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.

Machine learning techniques have shown their competence for representing and reasoning in symbolic systems such as language and phonology. In Sinitic Historical Phonology, notable tasks that could benefit from machine learning include the comparison of dialects and reconstruction of proto-languages systems. Motivated by this, this paper provides an approach for obtaining multi-dialectal representations of Sinitic syllables, by constructing a knowledge graph from structured phonological data ,then applying the BoxE technique from knowledge base learning. We applied unsupervised clustering techniques to the obtained representations to observe that the representations capture phonemic contrast from the input dialects. Furthermore, we trained classifiers to perform inference of unobserved Middle Chinese labels, showing the representations’ potential for indicating archaic, proto-language features. The representations can be used for performing completion of fragmented Sinitic phonological knowledge bases, estimating divergences between different characters, or aiding the exploration and reconstruction of archaic features.

In recent years, pre-trained language models have garnered significant attention due to their effectiveness, which stems from the rich knowledge acquired during pre-training. To mitigate the inconsistency issues between pre-training tasks and downstream tasks and to facilitate the resolution of language-related issues, prompt-based approaches have been introduced, which are particularly useful in low-resource scenarios. However, existing approaches mostly rely on verbalizers to translate the predicted vocabulary to task-specific labels. The major limitations of this approach are the ignorance of potentially relevant domain-specific words and being biased by the pre-training data. To address these limitations, we propose a framework that incorporates conceptual knowledge for text classification in the extreme zero-shot setting. The framework includes prompt-based keyword extraction, weight assignment to each prompt keyword, and final representation estimation in the knowledge graph embedding space. We evaluated the method on four widely-used datasets for sentiment analysis and topic detection, demonstrating that it consistently outperforms recently-developed prompt-based approaches in the same experimental settings.

This paper investigates the effect of tokenizers on the downstream performance of pretrained language models (PLMs) in scriptio continua languages where no explicit spaces exist between words, using Japanese as a case study. The tokenizer for such languages often consists of a morphological analyzer and a subword tokenizer, requiring us to conduct a comprehensive study of all possible pairs. However, previous studies lack this comprehensiveness. We therefore train extensive sets of tokenizers, build a PLM using each, and measure the downstream performance on a wide range of tasks. Our results demonstrate that each downstream task has a different optimal morphological analyzer, and that it is better to use Byte-Pair-Encoding or Unigram rather than WordPiece as a subword tokenizer, regardless of the type of task.

Event-Level Video Question Answering (EVQA) requires complex reasoning across video events to obtain the visual information needed to provide optimal answers. However, despite significant progress in model performance, few studies have focused on using the explicit semantic connections between the question and visual information especially at the event level. There is need for using such semantic connections to facilitate complex reasoning across video frames. Therefore, we propose a semantic-aware dynamic retrospective-prospective reasoning approach for video-based question answering. Specifically, we explicitly use the Semantic Role Labeling (SRL) structure of the question in the dynamic reasoning process where we decide to move to the next frame based on which part of the SRL structure (agent, verb, patient, etc.) of the question is being focused on. We conduct experiments on a benchmark EVQA dataset - TrafficQA. Results show that our proposed approach achieves superior performance compared to previous state-of-the-art models. Our code is publicly available at https://github.com/lyuchenyang/Semantic-aware-VideoQA.

Natural Language Inference (NLI) tasks involving temporal inference remain challenging for pre-trained language models (LMs). Although various datasets have been created for this task, they primarily focus on English and do not address the need for resources in other languages. It is unclear whether current LMs realize the generalization capacity for temporal inference across languages. In this paper, we present Jamp, a Japanese NLI benchmark focused on temporal inference. Our dataset includes a range of temporal inference patterns, which enables us to conduct fine-grained analysis. To begin the data annotation process, we create diverse inference templates based on the formal semantics test suites. We then automatically generate diverse NLI examples by using the Japanese case frame dictionary and well-designed templates while controlling the distribution of inference patterns and gold labels. We evaluate the generalization capacities of monolingual/multilingual LMs by splitting our dataset based on tense fragments (i.e., temporal inference patterns). Our findings demonstrate that LMs struggle with specific linguistic phenomena, such as habituality, indicating that there is potential for the development of more effective NLI models across languages.

