Natural Language Generation System Research Articles

YouTube Scholar: Advancing Education with AI Assistance

In the contemporary landscape of online educa- tion, the proliferation of educational content on platforms like YouTube has opened new avenues for learners. However, the sheer volume of videos presents a challenge, requiring innovative solutions. This project takes a pioneering approach by developing a specialized AI assistant for educational YouTube lectures, lever- aging advanced language models such as Generative Pretrained Transformers (GPT). This project focuses on the development of an AI assistant tailored specifically for educational YouTube lectures and courses. The primary goal is to enhance the learning experience by leveraging the capabilities of GPT-based models to provide personalised and contextually relevant interactions for learners. By harnessing the power of natural language understanding, generation, and dialog systems, the AI assistant aims to assist learners in navigating, understanding, and engaging with educa- tional content more effectively. The ultimate goal is to bridge the gap between the abundance of online educational content and learners’ comprehension, thereby revolutionizing the educational landscape in the digital age. By providing a sophisticated AI-driven companion, this project aspires to empower learners, making the educational journey more accessible, engaging, and conducive to deeper learning experiences. Through the integration of cutting-edge technologies, it envisions a future where learners can unlock the full potential of online education with confidence and ease

Human versus Machine Intelligence: Assessing Natural Language Generation Models through Complex Systems Theory.

The introduction of Transformer architectures - with the self-attention mechanism - in automatic Natural Language Generation (NLG) is a breakthrough in solving general task-oriented problems, such as the simple production of long text excerpts that resemble ones written by humans. While the performance of GPT-X architectures is there for all to see, many efforts are underway to penetrate the secrets of these black-boxes in terms of intelligent information processing whose output statistical distributions resemble that of natural language. In this work, through the complexity science framework, a comparative study of the stochastic processes underlying the texts produced by the English version of GPT-2 with respect to texts produced by human beings, notably novels in English and programming codes, is offered. The investigation, of a methodological nature, consists first of all of an analysis phase in which the Multifractal Detrended Fluctuation Analysis and the Recurrence Quantification Analysis - together with Zipf's law and approximate entropy - are adopted to characterize long-term correlations, regularities and recurrences in human and machine-produced texts. Results show several peculiarities and trends in terms of long-range correlations and recurrences in the last case. The synthesis phase, on the other hand, uses the complexity measures to build synthetic text descriptors - hence a suitable text embedding - which serve to constitute the features for feeding a machine learning system designed to operate feature selection through an evolutionary technique. Using multivariate analysis, it is then shown the grouping tendency of the three analyzed text types, allowing to place GTP-2 texts in between natural language texts and computer codes. Similarly, the classification task demonstrates that, given the high accuracy obtained in the automatic discrimination of text classes, the proposed set of complexity measures is highly informative. These interesting results allow us to add another piece to the theoretical understanding of the surprising results obtained by NLG systems based on deep learning and let us to improve the design of new informetrics or text mining systems for text classification, fake news detection, or even plagiarism detection.

Potential and Limitations of ChatGPT 3.5 and 4.0 as a Source of COVID-19 Information: Comprehensive Comparative Analysis of Generative and Authoritative Information.

