This paper presents teaching materials, particularly assignments and ideas for classroom activities, from a new course on large language modelsThe assignments include experiments with LLM inference for weather report generation and machine translation.The classroom activities include class quizzes, focused research on downstream tasks and datasets, and an interactive “best paper” session aimed at reading and comprehension of research papers.
Linguistic entrainment, or alignment, represents a phenomenon where linguistic patterns employed by conversational participants converge to one another. While entrainment has been shown to produce a more natural user experience, most dialogue systems do not have any provisions for it. In this work, we introduce methods for achieving dialogue entrainment in a GPT-2-based end-to-end task-oriented dialogue system through the utilization of shared vocabulary. We experiment with training instance weighting, entrainment-specific loss, and additional conditioning to generate responses that align with the user. We demonstrate that all three approaches produce significantly better entrainment than the base, non-entrainment-optimized model, as confirmed by both automated and manual evaluation metrics.
Existing explanation methods for image classification struggle to provide faithful and plausible explanations. This paper addresses this issue by proposing a post-hoc natural language explanation method that can be applied to any CNN-based classifier without altering its training process or affecting predictive performance. By analysing influential neurons and the corresponding activation maps, the method generates a faithful description of the classifier’s decision process in the form of a structured meaning representation, which is then converted into text by a language model. Through this pipeline approach, the generated explanations are grounded in the neural network architecture, providing accurate insight into the classification process while remaining accessible to non-experts. Experimental results show that the NLEs constructed by our method are significantly more plausible and faithful than baselines. In particular, user interventions in the neural network structure (masking of neurons) are three times more effective.
Large Language Models (LLMs) are being employed by end-users for various tasks, including sensitive ones such as health counseling, disregarding potential safety concerns. It is thus necessary to understand how adequately LLMs perform in such domains. We conduct a case study on ChatGPT in nutrition counseling, a popular use-case where the model supports a user with their dietary struggles. We crowd-source real-world diet-related struggles, then work with nutrition experts to generate supportive text using ChatGPT. Finally, experts evaluate the safety and text quality of ChatGPT’s output. The result is the HAI-coaching dataset, containing ~2.4K crowdsourced dietary struggles and ~97K corresponding ChatGPT-generated and expert-annotated supportive texts. We analyse ChatGPT’s performance, discovering potentially harmful behaviours, especially for sensitive topics like mental health. Finally, we use HAI-coaching to test open LLMs on various downstream tasks, showing that even the latest models struggle to achieve good performance. HAI-coaching is available at https://github.com/uccollab/hai-coaching/
Text style transfer (TST) involves altering the linguistic style of a text while preserving its style-independent content. This paper focuses on sentiment transfer, a popular TST subtask, across a spectrum of Indian languages: Hindi, Magahi, Malayalam, Marathi, Punjabi, Odia, Telugu, and Urdu, expanding upon previous work on English-Bangla sentiment transfer. We introduce dedicated datasets of 1,000 positive and 1,000 negative style-parallel sentences for each of these eight languages. We then evaluate the performance of various benchmark models categorized into parallel, non-parallel, cross-lingual, and shared learning approaches, including the Llama2 and GPT-3.5 large language models (LLMs). Our experiments highlight the significance of parallel data in TST and demonstrate the effectiveness of the Masked Style Filling (MSF) approach in non-parallel techniques. Moreover, cross-lingual and joint multilingual learning methods show promise, offering insights into selecting optimal models tailored to the specific language and task requirements. To the best of our knowledge, this work represents the first comprehensive exploration of the TST task as sentiment transfer across a diverse set of languages.
We analyze the performance of large language models (LLMs) on Text Style Transfer (TST), specifically focusing on sentiment transfer and text detoxification across three languages: English, Hindi, and Bengali. Text Style Transfer involves modifying the linguistic style of a text while preserving its core content. We evaluate the capabilities of pre-trained LLMs using zero-shot and few-shot prompting as well as parameter-efficient finetuning on publicly available datasets. Our evaluation using automatic metrics, GPT-4 and human evaluations reveals that while some prompted LLMs perform well in English, their performance in on other languages (Hindi, Bengali) remains average. However, finetuning significantly improves results compared to zero-shot and few-shot prompting, making them comparable to previous state-of-the-art. This underscores the necessity of dedicated datasets and specialized models for effective TST.
Automatic metrics are extensively used to evaluate Natural Language Processing systems. However, there has been increasing focus on how the are used and reported by practitioners within the field. In this paper, we have conducted a survey on the use of automatic metrics, focusing particularly on natural language generation tasks. We inspect which metrics are used as well as why they are chosen and how their use is reported. Our findings from this survey reveal significant shortcomings, including inappropriate metric usage, lack of implementation details and missing correlations with human judgements. We conclude with recommendations that we believe authors should follow to enable more rigour within the field.
We introduce a simple approach that uses a large language model (LLM) to automatically implement a fully interpretable rule-based data-to-text system in pure Python. Experimental evaluation on the WebNLG dataset showed that such a constructed system produces text of better quality (according to the BLEU and BLEURT metrics) than the same LLM prompted to directly produce outputs, and produces fewer hallucinations than a BART language model fine-tuned on the same data. Furthermore, at runtime, the approach generates text in a fraction of the processing time required by neural approaches, using only a single CPU.