Code search is a task to find programming codes that semantically match the given natural language queries. Even though some of the existing datasets for this task are multilingual on the programming language side, their query data are only in English. In this research, we create a multilingual code search dataset in four natural and four programming languages using a neural machine translation model. Using our dataset, we pre-train and fine-tune the Transformer-based models and then evaluate them on multiple code search test sets. Our results show that the model pre-trained with all natural and programming language data has performed best in most cases. By applying back-translation data filtering to our dataset, we demonstrate that the translation quality affects the model’s performance to a certain extent, but the data size matters more.

This study proposes a new multimodal neural machine translation (MNMT) model using synthetic images transformed by a latent diffusion model. MNMT translates a source language sentence based on its related image, but the image usually contains noisy information that are not relevant to the source language sentence. Our proposed method first generates a synthetic image corresponding to the content of the source language sentence by using a latent diffusion model and then performs translation based on the synthetic image. The experiments on the English-German translation tasks using the Multi30k dataset demonstrate the effectiveness of the proposed method.

Despite the rapid development of neural-based models, syntax still plays a crucial role in modern natural language processing. However, few studies have incorporated syntactic information into ancient Chinese understanding tasks due to the lack of syntactic annotation. This paper explores the role of syntax in ancient Chinese understanding based on the noisy syntax trees from unsupervised derivation and modern Chinese syntax parsers. On top of that, we propose a novel syntax encoding component – confidence-based syntax encoding network (cSEN) to alleviate the side effects from the existing noise caused by unsupervised syntax derivation and the incompatibility between ancient and modern Chinese. Experiments on two typical ancient Chinese understanding tasks, ancient poetry theme classification and ancient-modern Chinese translation, demonstrate that syntactic information can effectively enhance the understanding of ancient Chinese over strong baselines, and that the proposed cSEN plays an important role in noisy scenarios.

In this paper, we study the viability of the deployment of language models towards non-playable character (NPC) scripts, by introducing a novel pipeline for the automatic construction of NPC scripts using Transformer-based believable scripts for a variety of game genres and specifications. In addition, we propose a self-diagnosis method inspired by previous work to develop language models, tailored specifically to desirable NPC qualities such as coherency, believability, and degree of repetition. Finally, we propose a new benchmark, called The Turing Quest, which we use to show that the pipeline, when applied to GPT-3, can generate for a variety of game genres and contexts, NPC scripts that can fool judges in thinking they have been written by humans. We believe that these findings can greatly benefit both the gaming industry and its global community of users, since many current games continue to base their NPCs on manually-curated scripts that are resource-demanding and may curb the immersiveness and enjoyment of the user.

Recent advances in large language models have led to renewed interest in natural language processing in healthcare using the free text of clinical notes. One distinguishing characteristic of clinical notes is their long time span over multiple long documents. The unique structure of clinical notes creates a new design choice: when the context length for a language model predictor is limited, which part of clinical notes should we choose as the input? Existing studies either choose the inputs with domain knowledge or simply truncate them. We propose a framework to analyze the sections with high predictive power. Using MIMIC-III, we show that: 1) predictive power distribution is different between nursing notes and discharge notes and 2) combining different types of notes could improve performance when the context length is large. Our findings suggest that a carefully selected sampling function could enable more efficient information extraction from clinical notes.

The Ninth Revision of the International Classification of Diseases (ICD-9) is a standardized coding system used to classify health conditions. It is used for billing, tracking individual patient conditions, and for epidemiology. The highly detailed and technical nature of the codes and their associated medical conditions make it difficult for humans to accurately record them. Researchers have explored the use of neural networks, particularly language models, for automated ICD-9 code assignment. However, the imbalanced distribution of ICD-9 codes leads to poor performance. One solution is to use domain knowledge to incorporate a useful prior. This paper evaluates the usefulness of the correlation bias: we hypothesize that correlations between ICD-9 codes and other medical codes could help improve language models’ performance. We showed that while the correlation bias worsens the overall performance, the effect on individual class can be negative or positive. Performance on classes that are more imbalanced and less correlated with other codes is more sensitive to incorporating the correlation bias. This suggests that while the correlation bias has potential to improve ICD-9 code assignment in certain cases, the applicability criteria need to be more carefully studied.

