition, WebNLG 2023 once again includes Russian. In this paper, we present the organi- sation of the shared task (data, timeline, eval- uation), briefly describe the participating sys- tems and summarise results for participating systems.

We present our solution for the Russian RDF002 to-text generation task of the WebNLG Challenge 2023. We use the pretrained large language model named FRED-T5 (Zmitrovich et al., 2023) to finetune on the train dataset. Also, we propose several types of prompt and run experiments to analyze their effectiveness. Our submission achieves 0.373 TER on the test dataset, taking the first place according to the results of the automatic evaluation and outperforming the best result of the previous challenge by 0.025. The code of our solution is available at the following link: https://github.com/Ivan30003/webnlg_interno

This paper presents system descriptions of our submitted outputs for WebNLG Challenge 2023. We use mT5 in multi-task and multilingual settings to generate more fluent and reliable verbalizations of the given RDF triples. Furthermore, we introduce a partial decoding technique to produce more elaborate yet simplified outputs. Additionally, we demonstrate the significance of employing better translation systems in creating training data.

LLMs are great at tasks involving English which dominates in their training data. We explore their ability to address tasks involving languages that are severely under-represented in their training data. More specifically, we do this in the context of data-to-text generation for Irish, Maltese, Welsh and Breton. During the prompt-engineering phase we tested GPT-3.5 and~4 with a range of prompt types and formats on a small sample of example input/output pairs. We then fully evaluated the two most promising prompts in two scenarios: (i) direct generation into the under-resourced languages, and (ii) generation into English followed by translation into the under-resourced languages. We find that few-shot prompting works better for direct generation into under-resourced languages, but that the difference disappears when pivoting via English. The few-shot + translation system variants were submitted to the WebNLG 2023 shared task where they outperformed all other systems by substantial margins in all languages on all automatic metrics. We conclude that good performance can be achieved with state-of-the-art LLMs out-of-the box for under-resourced languages. However, best results (for Welsh) of BLEU 25.12, ChrF++ 0.55, and TER 0.64 are well below the lowest ranked English system at WebNLG’20 with BLEU 0.391, ChrF++ 0.579, and TER 0.564.

In this paper, we describe the submission of Dublin City University (DCU) and Trinity College Dublin (TCD) for the WebNLG 2023 shared task. We present a fully rule-based pipeline for generating Irish texts from DBpedia triple sets which comprises 4 components: triple lexicalisation, generation of noninflected Irish text, inflection generation, and post-processing.

This paper presents our submission to the WebNLG Challenge 2023 for generating text in several low-resource languages from RDF-triples. Our submission focuses on using machine translation for generating texts in Irish, Maltese, Welsh and Russian. While a simple and straightfoward approach, recent works have shown that using monolingual models for inference for multilingual tasks with the help of machine translation (translate-test) can out-perform multilingual models and training multilingual models on machine-translated data (translate-train) through careful tuning of the MT component. Our results show that this approach demonstrates competitive performance for this task even with limited data.