We present ‘factgenie‘: a framework for annotating and visualizing word spans in textual model outputs. Annotations can capture various span-based phenomena such as semantic inaccuracies or irrelevant text. With ‘factgenie‘, the annotations can be collected both from human crowdworkers and large language models. Our framework consists of a web interface for data visualization and gathering text annotations, powered by an easily extensible codebase.
Natural Language Processing (NLP) research is increasingly focusing on the use of Large Language Models (LLMs), with some of the most popular ones being either fully or partially closed-source. The lack of access to model details, especially regarding training data, has repeatedly raised concerns about data contamination among researchers. Several attempts have been made to address this issue, but they are limited to anecdotal evidence and trial and error. Additionally, they overlook the problem of indirect data leaking, where modelsare iteratively improved by using data coming from users. In this work, we conduct the first systematic analysis of work using OpenAI’s GPT-3.5 and GPT-4, the most prominently used LLMs today, in the context of data contamination. By analysing 255 papers and considering OpenAI’s data usage policy, we extensively document the amount of data leaked to these models during the first year after the model’s release. We report that these models have been globally exposed to ∼4.7M samples from 263 benchmarks. At the same time, we document a number of evaluation malpractices emerging in the reviewed papers, such as unfair or missing baseline comparisons and reproducibility issues. We release our results as a collaborative project on https://leak-llm.github.io/, where other researchers can contribute to our efforts.
In this paper, we describe several reproductions of a human evaluation experiment measuring the quality of automatic dialogue summarization (Feng et al., 2021). We investigate the impact of the annotators’ highest level of education, field of study, and native language on the evaluation of the informativeness of the summary. We find that the evaluation is relatively consistent regardless of these factors, but the biggest impact seems to be a prior specific background in natural language processing (as opposed to, e.g. a background in computer sci- ence). We also find that the experiment setup (asking for single vs. multiple criteria) may have an impact on the results.
We analyze the behaviors of open large language models (LLMs) on the task of data-to-text (D2T) generation, i.e., generating coherent and relevant text from structured data. To avoid the issue of LLM training data contamination with standard benchmarks, we design Quintd - a tool for collecting novel structured data records from public APIs. We find that open LLMs (Llama 2, Mistral, and Zephyr) can generate fluent and coherent texts in zero-shot settings from data in common formats collected with Quintd. However, we show that the semantic accuracy of the outputs is a major issue: both according to human annotators and our reference-free metric based on GPT-4, more than 80% of the outputs of open LLMs contain at least one semantic error. We publicly release the code, data, and model outputs.
Heterogenity of data-to-text generation datasets limits the research on data-to-text generation systems. We present TabGenie – a toolkit which enables researchers to explore, preprocess, and analyze a variety of data-to-text generation datasets through the unified framework of table-to-text generation. In TabGenie, all inputs are represented as tables with associated metadata. The tables can be explored through a web interface, which also provides an interactive mode for debugging table-to-text generation, facilitates side-by-side comparison of generated system outputs, and allows easy exports for manual analysis. Furthermore, TabGenie is equipped with command line processing tools and Python bindings for unified dataset loading and processing. We release TabGenie as a PyPI package and provide its open-source code and a live demo at https://github.com/kasnerz/tabgenie.
Instruction-finetuned large language models (LLMs) gained a huge popularity recently, thanks to their ability to interact with users through conversation. In this work, we aim to evaluate their ability to complete multi-turn tasks and interact with external databases in the context of established task-oriented dialogue benchmarks. We show that in explicit belief state tracking, LLMs underperform compared to specialized task-specific models. Nevertheless, they show some ability to guide the dialogue to a successful ending through their generated responses if they are provided with correct slot values. Furthermore, this ability improves with few-shot in-domain examples.
Text style transfer (TST) involves modifying the linguistic style of a given text while retaining its core content. This paper addresses the challenging task of text style transfer in the Bangla language, which is low-resourced in this area. We present a novel Bangla dataset that facilitates text sentiment transfer, a subtask of TST, enabling the transformation of positive sentiment sentences to negative and vice versa. To establish a high-quality base for further research, we refined and corrected an existing English dataset of 1,000 sentences for sentiment transfer based on Yelp reviews, and we introduce a new human-translated Bangla dataset that parallels its English counterpart. Furthermore, we offer multiple benchmark models that serve as a validation of the dataset and baseline for further research.
In this paper, we present UFAL-ULD team’s system, desinged as a part of the BLP Shared Task 1: Violence Inciting Text Detection (VITD). This task aims to classify text, with a particular challenge of identifying incitement to violence into Direct, Indirect or Non-violence levels. We experimented with several pre-trained sequence classification models, including XLM-RoBERTa, BanglaBERT, Bangla BERT Base, and Multilingual BERT. Our best-performing model was based on the XLM-RoBERTa-base architecture, which outperformed the baseline models. Our system was ranked 20th among the 27 teams that participated in the task.