The COVID-19 pandemic, caused by the SARS-CoV-2 virus, has necessitated reliable and authoritative information for public guidance. The World Health Organization (WHO) has been a primary source of such information, disseminating it through a question and answer format on its official website. Concurrently, ChatGPT 3.5 and 4.0, a deep learning-based natural language generation system, has shown potential in generating diverse text types based on user input. This study evaluates the accuracy of COVID-19 information generated by ChatGPT 3.5 and 4.0, assessing its potential as a supplementary public information source during the pandemic. We extracted 487 COVID-19-related questions from the WHO's official website and used ChatGPT 3.5 and 4.0 to generate corresponding answers. These generated answers were then compared against the official WHO responses for evaluation. Two clinical experts scored the generated answers on a scale of 0-5 across 4 dimensions-accuracy, comprehensiveness, relevance, and clarity-with higher scores indicating better performance in each dimension. The WHO responses served as the reference for this assessment. Additionally, we used the BERT (Bidirectional Encoder Representations from Transformers) model to generate similarity scores (0-1) between the generated and official answers, providing a dual validation mechanism. The mean (SD) scores for ChatGPT 3.5-generated answers were 3.47 (0.725) for accuracy, 3.89 (0.719) for comprehensiveness, 4.09 (0.787) for relevance, and 3.49 (0.809) for clarity. For ChatGPT 4.0, the mean (SD) scores were 4.15 (0.780), 4.47 (0.641), 4.56 (0.600), and 4.09 (0.698), respectively. All differences were statistically significant (P<.001), with ChatGPT 4.0 outperforming ChatGPT 3.5. The BERT model verification showed mean (SD) similarity scores of 0.83 (0.07) for ChatGPT 3.5 and 0.85 (0.07) for ChatGPT 4.0 compared with the official WHO answers. ChatGPT 3.5 and 4.0 can generate accurate and relevant COVID-19 information to a certain extent. However, compared with official WHO responses, gaps and deficiencies exist. Thus, users of ChatGPT 3.5 and 4.0 should also reference other reliable information sources to mitigate potential misinformation risks. Notably, ChatGPT 4.0 outperformed ChatGPT 3.5 across all evaluated dimensions, a finding corroborated by BERT model validation.

The Proof is in the Pudding

This paper presents FASTFOOD, a rule-based natural language generation (NLG) program for cooking recipes. We consider the representation of cooking recipes as discourse representation, because the meaning of each sentence needs to consider the context of the others. Our discourse representation system is based on states of affairs and transtions between states of affairs, and does not use discourse referents. Recipes are generated by using an automated theorem-proving procedure to select the ingredients and instructions, with ingredients corresponding to axioms and instructions to implications. FASTFOOD also contains a temporal optimization module which can rearrange the recipe to make it more time efficient for the user, e.g. the recipe specifies to chop the vegetables while the rice is boiling. The system is described in detail, including the decision to forgo discourse referents and how plausible representations of nouns and verbs emerge purely as a by-product of the practical requirements of efficiently representing recipe content. A comparison is then made with existing recipe generation systems, NLG systems more generally, and automated theorem provers.

Defending against Backdoor Attacks in Natural Language Generation

The frustratingly fragile nature of neural network models make current natural language generation (NLG) systems prone to backdoor attacks and generate malicious sequences that could be sexist or offensive. Unfortunately, little effort has been invested to how backdoor attacks can affect current NLG models and how to defend against these attacks. In this work, by giving a formal definition of backdoor attack and defense, we investigate this problem on two important NLG tasks, machine translation and dialog generation. Tailored to the inherent nature of NLG models (e.g., producing a sequence of coherent words given contexts), we design defending strategies against attacks. We find that testing the backward probability of generating sources given targets yields effective defense performance against all different types of attacks, and is able to handle the one-to-many issue in many NLG tasks such as dialog generation. We hope that this work can raise the awareness of backdoor risks concealed in deep NLG systems and inspire more future work (both attack and defense) towards this direction.

Evaluating factual accuracy in complex data-to-text

It is essential that data-to-text Natural Language Generation (NLG) systems produce texts which are factually accurate. We examine accuracy issues in the task of generating summaries of basketball games, including what accuracy means in this context, how accuracy errors can be detected by human annotators, as well as the types of accuracy mistakes made by both neural NLG systems and human authors. We also look at the effectiveness of automatic metrics in measuring factual accuracy.

InfoLM: A New Metric to Evaluate Summarization &amp; Data2Text Generation

Assessing the quality of natural language generation (NLG) systems through human annotation is very expensive. Additionally, human annotation campaigns are time-consuming and include non-reusable human labour. In practice, researchers rely on automatic metrics as a proxy of quality. In the last decade, many string-based metrics (e.g., BLEU or ROUGE) have been introduced. However, such metrics usually rely on exact matches and thus, do not robustly handle synonyms. In this paper, we introduce InfoLM a family of untrained metrics that can be viewed as a string-based metric that addresses the aforementioned flaws thanks to a pre-trained masked language model. This family of metrics also makes use of information measures allowing the possibility to adapt InfoLM to different evaluation criteria. Using direct assessment, we demonstrate that InfoLM achieves statistically significant improvement and two figure correlation gains in many configurations compared to existing metrics on both summarization and data2text generation tasks.

