Pre-trained language models have shown impressive abilities of understanding and generating natural languages. However, they typically inherit undesired human-like bias and stereotypes from training data, which raises concerns about putting these models into use in real-world scenarios. Although prior research has proposed to reduce bias using different fairness objectives, they usually fail to capture different representations of bias and, therefore, struggle with fully debiasing models. In this work, we introduce a multi-objective probability alignment approach to overcome current challenges by incorporating multiple debiasing losses to locate and penalize bias in different forms. Compared to existing methods, our proposed method can more effectively and comprehensively reduce stereotypical bias, and maintains the language ability of pre-trained models at the same time. Besides, we adopt prefix-tuning to optimize fairness objectives, and results show that it can achieve better bias removal than full fine-tuning while requiring much fewer computational resources. Our code and data are available at https://github.com/Ewanwong/debias_NLG.

The availability of corpora annotated for discourse relations is limited and discourse relation classification performance varies greatly depending on both language and domain. This is a problem for downstream applications that are intended for a language (i.e., not English) or a domain (i.e., not financial news) with comparatively low coverage for discourse annotations. In this paper, we experiment with a state-of-the-art model for discourse relation classification, originally developed for English, extend it to a multi-lingual setting (testing on Italian, Portuguese and Turkish), and employ a simple, yet effective method to mark out-of-domain training instances. By doing so, we aim to contribute to better generalization and more robust discourse relation classification performance across both language and domain.

Pre-trained large language models, such as ChatGPT, archive outstanding performance in various reasoning tasks without supervised training and were found to have outperformed crowdsourcing workers. Nonetheless, ChatGPT’s performance in the task of implicit discourse relation classification, prompted by a standard multiple-choice question, is still far from satisfactory and considerably inferior to state-of-the-art supervised approaches. This work investigates several proven prompting techniques to improve ChatGPT’s recognition of discourse relations. In particular, we experimented with breaking down the classification task that involves numerous abstract labels into smaller subtasks. Nonetheless, experiment results show that the inference accuracy hardly changes even with sophisticated prompt engineering, suggesting that implicit discourse relation classification is not yet resolvable under zero-shot or few-shot settings.

For long document summarization, discourse structure is important to discern the key content of the text and the differences in importance level between sentences. Unfortunately, the integration of rhetorical structure theory (RST) into parameter-efficient fine-tuning strategies for long document summarization remains unexplored. Therefore, this paper introduces RST-LoRA and proposes four RST-aware variants to explicitly incorporate RST into the LoRA model. Our empirical evaluation demonstrates that incorporating the type and uncertainty of rhetorical relations can complementarily enhance the performance of LoRA in summarization tasks. Furthermore, the best-performing variant we introduced outperforms the vanilla LoRA and full-parameter fine-tuning models, as confirmed by multiple automatic and human evaluations, and even surpasses previous state-of-the-art methods.

Large Language Models (LLMs) can generate text by transferring style attributes like formality resulting in formal or informal text.However, instructing LLMs to generate text that when spoken, is more intelligible in an acoustically difficult environment, is an under-explored topic.We conduct the first study to evaluate LLMs on a novel task of generating acoustically intelligible paraphrases for better human speech perception in noise.Our experiments in English demonstrated that with standard prompting, LLMs struggle to control the non-textual attribute, i.e., acoustic intelligibility, while efficiently capturing the desired textual attributes like semantic equivalence. To remedy this issue, we propose a simple prompting approach, prompt-and-select, which generates paraphrases by decoupling the desired textual and non-textual attributes in the text generation pipeline.Our approach resulted in a 40% relative improvement in human speech perception, by paraphrasing utterances that are highly distorted in a listening condition with babble noise at signal-to-noise ratio (SNR) -5 dB. This study reveals the limitation of LLMs in capturing non-textual attributes, and our proposed method showcases the potential of using LLMs for better human speech perception in noise.

Recently, large pre-trained language models (LLMs) have demonstrated superior language understanding abilities, including zero-shot causal reasoning. However, it is unclear to what extent their capabilities are similar to human ones. We here study the processing of an event B in a script-based story, which causally depends on a previous event A. In our manipulation, event A is stated, negated, or omitted in an earlier section of the text. We first conducted a self-paced reading experiment, which showed that humans exhibit significantly longer reading times when causal conflicts exist (¬ A → B) than under logical conditions (A → B). However, reading times remain similar when cause A is not explicitly mentioned, indicating that humans can easily infer event B from their script knowledge. We then tested a variety of LLMs on the same data to check to what extent the models replicate human behavior. Our experiments show that 1) only recent LLMs, like GPT-3 or Vicuna, correlate with human behavior in the ¬ A → B condition. 2) Despite this correlation, all models still fail to predict that nil → B is less surprising than ¬ A → B, indicating that LLMs still have difficulties integrating script knowledge.

Recent studies have claimed that large language models (LLMs) are capable of drawing pragmatic inferences (Qiu et al., 2023; Hu et al., 2022; Barattieri di San Pietro et al., 2023). The present paper sets out to test LLM’s abilities on atypicality inferences, a type of pragmatic inference that is triggered through informational redundancy. We test several state-of-the-art LLMs in a zero-shot setting and find that LLMs fail to systematically fail to derive atypicality inferences. Our robustness analysis indicates that when inferences are seemingly derived in a few-shot settings, these results can be attributed to shallow pattern matching and not pragmatic inferencing. We also analyse the performance of the LLMs at the different derivation steps required for drawing atypicality inferences – our results show that models have access to script knowledge and can use it to identify redundancies and accommodate the atypicality inference. The failure instead seems to stem from not reacting to the subtle maxim of quantity violations introduced by the informationally redundant utterances.