In this paper, we present the UFAL-ULD team’s system for the BLP Shared Task 2: Sentiment Analysis of Bangla Social Media Posts. The Task 2 involves classifying text into Positive, Negative, or Neutral sentiments. As a part of this task, we conducted a series of experiments with several pre-trained sequence classification models – XLM-RoBERTa, BanglaBERT, Bangla BERT Base and Multilingual BERT. Among these, our best-performing model was based on the XLM-RoBERTa-base architecture, which outperforms baseline models. Our system was ranked 19th among the 30 teams that participated in the task.
Hallucination of text ungrounded in the input is a well-known problem in neural data-to-text generation. Many methods have been proposed to mitigate it, but they typically require altering model architecture or collecting additional data, and thus cannot be easily applied to an existing model. In this paper, we explore a new way to mitigate hallucinations by combining the probabilistic output of a generator language model (LM) with the output of a special “text critic” classifier, which guides the generation by assessing the match between the input data and the text generated so far. Our method does not need any changes to the underlying LM’s architecture or training procedure and can thus be combined with any model and decoding operating on word probabilities. The critic does not need any additional training data, using the base LM’s training data and synthetic negative examples. Our experimental results show that our method improves over the baseline on the WebNLG and OpenDialKG benchmarks.
Pretrained language models (PLMs) for data-to-text (D2T) generation can use human-readable data labels such as column headings, keys, or relation names to generalize to out-of-domain examples. However, the models are well-known in producing semantically inaccurate outputs if these labels are ambiguous or incomplete, which is often the case in D2T datasets. In this paper, we expose this issue on the task of descibing a relation between two entities. For our experiments, we collect a novel dataset for verbalizing a diverse set of 1,522 unique relations from three large-scale knowledge graphs (Wikidata, DBPedia, YAGO). We find that although PLMs for D2T generation expectedly fail on unclear cases, models trained with a large variety of relation labels are surprisingly robust in verbalizing novel, unseen relations. We argue that using data with a diverse set of clear and meaningful labels is key to training D2T generation systems capable of generalizing to novel domains.
Generating polite responses is essential to build intelligent and engaging dialogue systems. However, this task is far from well-explored due to the difficulties of rendering a particular style in coherent responses, especially when parallel datasets for regular-to-polite pairs are usually unavailable. This paper proposes a polite chatbot that can produce responses that are polite and coherent to the given context. In this study, a politeness transfer model is first used to generate polite synthetic dialogue pairs of contexts and polite utterances. Then, these synthetic pairs are employed to train a dialogue model. Automatic and human evaluations demonstrate that our method outperforms baselines in producing polite dialogue responses while staying competitive in terms of coherent to the given context.
This work presents our efforts to reproduce the results of the human evaluation experiment presented in the paper of Vamvas and Sennrich (2022), which evaluated an automatic system detecting over- and undertranslations (translations containing more or less information than the original) in machine translation (MT) outputs. Despite the high quality of the documentation and code provided by the authors, we discuss some problems we found in reproducing the exact experimental setup and offer recommendations for improving reproducibility. Our replicated results generally confirm the conclusions of the original study, but in some cases statistically significant differences were observed, suggesting a high variability of human annotation.
Earlier research has shown that few studies in Natural Language Generation (NLG) evaluate their system outputs using an error analysis, despite known limitations of automatic evaluation metrics and human ratings. This position paper takes the stance that error analyses should be encouraged, and discusses several ways to do so. This paper is based on our shared experience as authors as well as a survey we distributed as a means of public consultation. We provide an overview of existing barriers to carrying out error analyses, and propose changes to improve error reporting in the NLG literature.
Data augmentation is an important method for evaluating the robustness of and enhancing the diversity of training data for natural language processing (NLP) models. In this paper, we present NL-Augmenter, a new participatory Python-based natural language (NL) augmentation framework which supports the creation of transformations (modifications to the data) and filters (data splits according to specific features). We describe the framework and an initial set of 117 transformations and 23 filters for a variety of NL tasks annotated with noisy descriptive tags. The transformations incorporate noise, intentional and accidental human mistakes, socio-linguistic variation, semantically-valid style, syntax changes, as well as artificial constructs that are unambiguous to humans. We demonstrate the efficacy of NL-Augmenter by using its transformations to analyze the robustness of popular language models. We find different models to be differently challenged on different tasks, with quasi-systematic score decreases. The infrastructure, datacards, and robustness evaluation results are publicly available on GitHub for the benefit of researchers working on paraphrase generation, robustness analysis, and low-resource NLP.
This paper focuses on text detoxification, i.e., automatically converting toxic text into nontoxic text. This task contributes to safer and more respectful online communication and can be considered a Text Style Transfer (TST) task, where the text’s style changes while its content is preserved. We present three approaches: (i) knowledge transfer from a similar task (ii) multi-task learning approach, combining sequence-to-sequence modeling with various toxicity classification tasks, and (iii) delete and reconstruct approach. To support our research, we utilize a dataset provided by Dementieva et al. (2021), which contains multiple versions of detoxified texts corresponding to toxic texts. In our experiments, we selected the best variants through expert human annotators, creating a dataset where each toxic sentence is paired with a single, appropriate detoxified version. Additionally, we introduced a small Hindi parallel dataset, aligning with a part of the English dataset, suitable for evaluation purposes. Our results demonstrate that our approach effectively balances text detoxification while preserving the actual content and maintaining fluency.