State-of-the-art models can perform well in controlled environments, but they often struggle when presented with out-of-distribution (OOD) examples, making OOD detection a critical component of NLP systems. In this paper, we focus on highlighting the limitations of existing approaches to OOD detection in NLP. Specifically, we evaluated eight OOD detection methods that are easily integrable into existing NLP systems and require no additional OOD data or model modifications. One of our contributions is providing a well-structured research environment that allows for full reproducibility of the results. Additionally, our analysis shows that existing OOD detection methods for NLP tasks are not yet sufficiently sensitive to capture all samples characterized by various types of distributional shifts. Particularly challenging testing scenarios arise in cases of background shift and randomly shuffled word order within in domain texts. This highlights the need for future work to develop more effective OOD detection approaches for the NLP problems, and our work provides a well-defined foundation for further research in this area.

It has been suggested that pretrained language models can be viewed as knowledge bases. One of the prerequisites for using language models as knowledge bases is how accurately they can store and retrieve world knowledge. It is already revealed that language models can store much 1-to-1 relational knowledge, such as ”country and its capital,” with high memorization accuracy. On the other hand, world knowledge includes not only 1-to-1 but also 1-to-N relational knowledge, such as ”parent and children.”However, it is not clear how accurately language models can handle 1-to-N relational knowledge. To investigate language models’ abilities toward 1-to-N relational knowledge, we start by designing the problem settings. Specifically, we organize the character of 1-to-N relational knowledge and define two essential skills: (i) memorizing multiple objects individually and (ii) retrieving multiple stored objects without excesses or deficiencies at once. We inspect LMs’ ability to handle 1-to-N relational knowledge on the controlled synthesized data. As a result, we report that it is possible to memorize multiple objects with high accuracy, but generalizing the retrieval ability (expressly, enumeration) is challenging.

While embedding-based methods have been dominant in language clustering for multilingual tasks, clustering based on linguistic features has not yet been explored much, as it remains baselines (Tan et al., 2019; Shaffer, 2021). This study investigates whether and how theoretical linguistics improves language clustering for multilingual named entity recognition (NER). We propose two types of language groupings: one based on morpho-syntactic features in a nominal domain and one based on a head parameter. Our NER experiments show that the proposed methods largely outperform a state-of-the-art embedding-based model, suggesting that theoretical linguistics plays a significant role in multilingual learning tasks.

Numerous studies found that the linguistic properties of a person’s native language affect the cognitive processing of other languages. However, only one study has shown that it was possible to identify the native language based on eye-tracking records of natural L2 reading using machine learning. A new corpus allows us to replicate these results on a more interrelated and larger set of native languages. Our results show that comparable classification performance is maintained despite using less data. However, analysis shows that the correlation between L2 eye movements and native language similarity may be more complex than the original study found.

Discharge summaries are comprehensive medical records that encompass vital information about a patient’s hospital stay. A crucial aspect of discharge summaries is the temporal information of treatments administered throughout the patient’s illness. With an extensive volume of clinical documents, manually extracting and compiling a patient’s medication list can be laborious, time-consuming, and susceptible to errors. The objective of this paper is to build upon the recent development on clinical NLP by temporally classifying treatments in clinical texts, specifically determining whether a treatment was administered between the time of admission and discharge from the hospital. State-of-the-art NLP methods including prompt-based learning on Generative Pre-trained Transformers (GPTs) models and fine-tuning on pre-trained language models (PLMs) such as BERT were employed to classify temporal relations between treatments and hospitalisation periods in discharge summaries. Fine-tuning with the BERT model achieved an F1 score of 92.45% and a balanced accuracy of 77.56%, while prompt learning using the T5 model and mixed templates resulted in an F1 score of 90.89% and a balanced accuracy of 72.07%.Our codes and data are available at https://github.com/HECTA-UoM/MedTem.