A Survey of Evaluation Metrics Used for NLG Systems

In the last few years, a large number of automatic evaluation metrics have been proposed for evaluating Natural Language Generation (NLG) systems. The rapid development and adoption of such automatic evaluation metrics in a relatively short time has created the need for a survey of these metrics. In this survey, we (i) highlight the challenges in automatically evaluating NLG systems, (ii) propose a coherent taxonomy for organising existing evaluation metrics, (iii) briefly describe different existing metrics, and finally (iv) discuss studies criticising the use of automatic evaluation metrics. We then conclude the article highlighting promising future directions of research.

WakaVT: A Sequential Variational Transformer for Waka Generation

Poetry generation has long been a challenge for artificial intelligence. In the scope of Japanese poetry generation, many researchers have paid attention to Haiku generation, but few have focused on Waka generation. To further explore the creative potential of natural language generation systems in Japanese poetry creation, we propose a novel Waka generation model, WakaVT, which automatically produces Waka poems given user-specified keywords. Firstly, an additive mask-based approach is presented to satisfy the form constraint. Secondly, the structures of Transformer and variational autoencoder are integrated to enhance the quality of generated content. Specifically, to obtain novelty and diversity, WakaVT employs a sequence of latent variables, which effectively captures word-level variability in Waka data. To improve linguistic quality in terms of fluency, coherence, and meaningfulness, we further propose the fused multilevel self-attention mechanism, which properly models the hierarchical linguistic structure of Waka. To the best of our knowledge, we are the first to investigate Waka generation with models based on Transformer and/or variational autoencoder. Both objective and subjective evaluation results demonstrate that our model outperforms baselines significantly.

Automatic linguistic reporting of customer activity patterns in open malls

In this work, we present a complete system to produce an automatic linguistic reporting about the customer activity patterns inside open malls, a mixed distribution of classical malls joined with the shops on the street. These reports can assist to design marketing campaigns by means of identifying the best places to catch the attention of customers. Activity patterns are estimated with process mining techniques and the key information of localization. Localization is obtained with a parallelized solution based on WiFi fingerprint system to speed up the solution. In agreement with the best practices for human evaluation of natural language generation systems, the linguistic quality of the generated report was evaluated by 41 experts who filled in an online questionnaire. Results are encouraging, since the average global score of the linguistic quality dimension is 6.17 (0.76 of standard deviation) in a 7-point Likert scale. This expresses a high degree of satisfaction of the generated reports and validates the adequacy of automatic natural language textual reports as a complementary tool to process model visualization.

Human evaluation of automatically generated text: Current trends and best practice guidelines

Currently, there is little agreement as to how Natural Language Generation (NLG) systems should be evaluated, with a particularly high degree of variation in the way that human evaluation is carried out. This paper provides an overview of how (mostly intrinsic) human evaluation is currently conducted and presents a set of best practices, grounded in the literature. These best practices are also linked to the stages that researchers go through when conducting an evaluation research (planning stage; execution and release stage), and the specific steps in these stages. With this paper, we hope to contribute to the quality and consistency of human evaluations in NLG.