We report our efforts in identifying a set of previous human evaluations in NLP that would be suitable for a coordinated study examining what makes human evaluations in NLP more/less reproducible. We present our results and findings, which include that just 13% of papers had (i) sufficiently low barriers to reproduction, and (ii) enough obtainable information, to be considered for reproduction, and that all but one of the experiments we selected for reproduction was discovered to have flaws that made the meaningfulness of conducting a reproduction questionable. As a result, we had to change our coordinated study design from a reproduce approach to a standardise-then-reproduce-twice approach. Our overall (negative) finding that the great majority of human evaluations in NLP is not repeatable and/or not reproducible and/or too flawed to justify reproduction, paints a dire picture, but presents an opportunity for a rethink about how to design and report human evaluations in NLP.
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.
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.
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.
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.
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.
We present a free online demo of THEaiTRobot, an open-source bilingual tool for interactively generating theatre play scripts, in two versions. THEaiTRobot 1.0 uses the GPT-2 language model with minimal adjustments. THEaiTRobot 2.0 uses two models created by fine-tuning GPT-2 on purposefully collected and processed datasets and several other components, generating play scripts in a hierarchical fashion (title → synopsis → script). The underlying tool is used in the THEaiTRE project to generate scripts for plays, which are then performed on stage by a professional theatre.
In this paper, we present the results of two reproduction studies for the human evaluation originally reported by Dušek and Kasner (2020) in which the authors comparatively evaluated outputs produced by a semantic error detection system for data-to-text generation against reference outputs. In the first reproduction, the original evaluators repeat the evaluation, in a test of the repeatability of the original evaluation. In the second study, two new evaluators carry out the evaluation task, in a test of the reproducibility of the original evaluation under otherwise identical conditions. We describe our approach to reproduction, and present and analyse results, finding different degrees of reproducibility depending on result type, data and labelling task. Our resources are available and open-sourced.
In data-to-text (D2T) generation, training on in-domain data leads to overfitting to the data representation and repeating training data noise. We examine how to avoid finetuning pretrained language models (PLMs) on D2T generation datasets while still taking advantage of surface realization capabilities of PLMs. Inspired by pipeline approaches, we propose to generate text by transforming single-item descriptions with a sequence of modules trained on general-domain text-based operations: ordering, aggregation, and paragraph compression. We train PLMs for performing these operations on a synthetic corpus WikiFluent which we build from English Wikipedia. Our experiments on two major triple-to-text datasets—WebNLG and E2E—show that our approach enables D2T generation from RDF triples in zero-shot settings.
We introduce AARGH, an end-to-end task-oriented dialog system combining retrieval and generative approaches in a single model, aiming at improving dialog management and lexical diversity of outputs. The model features a new response selection method based on an action-aware training objective and a simplified single-encoder retrieval architecture which allow us to build an end-to-end retrieval-enhanced generation model where retrieval and generation share most of the parameters. On the MultiWOZ dataset, we show that our approach produces more diverse outputs while maintaining or improving state tracking and context-to-response generation performance, compared to state-of-the-art baselines.
Evaluations in machine learning rarely use the latest metrics, datasets, or human evaluation in favor of remaining compatible with prior work. The compatibility, often facilitated through leaderboards, thus leads to outdated but standardized evaluation practices. We pose that the standardization is taking place in the wrong spot. Evaluation infrastructure should enable researchers to use the latest methods and what should be standardized instead is how to incorporate these new evaluation advances. We introduce GEMv2, the new version of the Generation, Evaluation, and Metrics Benchmark which uses a modular infrastructure for dataset, model, and metric developers to benefit from each other’s work. GEMv2 supports 40 documented datasets in 51 languages, ongoing online evaluation for all datasets, and our interactive tools make it easier to add new datasets to the living benchmark.
Every model is only as strong as the data that it is trained on. In this paper, we present a new dataset, obtained by merging four publicly available annotated corpora for task-oriented dialogues in several domains (MultiWOZ 2.2, CamRest676, DSTC2 and Schema-Guided Dialogue Dataset). This way, we assess the feasibility of providing a unified ontology and annotation schema covering several domains with a relatively limited effort. We analyze the characteristics of the resulting dataset along three main dimensions: language, information content and performance. We focus on aspects likely to be pertinent for improving dialogue success, e.g. dialogue consistency. Furthermore, to assess the usability of this new corpus, we thoroughly evaluate dialogue generation performance under various conditions with the help of two prominent recent end-to-end dialogue models: MarCo and GPT-2. These models were selected as popular open implementations representative of the two main dimensions of dialogue modelling. While we did not observe a significant gain for dialogue state tracking performance, we show that using more training data from different sources can improve language modelling capabilities and positively impact dialogue flow (consistency). In addition, we provide the community with one of the largest open dataset for machine learning experiments.