A wide body of evidence shows that human language processing difficulty is predicted by the information-theoretic measure surprisal, a word’s negative log probability in context. However, it is still unclear how to best estimate these probabilities needed for predicting human processing difficulty – while a long-standing belief held that models with lower perplexity would provide more accurate estimates of word predictability, and therefore lead to better reading time predictions, recent work has shown that for very large models, psycholinguistic predictive power decreases. One reason could be that language models might be more confident of their predictions than humans, because they have had exposure to several magnitudes more data. In this paper, we test what effect temperature-scaling of large language model (LLM) predictions has on surprisal estimates and their predictive power of reading times of English texts. Firstly, we show that calibration of large language models typically improves with model size, i.e. poorer calibration cannot account for poorer fit to reading times. Secondly, we find that temperature-scaling probabilities lead to a systematically better fit to reading times (up to 89% improvement in delta log likelihood), across several reading time corpora. Finally, we show that this improvement in fit is chiefly driven by words that are composed of multiple subword tokens.

We present DiscoGeM 2.0, a crowdsourced, parallel corpus of 12,834 implicit discourse relations, with English, German, French and Czech data. We propose and validate a new single-step crowdsourcing annotation method and apply it to collect new annotations in German, French and Czech. The corpus was constructed by having crowdsourced annotators choose a suitable discourse connective for each relation from a set of unambiguous candidates. Every instance was annotated by 10 workers. Our corpus hence represents the first multi-lingual resource that contains distributions of discourse interpretations for implicit relations. The results show that the connective insertion method of discourse annotation can be reliably extended to other languages. The resulting multi-lingual annotations also reveal that implicit relations inferred in one language may differ from those inferred in the translation, meaning the annotations are not always directly transferable. DiscoGem 2.0 promotes the investigation of cross-linguistic differences in discourse marking and could improve automatic discourse parsing applications. It is openly downloadable here: https://github.com/merelscholman/DiscoGeM.

Nigerian Pidgin is an English-derived contact language and is traditionally an oral language, spoken by approximately 100 million people. No orthographic standard has yet been adopted, and thus the few available Pidgin datasets that exist are characterised by noise in the form of orthographic variations. This contributes to under-performance of models in critical NLP tasks. The current work is the first to describe various types of orthographic variations commonly found in Nigerian Pidgin texts, and model this orthographic variation. The variations identified in the dataset form the basis of a phonetic-theoretic framework for word editing, which is used to generate orthographic variations to augment training data. We test the effect of this data augmentation on two critical NLP tasks: machine translation and sentiment analysis. The proposed variation generation framework augments the training data with new orthographic variants which are relevant for the test set but did not occur in the training set originally. Our results demonstrate the positive effect of augmenting the training data with a combination of real texts from other corpora as well as synthesized orthographic variation, resulting in performance improvements of 2.1 points in sentiment analysis and 1.4 BLEU points in translation to English.

Data-to-text (D2T) generation describes the task of verbalizing data, often given as attribute-value pairs. While this task is relevant for many different data domains beyond the traditionally well-explored tasks of weather forecasting, restaurant recommendations, and sports reporting, a major challenge to the applicability of data-to-text generation methods is typically data sparsity. For many applications, there is extremely little training data in terms of attribute-value inputs and target language outputs available for training a model. Given the sparse data setting, recently developed prompting methods seem most suitable for addressing D2T tasks since they do not require substantial amounts of training data, unlike finetuning approaches. However, prompt-based approaches are also challenging, as a) the design and search of prompts are non-trivial; and b) hallucination problems may occur because of the strong inductive bias of these models. In this paper, we propose a retrieval-augmented modular prompt tuning () method, which constructs prompts that fit the input data closely, thereby bridging the domain gap between the large-scale language model and the structured input data. Experiments show that our method generates texts with few hallucinations and achieves state-of-the-art performance on a dataset for drone handover message generation.

Scientific news reports serve as a bridge, adeptly translating complex research articles into reports that resonate with the broader public. The automated generation of such narratives enhances the accessibility of scholarly insights. In this paper, we present a new corpus to facilitate this paradigm development. Our corpus comprises a parallel compilation of academic publications and their corresponding scientific news reports across nine disciplines. To demonstrate the utility and reliability of our dataset, we conduct an extensive analysis, highlighting the divergences in readability and brevity between scientific news narratives and academic manuscripts. We benchmark our dataset employing state-of-the-art text generation models. The evaluation process involves both automatic and human evaluation, which lays the groundwork for future explorations into the automated generation of scientific news reports. The dataset and code related to this work are available at https://dongqi.me/projects/SciNews.

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

Automatic generation of code from natural language descriptions has emerged as one of the main use cases of large language models (LLMs). This has also led to a proliferation of datasets to track progress in the reliability of code generation models, including domains such as programming challenges and common data science tasks. However, existing datasets primarily target the use of code generation models to aid expert programmers in writing code. In this work, we consider a domain of code generation which is more frequently used by users without sophisticated programming skills: translating English descriptions to spreadsheet formulas that can be used to do everyday data processing tasks. We extract naturalistic instructions from StackOverflow posts and manually verify and standardize the corresponding spreadsheet formulas. We use this dataset to evaluate an off-the-shelf code generation model (GPT 3.5 text-davinci-003) as well as recently proposed pragmatic code generation procedures and find that Code Reviewer reranking (Zhang et al., 2022) performs best among the evaluated methods but still frequently generates formulas that differ from human-generated ones.

For text summarization, the role of discourse structure is pivotal in discerning the core content of a text. Regrettably, prior studies on incorporating Rhetorical Structure Theory (RST) into transformer-based summarization models only consider the nuclearity annotation, thereby overlooking the variety of discourse relation types. This paper introduces the ‘RSTformer’, a novel summarization model that comprehensively incorporates both the types and uncertainty of rhetorical relations. Our RST-attention mechanism, rooted in document-level rhetorical structure, is an extension of the recently devised Longformer framework. Through rigorous evaluation, the model proposed herein exhibits significant superiority over state-of-the-art models, as evidenced by its notable performance on several automatic metrics and human evaluation.