Building a Persuasive Virtual Dietitian

This paper describes the Multimedia Application for Diet Management (MADiMan), a system that supports users in managing their diets while admitting diet transgressions. MADiMan consists of a numerical reasoner that takes into account users’ dietary constraints and automatically adapts the users’ diet, and of a natural language generation (NLG) system that automatically creates textual messages for explaining the results provided by the reasoner with the aim of persuading users to stick to a healthy diet. In the first part of the paper, we introduce the MADiMan system and, in particular, the basic mechanisms related to reasoning, data interpretation and content selection for a numeric data-to-text NLG system. We also discuss a number of factors influencing the design of the textual messages produced. In particular, we describe in detail the design of the sentence-aggregation procedure, which determines the compactness of the final message by applying two aggregation strategies. In the second part of the paper, we present the app that we developed, CheckYourMeal!, and the results of two human-based quantitative evaluations of the NLG module conducted using CheckYourMeal! in a simulation. The first evaluation, conducted with twenty users, ascertained both the perceived usefulness of graphics/text and the appeal, easiness and persuasiveness of the textual messages. The second evaluation, conducted with thirty-nine users, ascertained their persuasive power. The evaluations were based on the analysis of questionnaires and of logged data of users’ behaviour. Both evaluations showed significant results.

Personality-dependent content selection in natural language generation systems

This paper focuses on the computer side of human-computer interaction through natural language, which is the domain of natural language generation (NLG) studies. From a given (usually non-linguistic) input, NLG systems will in principle generate the same fixed text as an output and in order to attain more natural or human-like interaction will often resort to a wide range of strategies for stylistic variation. Among these, the use of computational models of human personality has emerged as a popular alternative in the field and will be the focus of the present work as well. More specifically, the present study describes two machine learning experiments to establish possible relations between personality and content selection (as opposed to the more well-documented relation between personality and surface realisation), and it is, to the best of our knowledge, the first of its kind to address this issue at both macro and micro planning levels, which may arguably pave the way for the future development of more robust personality-dependent systems of this kind.

Scalable Micro-planned Generation of Discourse from Structured Data

We present a framework for generating natural language description from structured data such as tables; the problem comes under the category of data-to-text natural language generation (NLG). Modern data-to-text NLG systems typically use end-to-end statistical and neural architectures that learn from a limited amount of task-specific labeled data, and therefore exhibit limited scalability, domain-adaptability, and interpretability. Unlike these systems, ours is a modular, pipeline-based approach, and does not require task-specific parallel data. Rather, it relies on monolingual corpora and basic off-the-shelf NLP tools. This makes our system more scalable and easily adaptable to newer domains. Our system utilizes a three-staged pipeline that: (i) converts entries in the structured data to canonical form, (ii) generates simple sentences for each atomic entry in the canonicalized representation, and (iii) combines the sentences to produce a coherent, fluent, and adequate paragraph description through sentence compounding and co-reference replacement modules. Experiments on a benchmark mixed-domain data set curated for paragraph description from tables reveals the superiority of our system over existing data-to-text approaches. We also demonstrate the robustness of our system in accepting other popular data sets covering diverse data types such as knowledge graphs and key-value maps.

Evaluating the state-of-the-art of End-to-End Natural Language Generation: The E2E NLG challenge

This paper provides a comprehensive analysis of the first shared task on End-to-End Natural Language Generation (NLG) and identifies avenues for future research based on the results. This shared task aimed to assess whether recent end-to-end NLG systems can generate more complex output by learning from datasets containing higher lexical richness, syntactic complexity and diverse discourse phenomena. Introducing novel automatic and human metrics, 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. Seq2seq-based systems have demonstrated a great potential for NLG in the challenge. We find that seq2seq systems generally score high in terms of word-overlap metrics and human evaluations of naturalness – with the winning Slug system (Juraska et al., 2018) being seq2seq-based. However, vanilla seq2seq models often fail to correctly express a given meaning representation if they lack a strong semantic control mechanism applied during decoding. Moreover, seq2seq models can be outperformed by hand-engineered systems in terms of overall quality, as well as complexity, length and diversity of outputs. This research has influenced, inspired and motivated a number of recent studies outwith the original competition, which we also summarise as part of this paper.