We present a novel architecture for explainable modeling of task-oriented dialogues with discrete latent variables to represent dialogue actions. Our model is based on variational recurrent neural networks (VRNN) and requires no explicit annotation of semantic information. Unlike previous works, our approach models the system and user turns separately and performs database query modeling, which makes the model applicable to task-oriented dialogues while producing easily interpretable action latent variables. We show that our model outperforms previous approaches with less supervision in terms of perplexity and BLEU on three datasets, and we propose a way to measure dialogue success without the need for expert annotation. Finally, we propose a novel way to explain semantics of the latent variables with respect to system actions.
We experiment with adapting generative language models for the generation of long coherent narratives in the form of theatre plays. Since fully automatic generation of whole plays is not currently feasible, we created an interactive tool that allows a human user to steer the generation somewhat while minimizing intervention. We pursue two approaches to long-text generation: a flat generation with summarization of context, and a hierarchical text-to-text two-stage approach, where a synopsis is generated first and then used to condition generation of the final script. Our preliminary results and discussions with theatre professionals show improvements over vanilla language model generation, but also identify important limitations of our approach.
We present AggGen (pronounced ‘again’) a data-to-text model which re-introduces two explicit sentence planning stages into neural data-to-text systems: input ordering and input aggregation. In contrast to previous work using sentence planning, our model is still end-to-end: AggGen performs sentence planning at the same time as generating text by learning latent alignments (via semantic facts) between input representation and target text. Experiments on the WebNLG and E2E challenge data show that by using fact-based alignments our approach is more interpretable, expressive, robust to noise, and easier to control, while retaining the advantages of end-to-end systems in terms of fluency. Our code is available at https://github.com/XinnuoXu/AggGen.
Task-oriented dialogue systems typically require manual annotation of dialogue slots in training data, which is costly to obtain. We propose a method that eliminates this requirement: We use weak supervision from existing linguistic annotation models to identify potential slot candidates, then automatically identify domain-relevant slots by using clustering algorithms. Furthermore, we use the resulting slot annotation to train a neural-network-based tagger that is able to perform slot tagging with no human intervention. This tagger is trained solely on the outputs of our method and thus does not rely on any labeled data. Our model demonstrates state-of-the-art performance in slot tagging without labeled training data on four different dialogue domains. Moreover, we find that slot annotations discovered by our model significantly improve the performance of an end-to-end dialogue response generation model, compared to using no slot annotation at all.
The MultiWOZ dataset (Budzianowski et al.,2018) is frequently used for benchmarkingcontext-to-response abilities of task-orienteddialogue systems. In this work, we identifyinconsistencies in data preprocessing and re-porting of three corpus-based metrics used onthis dataset, i.e., BLEU score and Inform &Success rates. We point out a few problemsof the MultiWOZ benchmark such as unsat-isfactory preprocessing, insufficient or under-specified evaluation metrics, or rigid database. We re-evaluate 7 end-to-end and 6 policy opti-mization models in as-fair-as-possible setups,and we show that their reported scores cannotbe directly compared. To facilitate compari-son of future systems, we release our stand-alone standardized evaluation scripts. We alsogive basic recommendations for corpus-basedbenchmarking in future works.
We introduce GEM, a living benchmark for natural language Generation (NLG), its Evaluation, and Metrics. Measuring progress in NLG relies on a constantly evolving ecosystem of automated metrics, datasets, and human evaluation standards. Due to this moving target, new models often still evaluate on divergent anglo-centric corpora with well-established, but flawed, metrics. This disconnect makes it challenging to identify the limitations of current models and opportunities for progress. Addressing this limitation, GEM provides an environment in which models can easily be applied to a wide set of tasks and in which evaluation strategies can be tested. Regular updates to the benchmark will help NLG research become more multilingual and evolve the challenge alongside models. This paper serves as the description of the data for the 2021 shared task at the associated GEM Workshop.
We observe a severe under-reporting of the different kinds of errors that Natural Language Generation systems make. This is a problem, because mistakes are an important indicator of where systems should still be improved. If authors only report overall performance metrics, the research community is left in the dark about the specific weaknesses that are exhibited by ‘state-of-the-art’ research. Next to quantifying the extent of error under-reporting, this position paper provides recommendations for error identification, analysis and reporting.
We present our Charles-UPF submission for the Shared Task on Evaluating Accuracy in Generated Texts at INLG 2021. Our system can detect the errors automatically using a combination of a rule-based natural language generation (NLG) system and pretrained language models (LMs). We first utilize a rule-based NLG system to generate sentences with facts that can be derived from the input. For each sentence we evaluate, we select a subset of facts which are relevant by measuring semantic similarity to the sentence in question. Finally, we finetune a pretrained language model on annotated data along with the relevant facts for fine-grained error detection. On the test set, we achieve 69% recall and 75% precision with a model trained on a mixture of human-annotated and synthetic data.