Large-scale language models, like ChatGPT, have garnered significant media attention and stunned the public with their remarkable capacity for generating coherent text from short natural language prompts. In this paper, we aim to conduct a systematic inspection of ChatGPT’s performance in two controllable generation tasks, with respect to ChatGPT’s ability to adapt its output to different target audiences (expert vs. layman) and writing styles (formal vs. informal). Additionally, we evaluate the faithfulness of the generated text, and compare the model’s performance with human-authored texts. Our findings indicate that the stylistic variations produced by humans are considerably larger than those demonstrated by ChatGPT, and the generated texts diverge from human samples in several characteristics, such as the distribution of word types. Moreover, we observe that ChatGPT sometimes incorporates factual errors or hallucinations when adapting the text to suit a specific style.

Knowing exactly how many data points need to be labeled to achieve a certain model performance is a hugely beneficial step towards reducing the overall budgets for annotation. It pertains to both active learning and traditional data annotation, and is particularly beneficial for low resource scenarios. Nevertheless, it remains a largely under-explored area of research in NLP. We therefore explored various techniques for estimating the training sample size necessary to achieve a targeted performance value. We derived a simple yet effective approach to predict the maximum achievable model performance based on small amount of training samples – which serves as an early indicator during data annotation for data quality and sample size determination. We performed ablation studies on four language understanding tasks, and showed that the proposed approach allows us to forecast model performance within a small margin of mean absolute error (~0.9%) with only 10% data.

Local coherence is essential for long-form text generation models. We identify two important aspects of local coherence within the visual storytelling task: (1) the model needs to represent re-occurrences of characters within the image sequence in order to mention them correctly in the story; (2) character representations should enable us to find instances of the same characters and distinguish different characters. In this paper, we propose a loss function inspired by a linguistic theory of coherence for self-supervised learning for image sequence representations. We further propose combining features from an object and a face detector to construct stronger character features. To evaluate input-output relevance that current reference-based metrics don’t measure, we propose a character matching metric to check whether the models generate referring expressions correctly for characters in input image sequences. Experiments on a visual story generation dataset show that our proposed features and loss function are effective for generating more coherent and visually grounded stories.

Discourse relations have different patterns of marking across different languages. As a result, discourse connectives are often added, omitted, or rephrased in translation. Prior work has shown a tendency for explicitation of discourse connectives, but such work was conducted using restricted sample sizes due to difficulty of connective identification and alignment. The current study exploits automatic methods to facilitate a large-scale study of connectives in English and German parallel texts. Our results based on over 300 types and 18000 instances of aligned connectives and an empirical approach to compare the cross-lingual specificity gap provide strong evidence of the Explicitation Hypothesis. We conclude that discourse relations are indeed more explicit in translation than texts written originally in the same language. Automatic annotations allow us to carry out translation studies of discourse relations on a large scale. Our methodology using relative entropy to study the specificity of connectives also provides more fine-grained insights into translation patterns.

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.

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.

Current work on image-based story generation suffers from the fact that the existing image sequence collections do not have coherent plots behind them. We improve visual story generation by producing a new image-grounded dataset, Visual Writing Prompts (VWP). VWP contains almost 2K selected sequences of movie shots, each including 5-10 images. The image sequences are aligned with a total of 12K stories which were collected via crowdsourcing given the image sequences and a set of grounded characters from the corresponding image sequence. Our new image sequence collection and filtering process has allowed us to obtain stories that are more coherent, diverse, and visually grounded compared to previous work. We also propose a character-based story generation model driven by coherence as a strong baseline. Evaluations show that our generated stories are more coherent, visually grounded, and diverse than stories generated with the current state-of-the-art model. Our code, image features, annotations and collected stories are available at https://vwprompt.github.io/.

Disagreement in natural language annotation has mostly been studied from a perspective of biases introduced by the annotators and the annotation frameworks. Here, we propose to analyze another source of bias—task design bias, which has a particularly strong impact on crowdsourced linguistic annotations where natural language is used to elicit the interpretation of lay annotators. For this purpose we look at implicit discourse relation annotation, a task that has repeatedly been shown to be difficult due to the relations’ ambiguity. We compare the annotations of 1,200 discourse relations obtained using two distinct annotation tasks and quantify the biases of both methods across four different domains. Both methods are natural language annotation tasks designed for crowdsourcing. We show that the task design can push annotators towards certain relations and that some discourse relation senses can be better elicited with one or the other annotation approach. We also conclude that this type of bias should be taken into account when training and testing models.

In discourse relation recognition, the classification labels are typically represented as one-hot vectors. However, the categories are in fact not all independent of one another on the contrary, there are several frameworks that describe the labels’ similarities (by e.g. sorting them into a hierarchy or describing them interms of features (Sanders et al., 2021)). Recently, several methods for representing the similarities between labels have been proposed (Zhang et al., 2018; Wang et al., 2018; Xiong et al., 2021). We here explore and extend the Label Confusion Model (Guo et al., 2021) for learning a representation for discourse relation labels. We explore alternative ways of informing the model about the similarities between relations, by representing relations in terms of their names (and parent category), their typical markers, or in terms of CCR features that describe the relations. Experimental results show that exploiting label similarity improves classification results.

Implicit discourse relations can convey more than one relation sense, but much of the research on discourse relations has focused on single relation senses. Recently, DiscoGeM, a novel multi-domain corpus, which contains 10 crowd-sourced labels per relational instance, has become available. In this paper, we analyse the co-occurrences of relations in DiscoGem and show that they are systematic and characteristic of text genre. We then test whether information on multi-label distributions in the data can help implicit relation classifiers. Our results show that incorporating multiple labels in parser training can improve its performance, and yield label distributions which are more similar to human label distributions, compared to a parser that is trained on just a single most frequent label per instance.