In this paper, we investigate a type of semantic shift that occurs when a sudden event radically changes public opinion on a topic. Looking at Sweden’s decision to apply for NATO membership in 2022, we use word embeddings to study how the associations users on Twitter have regarding NATO evolve. We identify several changes that we successfully validate against real-world events. However, the low engagement of the public with the issue often made it challenging to distinguish true signals from noise. We thus find that domain knowledge and data selection are of prime importance when using word embeddings to study semantic shifts.

A buzzer quiz is a genre of quiz in which multiple players simultaneously listen to a quiz being read aloud and respond it by buzzing in as soon as they can predict the answer. Because incorrect answers often result in penalties, a buzzer-quiz answering system must not only predict the answer from only part of a question but also estimate the predicted answer’s accuracy. In this paper, we introduce two types of buzzer-quiz answering systems: (1) a system that directly generates an answer from part of a question by using an autoregressive language model; and (2) a system that first reconstructs the entire question by using an autoregressive language model and then determines the answer according to the reconstructed question. We then propose a method to estimate the accuracy of the answers for each system by using the internal scores of each model.

Hyperbole is a common figure of speech, which is under-explored in NLP research. In this study, we conduct edge and minimal description length (MDL) probing experiments on three pre-trained language models (PLMs) in an attempt to explore the extent to which hyperbolic information is encoded in these models. We use both word-in-context and sentence-level representations as model inputs as a basis for comparison. We also annotate 63 hyperbole sentences from the HYPO dataset according to an operational taxonomy to conduct an error analysis to explore the encoding of different hyperbole categories. Our results show that hyperbole is to a limited extent encoded in PLMs, and mostly in the final layers. They also indicate that hyperbolic information may be better encoded by the sentence-level representations, which, due to the pragmatic nature of hyperbole, may therefore provide a more accurate and informative representation in PLMs. Finally, the inter-annotator agreement for our annotations, a Cohen’s Kappa of 0.339, suggest that the taxonomy categories may not be intuitive and need revision or simplification.

In this paper we describe the task of adapting NLP models to dialogue processing in the emergency response domain. Our goal is to provide a recipe for building a system that performs dialogue act classification and domain-specific slot tagging while being efficient, flexible and robust. We show that adapter models Pfeiffer et al. (2020) perform well in the emergency response domain and benefit from additional dialogue context and speaker information. Comparing adapters to standard fine-tuned Transformer models we show that they achieve competitive results and can easily accommodate new tasks without significant memory increase since the base model can be shared between the adapters specializing on different tasks. We also address the problem of scarce annotations in the emergency response domain and evaluate different data augmentation techniques in a low-resource setting.

Neural text generation models have achieved remarkable success in carrying on short open-domain conversations. However, their performance degrades significantly in the long term, especially in their ability to ask coherent questions. A significant issue is the generation of redundant questions where the answer has already been provided by the user. We adapt and evaluate different methods, including negative training, decoding, and classification, to mitigate the redundancy problem. We also propose a simple yet effective method for generating training data without the need for crowdsourcing human-human or human-bot conversations. Experiments with the BlenderBot model show that our combined method significantly reduces the rate of redundant questions from 27.2% to 8.7%, while improving the quality of the original model. The code, dataset, and trained models can be found at our repository.

Currently, most knowledge-grounded dialogue response generation models focus on reflecting given external knowledge. However, even when conveying external knowledge, humans integrate their own knowledge, experiences, and opinions with external knowledge to make their utterances engaging. In this study, we analyze such human behavior by annotating the utterances in an existing knowledge-grounded dialogue corpus. Each entity in the corpus is annotated with its information source, either derived from external knowledge (database-derived) or the speaker’s own knowledge, experiences, and opinions (speaker-derived). Our analysis shows that the presence of speaker-derived information in the utterance improves dialogue engagingness. We also confirm that responses generated by an existing model, which is trained to reflect the given knowledge, cannot include speaker-derived information in responses as often as humans do.