One of the most challenging aspects of current single-document news summarization is that the summary often contains ‘extrinsic hallucinations’, i.e., facts that are not present in the source document, which are often derived via world knowledge. This causes summarisation systems to act more like open-ended language models tending to hallucinate facts that are erroneous. In this paper, we mitigate this problem with the help of multiple supplementary resource documents assisting the task. We present a new dataset MiraNews and benchmark existing summarisation models. In contrast to multi-document summarization, which addresses multiple events from several source documents, we still aim at generating a summary for a single document. We show via data analysis that it’s not only the models which are to blame: more than 27% of facts mentioned in the gold summaries of MiraNews are better grounded on assisting documents than in the main source articles. An error analysis of generated summaries from pretrained models fine-tuned on MIRANEWS reveals that this has an even bigger effects on models: assisted summarisation reduces 55% of hallucinations when compared to single-document summarisation models trained on the main article only.
Attention-based pre-trained language models such as GPT-2 brought considerable progress to end-to-end dialogue modelling. However, they also present considerable risks for task-oriented dialogue, such as lack of knowledge grounding or diversity. To address these issues, we introduce modified training objectives for language model finetuning, and we employ massive data augmentation via back-translation to increase the diversity of the training data. We further examine the possibilities of combining data from multiples sources to improve performance on the target dataset. We carefully evaluate our contributions with both human and automatic methods. Our model substantially outperforms the baseline on the MultiWOZ data and shows competitive performance with state of the art in both automatic and human evaluation.
Définition et détection des incohérences du système dans les dialogues orientés tâche. Nous présentons des expériences sur la détection automatique des comportements incohérents des systèmes de dialogues orientés tâche à partir du contexte. Nous enrichissons les données bAbI/DSTC2 (Bordes et al., 2017) avec une annotation automatique des incohérences de dialogue, et nous démontrons que les incohérences sont en corrélation avec les dialogues ratés. Nous supposons que l’utilisation d’un historique de dialogue limité et la prédiction du prochain tour de l’utilisateur peuvent améliorer la classification des incohérences. Si les deux hypothèses sont confirmées pour un modèle de dialogue basé sur les réseaux de mémoire, elles ne le sont pas pour un entraînement basé sur le modèle de langage GPT-2, qui bénéficie le plus de l’utilisation de l’historique complet du dialogue et obtient un score de précision de 0,99.
We describe our system for the RDF-to-text generation task of the WebNLG Challenge 2020. We base our approach on the mBART model, which is pre-trained for multilingual denoising. This allows us to use a simple, identical, end-to-end setup for both English and Russian. Requiring minimal taskor languagespecific effort, our model placed in the first third of the leaderboard for English and first or second for Russian on automatic metrics, and it made it into the best or second-best system cluster on human evaluation.
We present our submission to the Simultaneous Translation And Paraphrase for Language Education (STAPLE) challenge. We used a standard Transformer model for translation, with a crosslingual classifier predicting correct translations on the output n-best list. To increase the diversity of the outputs, we used additional data to train the translation model, and we trained a paraphrasing model based on the Levenshtein Transformer architecture to generate further synonymous translations. The paraphrasing results were again filtered using our classifier. While the use of additional data and our classifier filter were able to improve results, the paraphrasing model produced too many invalid outputs to further improve the output quality. Our model without the paraphrasing component finished in the middle of the field for the shared task, improving over the best baseline by a margin of 10-22 % weighted F1 absolute.
We present a novel approach to data-to-text generation based on iterative text editing. Our approach maximizes the completeness and semantic accuracy of the output text while leveraging the abilities of recent pre-trained models for text editing (LaserTagger) and language modeling (GPT-2) to improve the text fluency. To this end, we first transform data items to text using trivial templates, and then we iteratively improve the resulting text by a neural model trained for the sentence fusion task. The output of the model is filtered by a simple heuristic and reranked with an off-the-shelf pre-trained language model. We evaluate our approach on two major data-to-text datasets (WebNLG, Cleaned E2E) and analyze its caveats and benefits. Furthermore, we show that our formulation of data-to-text generation opens up the possibility for zero-shot domain adaptation using a general-domain dataset for sentence fusion.
A major challenge in evaluating data-to-text (D2T) generation is measuring the semantic accuracy of the generated text, i.e. checking if the output text contains all and only facts supported by the input data. We propose a new metric for evaluating the semantic accuracy of D2T generation based on a neural model pretrained for natural language inference (NLI). We use the NLI model to check textual entailment between the input data and the output text in both directions, allowing us to reveal omissions or hallucinations. Input data are converted to text for NLI using trivial templates. Our experiments on two recent D2T datasets show that our metric can achieve high accuracy in identifying erroneous system outputs.
Abstractive summarisation is notoriously hard to evaluate since standard word-overlap-based metrics are insufficient. We introduce a new evaluation metric which is based on fact-level content weighting, i.e. relating the facts of the document to the facts of the summary. We fol- low the assumption that a good summary will reflect all relevant facts, i.e. the ones present in the ground truth (human-generated refer- ence summary). We confirm this hypothe- sis by showing that our weightings are highly correlated to human perception and compare favourably to the recent manual highlight- based metric of Hardy et al. (2019).