Obtaining linguistic annotation from novice crowdworkers is far from trivial. A case in point is the annotation of discourse relations, which is a complicated task. Recent methods have obtained promising results by extracting relation labels from either discourse connectives (DCs) or question-answer (QA) pairs that participants provide. The current contribution studies the effect of worker selection and training on the agreement on implicit relation labels between workers and gold labels, for both the DC and the QA method. In Study 1, workers were not specifically selected or trained, and the results show that there is much room for improvement. Study 2 shows that a combination of selection and training does lead to improved results, but the method is cost- and time-intensive. Study 3 shows that a selection-only approach is a viable alternative; it results in annotations of comparable quality compared to annotations from trained participants. The results generalized over both the DC and QA method and therefore indicate that a selection-only approach could also be effective for other crowdsourced discourse annotation tasks.

We present DiscoGeM, a crowdsourced corpus of 6,505 implicit discourse relations from three genres: political speech, literature, and encyclopedic texts. Each instance was annotated by 10 crowd workers. Various label aggregation methods were explored to evaluate how to obtain a label that best captures the meaning inferred by the crowd annotators. The results show that a significant proportion of discourse relations in DiscoGeM are ambiguous and can express multiple relation senses. Probability distribution labels better capture these interpretations than single labels. Further, the results emphasize that text genre crucially affects the distribution of discourse relations, suggesting that genre should be included as a factor in automatic relation classification. We make available the newly created DiscoGeM corpus, as well as the dataset with all annotator-level labels. Both the corpus and the dataset can facilitate a multitude of applications and research purposes, for example to function as training data to improve the performance of automatic discourse relation parsers, as well as facilitate research into non-connective signals of discourse relations.

We present Barch, a new English dataset of human-written summaries describing bar charts. This dataset contains 47 charts based on a selection of 18 topics. Each chart is associated with one of the four intended messages expressed in the chart title. Using crowdsourcing, we collected around 20 summaries per chart, or one thousand in total. The text of the summaries is aligned with the chart data as well as with analytical inferences about the data drawn by humans. Our datasets is one of the first to explore the effect of intended messages on the data descriptions in chart summaries. Additionally, it lends itself well to the task of training data-driven systems for chart-to-text generation. We provide results on the performance of state-of-the-art neural generation models trained on this dataset and discuss the strengths and shortcomings of different models.

Natural language generation in real-time settings with raw sensor data is a challenging task. We find that formulating the task as an end-to-end problem leads to two major challenges in content selection – the sensor data is both redundant and diverse across environments, thereby making it hard for the encoders to select and reason on the data. We here present a new corpus for a specific domain that instantiates these properties. It includes handover utterances that an assistant for a semi-autonomous drone uses to communicate with humans during the drone flight. The corpus consists of sensor data records and utterances in 8 different environments. As a structured intermediary representation between data records and text, we explore the use of description logic (DL). We also propose a neural generation model that can alert the human pilot of the system state and environment in preparation of the handover of control.

The Creative Summarization Shared Task at COLING 2022 aspires to generate summaries given long-form texts from creative writing. This paper presents the system architecture and the results of our participation in the Scriptbase track that focuses on generating movie plots given movie scripts. The core innovation in our model employs a two-stage hierarchical architecture for movie script summarization. In the first stage, a heuristic extraction method is applied to extract actions and essential dialogues, which reduces the average length of input movie scripts by 66% from about 24K to 8K tokens. In the second stage, a state-of-the-art encoder-decoder model, Longformer-Encoder-Decoder (LED), is trained with effective fine-tuning methods, BitFit and NoisyTune. Evaluations on the unseen test set indicate that our system outperforms both zero-shot LED baselines as well as other participants on various automatic metrics and ranks 1st in the Scriptbase track.

Neural natural language generation (NLG) and understanding (NLU) models are costly and require massive amounts of annotated data to be competitive. Recent data programming frameworks address this bottleneck by allowing human supervision to be provided as a set of labeling functions to construct generative models that synthesize weak labels at scale. However, these labeling functions are difficult to build from scratch for NLG/NLU models, as they often require complex rule sets to be specified. To this end, we propose a novel data programming framework that can jointly construct labeled data for language generation and understanding tasks – by allowing the annotators to modify an automatically-inferred alignment rule set between sequence labels and text, instead of writing rules from scratch. Further, to mitigate the effect of poor quality labels, we propose a dually-regularized denoising mechanism for optimizing the NLU and NLG models. On two benchmarks we show that the framework can generate high-quality data that comes within a 1.48 BLEU and 6.42 slot F1 of the 100% human-labeled data (42k instances) with just 100 labeled data samples – outperforming benchmark annotation frameworks and other semi-supervised approaches.

In many linguistic fields requiring annotated data, multiple interpretations of a single item are possible. Multi-label annotations more accurately reflect this possibility. However, allowing for multi-label annotations also affects the chance that two coders agree with each other. Calculating inter-coder agreement for multi-label datasets is therefore not trivial. In the current contribution, we evaluate different metrics for calculating agreement on multi-label annotations: agreement on the intersection of annotated labels, an augmented version of Cohen’s Kappa, and precision, recall and F1. We propose a bootstrapping method to obtain chance agreement for each measure, which allows us to obtain an adjusted agreement coefficient that is more interpretable. We demonstrate how various measures affect estimates of agreement on simulated datasets and present a case study of discourse relation annotations. We also show how the proportion of double labels, and the entropy of the label distribution, influences the measures outlined above and how a bootstrapped adjusted agreement can make agreement measures more comparable across datasets in multi-label scenarios.