Towards more effective online environmental information provision through tailored Natural Language Generation: Profiles of Scottish river user groups and an evaluative online experiment

As a result of societal transformations, political governance shifts, and advances in ICT, online information has become crucial in efforts by public authorities to make citizens better stewards of the environment. Yet, their environmental information provision may not always be attuned to end users' rationales, behaviours and appreciations. This study revolves around dynamic river level information provided by an environmental regulator – updated once a day or more, and collected by a sensor network of 333 gauging stations along 232 Scottish rivers. Employing an elaborate mixed methods approach with qualitative and quantitative elements, we examined if profiling of web page user groups and the subsequent employment of a specially designed Natural Language Generation (NLG) system could foster more effective online information provision. We identified profiles for the three main user groups: fishing, flood risk related, and paddling. The existence of well-distinguishable rationales and characteristics was in itself an argument for profiling; the same river level information was used in entirely different ways by the three groups. We subsequently constructed an advanced online experiment that implemented NLG based on live river level data. We found that textual information can be of much value in translating dynamic technical information into straightforward messages for the specific purposes of the user groups. We conclude that tailored NLG could be widely used in more effective online environmental information provision, and we provide five practical recommendations for public authorities and other information providers.

Probabilistic Verb Selection for Data-to-Text Generation

In data-to-text Natural Language Generation (NLG) systems, computers need to find the right words to describe phenomena seen in the data. This paper focuses on the problem of choosing appropriate verbs to express the direction and magnitude of a percentage change (e.g., in stock prices). Rather than simply using the same verbs again and again, we present a principled data-driven approach to this problem based on Shannon’s noisy-channel model so as to bring variation and naturalness into the generated text. Our experiments on three large-scale real-world news corpora demonstrate that the proposed probabilistic model can be learned to accurately imitate human authors’ pattern of usage around verbs, outperforming the state-of-the-art method significantly.

A library for automatic natural language generation of spanish texts

In this article we present a novel system for natural language generation (nlg) of Spanish sentences from a minimum set of meaningful words (such as nouns, verbs and adjectives) which, unlike other state-of-the-art solutions, performs the nlg task in a fully automatic way, exploiting both knowledge-based and statistical approaches. Relying on its linguistic knowledge of vocabulary and grammar, the system is able to generate complete, coherent and correctly spelled sentences from the main word sets presented by the user. The system, which was designed to be integrable, portable and efficient, can be easily adapted to other languages by design and can feasibly be integrated in a wide range of digital devices. During its development we also created a supplementary lexicon for Spanish, aLexiS, with wide coverage and high precision, as well as syntactic trees from a freely available definite-clause grammar. The resulting nlg library has been evaluated both automatically and manually (annotation). The system can potentially be used in different application domains such as augmentative communication and automatic generation of administrative reports or news.

Tool vs. agent: attributing agency to natural language generation systems

ABSTRACTThis paper argues that we should shift our consideration of natural language generation systems as tools for manifesting human intent to natural language generation systems as agents in themselves. Such a semantic shift would permit a more holistic conversation about the transformative social power of these systems’ output.

No Landslide for the Human Journalist - An Empirical Study of Computer-Generated Election News in Finland

In an age of struggling news media, automated generation of news via natural language generation (NLG) methods could be of great help, especially in areas where the amount of raw input data is big, and the structure of the data is known in advance. One such news automation system is the Valtteri NLG system, which generates news articles about the Finnish municipal elections of 2017. To evaluate the quality of Valtteri-produced articles and to identify aspects to improve, $n=152$ users were asked to evaluate the output of Valtteri. Each evaluator rated six preselected computer-generated articles, four control articles written by journalists, and four computer-generated articles of their own choice. All the articles were evaluated along four dimensions: credibility, liking, quality, and representativeness. As expected, the texts written by Valtteri received lower ratings than those written by journalists, but overall the ratings were satisfactory (average 2.9 versus 4.0 for journalists on a five-point scale). Valtteri’s best rating (3.6) was for credibility. The computer-written articles that the evaluators could freely select got slightly better ratings than the preselected computer-written articles. When looking at the results by demographic groups, males aged 55 or more liked the automatic articles best and females aged 34 or less liked them the least. Evaluators mistook 21% of the computer-written articles as written by humans and 10% of the human-written articles as computer-written. The share of users making these mistakes grew with the age. Overall, the male evaluators made less writer-identification mistakes than female evaluators did.