We present the first complete spoken dialogue system driven by a multiimensional statistical dialogue manager. This framework has been shown to substantially reduce data needs by leveraging domain-independent dimensions, such as social obligations or feedback, which (as we show) can be transferred between domains. In this paper, we conduct a user study and show that the performance of a multi-dimensional system, which can be adapted from a source domain, is equivalent to that of a one-dimensional baseline, which can only be trained from scratch.
We present a recurrent neural network based system for automatic quality estimation of natural language generation (NLG) outputs, which jointly learns to assign numerical ratings to individual outputs and to provide pairwise rankings of two different outputs. The latter is trained using pairwise hinge loss over scores from two copies of the rating network. We use learning to rank and synthetic data to improve the quality of ratings assigned by our system: We synthesise training pairs of distorted system outputs and train the system to rank the less distorted one higher. This leads to a 12% increase in correlation with human ratings over the previous benchmark. We also establish the state of the art on the dataset of relative rankings from the E2E NLG Challenge (Dusek et al., 2019), where synthetic data lead to a 4% accuracy increase over the base model.
Neural natural language generation (NNLG) systems are known for their pathological outputs, i.e. generating text which is unrelated to the input specification. In this paper, we show the impact of semantic noise on state-of-the-art NNLG models which implement different semantic control mechanisms. We find that cleaned data can improve semantic correctness by up to 97%, while maintaining fluency. We also find that the most common error is omitting information, rather than hallucination.
We present the first dataset targeted at end-to-end NLG in Czech in the restaurant domain, along with several strong baseline models using the sequence-to-sequence approach. While non-English NLG is under-explored in general, Czech, as a morphologically rich language, makes the task even harder: Since Czech requires inflecting named entities, delexicalization or copy mechanisms do not work out-of-the-box and lexicalizing the generated outputs is non-trivial. In our experiments, we present two different approaches to this this problem: (1) using a neural language model to select the correct inflected form while lexicalizing, (2) a two-step generation setup: our sequence-to-sequence model generates an interleaved sequence of lemmas and morphological tags, which are then inflected by a morphological generator.
Human evaluation for natural language generation (NLG) often suffers from inconsistent user ratings. While previous research tends to attribute this problem to individual user preferences, we show that the quality of human judgements can also be improved by experimental design. We present a novel rank-based magnitude estimation method (RankME), which combines the use of continuous scales and relative assessments. We show that RankME significantly improves the reliability and consistency of human ratings compared to traditional evaluation methods. In addition, we show that it is possible to evaluate NLG systems according to multiple, distinct criteria, which is important for error analysis. Finally, we demonstrate that RankME, in combination with Bayesian estimation of system quality, is a cost-effective alternative for ranking multiple NLG systems.
The overall objective of ‘social’ dialogue systems is to support engaging, entertaining, and lengthy conversations on a wide variety of topics, including social chit-chat. Apart from raw dialogue data, user-provided ratings are the most common signal used to train such systems to produce engaging responses. In this paper we show that social dialogue systems can be trained effectively from raw unannotated data. Using a dataset of real conversations collected in the 2017 Alexa Prize challenge, we developed a neural ranker for selecting ‘good’ system responses to user utterances, i.e. responses which are likely to lead to long and engaging conversations. We show that (1) our neural ranker consistently outperforms several strong baselines when trained to optimise for user ratings; (2) when trained on larger amounts of data and only using conversation length as the objective, the ranker performs better than the one trained using ratings – ultimately reaching a Precision@1 of 0.87. This advance will make data collection for social conversational agents simpler and less expensive in the future.
Multimodal search-based dialogue is a challenging new task: It extends visually grounded question answering systems into multi-turn conversations with access to an external database. We address this new challenge by learning a neural response generation system from the recently released Multimodal Dialogue (MMD) dataset (Saha et al., 2017). We introduce a knowledge-grounded multimodal conversational model where an encoded knowledge base (KB) representation is appended to the decoder input. Our model substantially outperforms strong baselines in terms of text-based similarity measures (over 9 BLEU points, 3 of which are solely due to the use of additional information from the KB).
In this work, we investigate the task of textual response generation in a multimodal task-oriented dialogue system. Our work is based on the recently released Multimodal Dialogue (MMD) dataset (Saha et al., 2017) in the fashion domain. We introduce a multimodal extension to the Hierarchical Recurrent Encoder-Decoder (HRED) model and show that this extension outperforms strong baselines in terms of text-based similarity metrics. We also showcase the shortcomings of current vision and language models by performing an error analysis on our system’s output.
This paper summarises the experimental setup and results of the first shared task on end-to-end (E2E) natural language generation (NLG) in spoken dialogue systems. Recent end-to-end generation systems are promising since they reduce the need for data annotation. However, they are currently limited to small, delexicalised datasets. The E2E NLG shared task aims to assess whether these novel approaches can generate better-quality output by learning from a dataset containing higher lexical richness, syntactic complexity and diverse discourse phenomena. We compare 62 systems submitted by 17 institutions, covering a wide range of approaches, including machine learning architectures – with the majority implementing sequence-to-sequence models (seq2seq) – as well as systems based on grammatical rules and templates.