Script knowledge is useful to a variety of NLP tasks. However, existing resources only cover a small number of activities, limiting its practical usefulness. In this work, we propose a zero-shot learning approach to script parsing, the task of tagging texts with scenario-specific event and participant types, which enables us to acquire script knowledge without domain-specific annotations. We (1) learn representations of potential event and participant mentions by promoting cluster consistency according to the annotated data; (2) perform clustering on the event / participant candidates from unannotated texts that belongs to an unseen scenario. The model achieves 68.1/74.4 average F1 for event / participant parsing, respectively, outperforming a previous CRF model that, in contrast, has access to scenario-specific supervision. We also evaluate the model by testing on a different corpus, where it achieved 55.5/54.0 average F1 for event / participant parsing.

The surging demand for multilingual dialogue systems often requires a costly labeling process for each language addition. For low resource languages, human annotators are continuously tasked with the adaptation of resource-rich language utterances for each new domain. However, this prohibitive and impractical process can often be a bottleneck for low resource languages that are still without proper translation systems nor parallel corpus. In particular, it is difficult to obtain task-specific low resource language annotations for the English-derived creoles (e.g. Nigerian and Cameroonian Pidgin). To address this issue, we utilize the pretrained language models i.e. BART which has shown great potential in language generation/understanding – we propose to finetune the BART model to generate utterances in Pidgin by leveraging the proximity of the source and target languages, and utilizing positive and negative examples in constrastive training objectives. We collected and released the first parallel Pidgin-English conversation corpus in two dialogue domains and showed that this simple and effective technique is suffice to yield impressive results for English-to-Pidgin generation, which are two closely-related languages.

The demand for multilingual dialogue systems often requires a costly labeling process, where human translators derive utterances in low resource languages from resource rich language annotation. To this end, we explore leveraging the inductive biases for target languages learned by numerous pretrained teacher models by transferring them to student models via sequence-level knowledge distillation. By assuming no target language text, the both the teacher and student models need to learn from the target distribution in a few/zero-shot manner. On the MultiATIS++ benchmark, we explore the effectiveness of our proposed technique to derive the multilingual text for 6 languages, using only the monolingual English data and the pretrained models. We show that training on the synthetic multilingual generation outputs yields close performance to training on human annotations in both slot F1 and intent accuracy; the synthetic text also scores high in naturalness and correctness based on human evaluation.

Large-scale pretrained language models have led to dramatic improvements in text generation. Impressive performance can be achieved by finetuning only on a small number of instances (few-shot setting). Nonetheless, almost all previous work simply applies random sampling to select the few-shot training instances. Little to no attention has been paid to the selection strategies and how they would affect model performance. In this work, we present a study on training instance selection in few-shot neural text generation. The selection decision is made based only on the unlabeled data so as to identify the most worthwhile data points that should be annotated under some budget of labeling cost. Based on the intuition that the few-shot training instances should be diverse and representative of the entire data distribution, we propose a simple selection strategy with K-means clustering. We show that even with the naive clustering-based approach, the generation models consistently outperform random sampling on three text generation tasks: data-to-text generation, document summarization and question generation. The code and training data are made available. We hope that this work will call for more attention on this largely unexplored area.

Implicit discourse relation classification is a challenging task, in particular when the text domain is different from the standard Penn Discourse Treebank (PDTB; Prasad et al., 2008) training corpus domain (Wall Street Journal in 1990s). We here tackle the task of implicit discourse relation classification on the biomedical domain, for which the Biomedical Discourse Relation Bank (BioDRB; Prasad et al., 2011) is available. We show that entity information can be used to improve discourse relational argument representation. In a first step, we show that explicitly marked instances that are content-wise similar to the target relations can be used to achieve good performance in the cross-domain setting using a simple unsupervised voting pipeline. As a further step, we show that with the linked entity information from the first step, a transformer which is augmented with entity-related information (KBERT; Liu et al., 2020) sets the new state of the art performance on the dataset, outperforming the large pre-trained BioBERT (Lee et al., 2020) model by 2% points.

In data-driven natural language generation, we typically know what relation should be expressed and need to select a connective to lexicalize it. In the current contribution, we analyse whether a sophisticated connective generation module is necessary to select a connective, or whether this can be solved with simple methods (such as random choice between connectives that are known to express a given relation, or usage of a generic language model). Comparing these methods to the distributions of connective choices from a human connective insertion task, we find mixed results: for some relations, it is acceptable to lexicalize them using any of the connectives that mark this relation. However, for other relations (temporals, concessives) either a more detailed relation distinction needs to be introduced, or a more sophisticated connective choice module would be necessary.

Cross-linguistic research on discourse structure and coherence marking requires discourse-annotated corpora and connective lexicons in a large number of languages. However, the availability of such resources is limited, especially for languages for which linguistic resources are scarce in general, such as Nigerian Pidgin. In this study, we demonstrate how a semi-automatic approach can be used to source connectives and their relation senses and develop a discourse-annotated corpus in a low-resource language. Connectives and their relation senses were extracted from a parallel corpus combining automatic (PDTB end-to-end parser) and manual annotations. This resulted in Naija-Lex, a lexicon of discourse connectives in Nigerian Pidgin with English translations. The lexicon shows that the majority of Nigerian Pidgin connectives are borrowed from its English lexifier, but that there are also some connectives that are unique to Nigerian Pidgin.

Existing parse methods use varying approaches to identify explicit discourse connectives, but their performance has not been consistently evaluated in comparison to each other, nor have they been evaluated consistently on text other than newspaper articles. We here assess the performance on explicit connective identification of three parse methods (PDTB e2e, Lin et al., 2014; the winner of CONLL2015, Wang et al., 2015; and DisSent, Nie et al., 2019), along with a simple heuristic. We also examine how well these systems generalize to different datasets, namely written newspaper text (PDTB), written scientific text (BioDRB), prepared spoken text (TED-MDB) and spontaneous spoken text (Disco-SPICE). The results show that the e2e parser outperforms the other parse methods in all datasets. However, performance drops significantly from the PDTB to all other datasets. We provide a more fine-grained analysis of domain differences and connectives that prove difficult to parse, in order to highlight the areas where gains can be made.