We present three enhancements to existing encoder-decoder models for open-domain conversational agents, aimed at effectively modeling coherence and promoting output diversity: (1) We introduce a measure of coherence as the GloVe embedding similarity between the dialogue context and the generated response, (2) we filter our training corpora based on the measure of coherence to obtain topically coherent and lexically diverse context-response pairs, (3) we then train a response generator using a conditional variational autoencoder model that incorporates the measure of coherence as a latent variable and uses a context gate to guarantee topical consistency with the context and promote lexical diversity. Experiments on the OpenSubtitles corpus show a substantial improvement over competitive neural models in terms of BLEU score as well as metrics of coherence and diversity.
The majority of NLG evaluation relies on automatic metrics, such as BLEU . In this paper, we motivate the need for novel, system- and data-independent automatic evaluation methods: We investigate a wide range of metrics, including state-of-the-art word-based and novel grammar-based ones, and demonstrate that they only weakly reflect human judgements of system outputs as generated by data-driven, end-to-end NLG. We also show that metric performance is data- and system-specific. Nevertheless, our results also suggest that automatic metrics perform reliably at system-level and can support system development by finding cases where a system performs poorly.
This paper describes the E2E data, a new dataset for training end-to-end, data-driven natural language generation systems in the restaurant domain, which is ten times bigger than existing, frequently used datasets in this area. The E2E dataset poses new challenges: (1) its human reference texts show more lexical richness and syntactic variation, including discourse phenomena; (2) generating from this set requires content selection. As such, learning from this dataset promises more natural, varied and less template-like system utterances. We also establish a baseline on this dataset, which illustrates some of the difficulties associated with this data.
We describe experiments in Machine Translation using word sense disambiguation (WSD) information. This work focuses on WSD in verbs, based on two different approaches – verbal patterns based on corpus pattern analysis and verbal word senses from valency frames. We evaluate several options of using verb senses in the source-language sentences as an additional factor for the Moses statistical machine translation system. Our results show a statistically significant translation quality improvement in terms of the BLEU metric for the valency frames approach, but in manual evaluation, both WSD methods bring improvements.
We present a dataset of telephone conversations in English and Czech, developed for training acoustic models for automatic speech recognition (ASR) in spoken dialogue systems (SDSs). The data comprise 45 hours of speech in English and over 18 hours in Czech. Large part of the data, both audio and transcriptions, was collected using crowdsourcing, the rest are transcriptions by hired transcribers. We release the data together with scripts for data pre-processing and building acoustic models using the HTK and Kaldi ASR toolkits. We publish also the trained models described in this paper. The data are released under the CC-BY-SA 3.0 license, the scripts are licensed under Apache 2.0. In the paper, we report on the methodology of collecting the data, on the size and properties of the data, and on the scripts and their use. We verify the usability of the datasets by training and evaluating acoustic models using the presented data and scripts.
This paper presents development and test sets for machine translation of search queries in cross-lingual information retrieval in the medical domain. The data consists of the total of 1,508 real user queries in English translated to Czech, German, and French. We describe the translation and review process involving medical professionals and present a baseline experiment where our data sets are used for tuning and evaluation of a machine translation system.
We describe a method for semi-automatic extraction of Slovak multiword expressions (MWEs) from a dependency treebank. The process uses an automatic conversion from dependency syntactic trees to deep syntax and automatic tagging of verbal argument nodes based on a valency dictionary. Both the valency dictionary and the treebank conversion were adapted from the corresponding Czech versions; the automatically translated valency dictionary has been manually proofread and corrected. There are two main achievements – a valency dictionary of Slovak MWEs with direct links to corresponding expressions in the Czech dictionary, PDT-Vallex, and a method of extraction of MWEs from the Slovak Dependency Treebank. The extraction reached very high precision but lower recall in a manual evaluation. This is a work in progress, the overall goal of which is twofold: to create a Slovak language valency dictionary paralleling the Czech one, with bilingual links; and to use the extracted verbal frames in a collocation dictionary of Slovak verbs.
CzEng 1.0 is an updated release of our Czech-English parallel corpus, freely available for non-commercial research or educational purposes. In this release, we approximately doubled the corpus size, reaching 15 million sentence pairs (about 200 million tokens per language). More importantly, we carefully filtered the data to reduce the amount of non-matching sentence pairs. CzEng 1.0 is automatically aligned at the level of sentences as well as words. We provide not only the plain text representation, but also automatic morphological tags, surface syntactic as well as deep syntactic dependency parse trees and automatic co-reference links in both English and Czech. This paper describes key properties of the released resource including the distribution of text domains, the corpus data formats, and a toolkit to handle the provided rich annotation. We also summarize the procedure of the rich annotation (incl. co-reference resolution) and of the automatic filtering. Finally, we provide some suggestions on exploiting such an automatically annotated sentence-parallel corpus.