Recent advancements in data-to-text generation largely take on the form of neural end-to-end systems. Efforts have been dedicated to improving text generation systems by changing the order of training samples in a process known as curriculum learning. Past research on sequence-to-sequence learning showed that curriculum learning helps to improve both the performance and convergence speed. In this work, we delve into the same idea surrounding the training samples consisting of structured data and text pairs, where at each update, the curriculum framework selects training samples based on the model’s competence. Specifically, we experiment with various difficulty metrics and put forward a soft edit distance metric for ranking training samples. On our benchmarks, it shows faster convergence speed where training time is reduced by 38.7% and performance is boosted by 4.84 BLEU.

For many new application domains for data-to-text generation, the main obstacle in training neural models consists of a lack of training data. While usually large numbers of instances are available on the data side, often only very few text samples are available. To address this problem, we here propose a novel few-shot approach for this setting. Our approach automatically augments the data available for training by (i) generating new text samples based on replacing specific values by alternative ones from the same category, (ii) generating new text samples based on GPT-2, and (iii) proposing an automatic method for pairing the new text samples with data samples. As the text augmentation can introduce noise to the training data, we use cycle consistency as an objective, in order to make sure that a given data sample can be correctly reconstructed after having been formulated as text (and that text samples can be reconstructed from data). On both the E2E and WebNLG benchmarks, we show that this weakly supervised training paradigm is able to outperform fully supervised sequence-to-sequence models with less than 10% of the training set. By utilizing all annotated data, our model can boost the performance of a standard sequence-to-sequence model by over 5 BLEU points, establishing a new state-of-the-art on both datasets.

Neural natural language generation (NLG) and understanding (NLU) models are data-hungry and require massive amounts of annotated data to be competitive. Recent frameworks address this bottleneck with generative models that synthesize weak labels at scale, where a small amount of training labels are expert-curated and the rest of the data is automatically annotated. We follow that approach, by automatically constructing a large-scale weakly-labeled data with a fine-tuned GPT-2, and employ a semi-supervised framework to jointly train the NLG and NLU models. The proposed framework adapts the parameter updates to the models according to the estimated label-quality. On both the E2E and Weather benchmarks, we show that this weakly supervised training paradigm is an effective approach under low resource scenarios with as little as 10 data instances, and outperforming benchmark systems on both datasets when 100% of the training data is used.

We propose a shared task on training instance selection for few-shot neural text generation. Large-scale pretrained language models have led to dramatic improvements in few-shot text generation. Nonetheless, almost all previous work simply applies random sampling to select the few-shot training instances. Little to no attention has been paid to the selection strategies and how they would affect model performance. Studying the selection strategy can help us (1) make the most use of our annotation budget in downstream tasks and (2) better benchmark few-shot text generative models. We welcome submissions that present their selection strategies and the effects on the generation quality.

In this paper, we aim to address the challenges surrounding the translation of ancient Chinese text: (1) The linguistic gap due to the difference in eras results in translations that are poor in quality, and (2) most translations are missing the contextual information that is often very crucial to understanding the text. To this end, we improve upon past translation techniques by proposing the following: We reframe the task as a multi-label prediction task where the model predicts both the translation and its particular era. We observe that this helps to bridge the linguistic gap as chronological context is also used as auxiliary information. We validate our framework on a parallel corpus annotated with chronology information and show experimentally its efficacy in producing quality translation outputs. We release both the code and the data for future research.

Automatically generated stories need to be not only coherent, but also interesting. Apart from realizing a story line, the text also needs to include rich details to engage the readers. We propose a model that features two different generation components: an outliner, which proceeds the main story line to realize global coherence; a detailer, which supplies relevant details to the story in a locally coherent manner. Human evaluations show our model substantially improves the informativeness of generated text while retaining its coherence, outperforming various baselines.

We present a lightweight annotation tool, the Data AnnotatoR Tool (DART), for the general task of labeling structured data with textual descriptions. The tool is implemented as an interactive application that reduces human efforts in annotating large quantities of structured data, e.g. in the format of a table or tree structure. By using a backend sequence-to-sequence model, our system iteratively analyzes the annotated labels in order to better sample unlabeled data. In a simulation experiment performed on annotating large quantities of structured data, DART has been shown to reduce the total number of annotations needed with active learning and automatically suggesting relevant labels.

A problem in automatically generated stories for image sequences is that they use overly generic vocabulary and phrase structure and fail to match the distributional characteristics of human-generated text. We address this problem by introducing explicit representations for objects and their relations by extracting scene graphs from the images. Utilizing an embedding of this scene graph enables our model to more explicitly reason over objects and their relations during story generation, compared to the global features from an object classifier used in previous work. We apply metrics that account for the diversity of words and phrases of generated stories as well as for reference to narratively-salient image features and show that our approach outperforms previous systems. Our experiments also indicate that our models obtain competitive results on reference-based metrics.

Implicit discourse relation classification is one of the most difficult tasks in discourse parsing. Previous studies have generally focused on extracting better representations of the relational arguments. In order to solve the task, it is however additionally necessary to capture what events are expected to cause or follow each other. Current discourse relation classifiers fall short in this respect. We here show that this shortcoming can be effectively addressed by using the bidirectional encoder representation from transformers (BERT) proposed by Devlin et al. (2019), which were trained on a next-sentence prediction task, and thus encode a representation of likely next sentences. The BERT-based model outperforms the current state of the art in 11-way classification by 8% points on the standard PDTB dataset. Our experiments also demonstrate that the model can be successfully ported to other domains: on the BioDRB dataset, the model outperforms the state of the art system around 15% points.

The Multilingual Surface Realization Shared Task 2019 focuses on generating sentences from lemmatized sets of universal dependency parses with rich features. This paper describes the results of our participation in the deep track. The core innovation in our approach is to use a graph convolutional network to encode the dependency trees given as input. Upon adding morphological features, our system achieves the third rank without using data augmentation techniques or additional components (such as a re-ranker).

Implicit discourse relation classification is one of the most difficult steps in discourse parsing. The difficulty stems from the fact that the coherence relation must be inferred based on the content of the discourse relational arguments. Therefore, an effective encoding of the relational arguments is of crucial importance. We here propose a new model for implicit discourse relation classification, which consists of a classifier, and a sequence-to-sequence model which is trained to generate a representation of the discourse relational arguments by trying to predict the relational arguments including a suitable implicit connective. Training is possible because such implicit connectives have been annotated as part of the PDTB corpus. Along with a memory network, our model could generate more refined representations for the task. And on the now standard 11-way classification, our method outperforms the previous state of the art systems on the PDTB benchmark on multiple settings including cross validation.

Implicit discourse relation classification is one of the most challenging and important tasks in discourse parsing, due to the lack of connectives as strong linguistic cues. A principle bottleneck to further improvement is the shortage of training data (ca. 18k instances in the Penn Discourse Treebank (PDTB)). Shi et al. (2017) proposed to acquire additional data by exploiting connectives in translation: human translators mark discourse relations which are implicit in the source language explicitly in the translation. Using back-translations of such explicitated connectives improves discourse relation parsing performance. This paper addresses the open question of whether the choice of the translation language matters, and whether multiple translations into different languages can be effectively used to improve the quality of the additional data.

Sentences like “Every child climbed a tree” have at least two interpretations depending on the precedence order of the universal quantifier and the indefinite. Previous experimental work explores the role that different mechanisms such as semantic reanalysis and world knowledge may have in enabling each interpretation. This paper discusses a web-based task that uses the verb-second characteristic of German main clauses to estimate the influence of word order variation over world knowledge.

Automatically generating globally coherent stories is a challenging problem. Neural text generation models have been shown to perform well at generating fluent sentences from data, but they usually fail to keep track of the overall coherence of the story after a couple of sentences. Existing work that incorporates a text planning module succeeded in generating recipes and dialogues, but appears quite data-demanding. We propose a novel story generation approach that generates globally coherent stories from a fairly small corpus. The model exploits a symbolic text planning module to produce text plans, thus reducing the demand of data; a neural surface realization module then generates fluent text conditioned on the text plan. Human evaluation showed that our model outperforms various baselines by a wide margin and generates stories which are fluent as well as globally coherent.

The perspective of being able to crowd-source coherence relations bears the promise of acquiring annotations for new texts quickly, which could then increase the size and variety of discourse-annotated corpora. It would also open the avenue to answering new research questions: Collecting annotations from a larger number of individuals per instance would allow to investigate the distribution of inferred relations, and to study individual differences in coherence relation interpretation. However, annotating coherence relations with untrained workers is not trivial. We here propose a novel two-step annotation procedure, which extends an earlier method by Scholman and Demberg (2017a). In our approach, coherence relation labels are inferred from connectives that workers insert into the text. We show that the proposed method leads to replicable coherence annotations, and analyse the agreement between the obtained relation labels and annotations from PDTB and RSTDT on the same texts.

A number of different discourse connectives can be used to mark the same discourse relation, but it is unclear what factors affect connective choice. One recent account is the Rational Speech Acts theory, which predicts that speakers try to maximize the informativeness of an utterance such that the listener can interpret the intended meaning correctly. Existing prior work uses referential language games to test the rational account of speakers’ production of concrete meanings, such as identification of objects within a picture. Building on the same paradigm, we design a novel Discourse Continuation Game to investigate speakers’ production of abstract discourse relations. Experimental results reveal that speakers significantly prefer a more informative connective, in line with predictions of the RSA model.

We evaluate corpus-based measures of linguistic complexity obtained using Universal Dependencies (UD) treebanks. We propose a method of estimating robustness of the complexity values obtained using a given measure and a given treebank. The results indicate that measures of syntactic complexity might be on average less robust than those of morphological complexity. We also estimate the validity of complexity measures by comparing the results for very similar languages and checking for unexpected differences. We show that some of those differences that arise can be diminished by using parallel treebanks and, more importantly from the practical point of view, by harmonizing the language-specific solutions in the UD annotation.

Developing conventional natural language generation systems requires extensive attention from human experts in order to craft complex sets of sentence planning rules. We propose a Bayesian nonparametric approach to learn sentence planning rules by inducing synchronous tree substitution grammars for pairs of text plans and morphosyntactically-specified dependency trees. Our system is able to learn rules which can be used to generate novel texts after training on small datasets.

Human world knowledge contains information about prototypical events and their participants and locations. In this paper, we train the first models using multi-task learning that can both predict missing event participants and also perform semantic role classification based on semantic plausibility. Our best-performing model is an improvement over the previous state-of-the-art on thematic fit modelling tasks. The event embeddings learned by the model can additionally be used effectively in an event similarity task, also outperforming the state-of-the-art.

Implicit discourse relation recognition is an extremely challenging task due to the lack of indicative connectives. Various neural network architectures have been proposed for this task recently, but most of them suffer from the shortage of labeled data. In this paper, we address this problem by procuring additional training data from parallel corpora: When humans translate a text, they sometimes add connectives (a process known as explicitation). We automatically back-translate it into an English connective and use it to infer a label with high confidence. We show that a training set several times larger than the original training set can be generated this way. With the extra labeled instances, we show that even a simple bidirectional Long Short-Term Memory Network can outperform the current state-of-the-art.

Inferring implicit discourse relations in natural language text is the most difficult subtask in discourse parsing. Many neural network models have been proposed to tackle this problem. However, the comparison for this task is not unified, so we could hardly draw clear conclusions about the effectiveness of various architectures. Here, we propose neural network models that are based on feedforward and long-short term memory architecture and systematically study the effects of varying structures. To our surprise, the best-configured feedforward architecture outperforms LSTM-based model in most cases despite thorough tuning. Further, we compare our best feedforward system with competitive convolutional and recurrent networks and find that feedforward can actually be more effective. For the first time for this task, we compile and publish outputs from previous neural and non-neural systems to establish the standard for further comparison.

While previous research on readability has typically focused on document-level measures, recent work in areas such as natural language generation has pointed out the need of sentence-level readability measures. Much of psycholinguistics has focused for many years on processing measures that provide difficulty estimates on a word-by-word basis. However, these psycholinguistic measures have not yet been tested on sentence readability ranking tasks. In this paper, we use four psycholinguistic measures: idea density, surprisal, integration cost, and embedding depth to test whether these features are predictive of readability levels. We find that psycholinguistic features significantly improve performance by up to 3 percentage points over a standard document-level readability metric baseline.

The task of implicit discourse relation classification has received increased attention in recent years, including two CoNNL shared tasks on the topic. Existing machine learning models for the task train on sections 2-21 of the PDTB and test on section 23, which includes a total of 761 implicit discourse relations. In this paper, we’d like to make a methodological point, arguing that the standard test set is too small to draw conclusions about whether the inclusion of certain features constitute a genuine improvement, or whether one got lucky with some properties of the test set, and argue for the adoption of cross validation for the discourse relation classification task by the community.

Recent research in psycholinguistics has provided increasing evidence that humans predict upcoming content. Prediction also affects perception and might be a key to robustness in human language processing. In this paper, we investigate the factors that affect human prediction by building a computational model that can predict upcoming discourse referents based on linguistic knowledge alone vs. linguistic knowledge jointly with common-sense knowledge in the form of scripts. We find that script knowledge significantly improves model estimates of human predictions. In a second study, we test the highly controversial hypothesis that predictability influences referring expression type but do not find evidence for such an effect.

Traditional discourse annotation tasks are considered costly and time-consuming, and the reliability and validity of these tasks is in question. In this paper, we investigate whether crowdsourcing can be used to obtain reliable discourse relation annotations. We also examine the influence of context on the reliability of the data. The results of a crowdsourced connective insertion task showed that the method can be used to obtain reliable annotations: The majority of the inserted connectives converged with the original label. Further, the method is sensitive to the fact that multiple senses can often be inferred for a single relation. Regarding the presence of context, the results show no significant difference in distributions of insertions between conditions overall. However, a by-item comparison revealed several characteristics of segments that determine whether the presence of context makes a difference in annotations. The findings discussed in this paper can be taken as evidence that crowdsourcing can be used as a valuable method to obtain insights into the sense(s) of relations.

Corpora of referring expressions elicited from human participants in a controlled environment are an important resource for research on automatic referring expression generation. We here present G-TUNA, a new corpus of referring expressions for German. Using the furniture stimuli set developed for the TUNA and D-TUNA corpora, our corpus extends on these corpora by providing data collected in a simulated driving dual-task setting, and additionally provides exact duration annotations for the spoken referring expressions. This corpus will hence allow researchers to analyze the interaction between referring expression length and speech rate, under conditions where the listener is under high vs. low cognitive load.

In discourse relation annotation, there is currently a variety of different frameworks being used, and most of them have been developed and employed mostly on written data. This raises a number of questions regarding interoperability of discourse relation annotation schemes, as well as regarding differences in discourse annotation for written vs. spoken domains. In this paper, we describe ouron annotating two spoken domains from the SPICE Ireland corpus (telephone conversations and broadcast interviews) according todifferent discourse annotation schemes, PDTB 3.0 and CCR. We show that annotations in the two schemes can largely be mappedone another, and discuss differences in operationalisations of discourse relation schemes which present a challenge to automatic mapping. We also observe systematic differences in the prevalence of implicit discourse relations in spoken data compared to written texts,find that there are also differences in the types of causal relations between the domains. Finally, we find that PDTB 3.0 addresses many shortcomings of PDTB 2.0 wrt. the annotation of spoken discourse, and suggest further extensions. The new corpus has roughly theof the CoNLL 2015 Shared Task test set, and we hence hope that it will be a valuable resource for the evaluation of automatic discourse relation labellers.

The search space in grammar-based natural language generation tasks can get very large, which is particularly problematic when generating long utterances or paragraphs. Using surface realization with OpenCCG as an example, we show that we can effectively detect partial solutions (edges) which cannot ultimately be part of a complete sentence because of their syntactic category. Formulating the completion of an edge into a sentence as finding a solution path in a large state-transition system, we demonstrate a connection to AI Planning which is concerned with this kind of problem. We design a compilation from OpenCCG into AI Planning allowing the detection of infeasible edges via AI Planning dead-end detection methods (proving the absence of a solution to the compilation). Our experiments show that this can filter out large fractions of infeasible edges in, and thus benefit the performance of, complex realization processes.

We present a treebank and lexicon for German and English, which have been developed for PLTAG parsing. PLTAG is a psycholinguistically motivated, incremental version of tree-adjoining grammar (TAG). The resources are however also applicable to parsing with other variants of TAG. The German PLTAG resources are based on the TIGER corpus and, to the best of our knowledge, constitute the first scalable German TAG grammar. The English PLTAG resources go beyond existing resources in that they include the NP annotation by (Vadas and Curran, 2007), and include the prediction lexicon necessary for PLTAG.