Pre-trained language models have achieved remarkable results on several NLP tasks. Most of them adopt masked language modeling to learn representations by randomly masking tokens and predicting them based on their context. However, this random selection of tokens to be masked is inefficient to learn some language patterns as it may not consider linguistic information that can be helpful for many NLP tasks, such as multimodal machine translation (MMT). Hence, we propose three novel masking strategies for cross-lingual visual pre-training - more informed visual masking, more informed textual masking, and more informed visual and textual masking - each one focusing on learning different linguistic patterns. We apply them to Vision Translation Language Modelling for video subtitles (Sato et al., 2022) and conduct extensive experiments on the Portuguese-English MMT task. The results show that our masking approaches yield significant improvements over the original random masking strategy for downstream MMT performance. Our models outperform the MMT baseline and we achieve state-of-the-art accuracy (52.70 in terms of BLEU score) on the How2 dataset, indicating that more informed masking helps in acquiring an understanding of specific language structures and has great potential for language understanding.

Procedural knowledge understanding (PKU) underlies the ability to infer goal-step relations. The task of Visual Goal–Step Inference addresses this ability in the multimodal domain. It requires to identify images that represent the steps towards achieving a textually expressed goal. The best existing methods encode texts and images either with independent encoders, or with object-level multimodal encoders using blackbox transformers. This stands in contrast to early, linguistically inspired methods for event representations, which focus on capturing the most crucial information, namely actions and the participants, to learn stereotypical event sequences and hence procedural knowledge. In this work, we study various methods and their effects on PKU of injecting the early shallow event representations to nowadays multimodal deep learning-based models. We find that the early, linguistically inspired methods for representing event knowledge does contribute to understand procedures in combination with modern vision-and-language models. In the future, we are going to explore more complex structure of events and study how to exploit it on top of large language models.

Although Shapley values have been shown to be highly effective for identifying harmful training instances, dataset size and model complexity constraints limit the ability to apply Shapley-based data valuation to fine-tuning large pre-trained language models. To address this, we propose TS-DShapley, an algorithm that reduces computational cost of Shapley-based data valuation through: 1) an efficient sampling-based method that aggregates Shapley values computed from subsets for valuation of the entire training set, and 2) a value transfer method that leverages value information extracted from a simple classifier trained using representations from the target language model. Our experiments applying TS-DShapley to select data for fine-tuning BERT-based language models on benchmark natural language understanding (NLU) datasets show that TS-DShapley outperforms existing data selection methods. Further, TS-DShapley can filter fine-tuning data to increase language model performance compared to training with the full fine-tuning dataset.

This study addresses the automatic generation of distractors for English fill-in-the-blank exercises in the entrance examinations for Japanese universities. While previous studies applied the same method to all questions, actual entrance examinations have multiple question types that reflect the purpose of the questions. Therefore, we define three types of questions (grammar, function word, and context) and propose a method to generate distractors according to the characteristics of each question type. Experimental results on 500 actual questions show the effectiveness of the proposed method for both automatic and manual evaluation.

Large Language Models (LLMs) have demonstrated impressive capabilities in generating fluent text, as well as tendencies to reproduce undesirable social biases. This work investigates whether LLMs reproduce the moral biases associated with political groups in the United States, an instance of a broader capability herein termed moral mimicry. This work explores this hypothesis in the GPT-3/3.5 and OPT families of Transformer-based LLMs. Using tools from Moral Foundations Theory, this work shows that these LLMs are indeed moral mimics. When prompted with a liberal or conservative political identity, the models generate text reflecting corresponding moral biases. This study also explores the relationship between moral mimicry and model size, and similarity between human and LLM moral word use.

Recently, dynamic early exiting has attracted much attention since it can accelerate the inference speed of pre-trained models (PTMs). However, previous work on early exiting has neglected the intermediate exits’ architectural designs. In this work, we propose a novel framework, Learned Exits and COmparison-based early exiting (LECO) to improve PTMs’ early exiting performances. First, to fully uncover the potentials of multi-exit BERT, we design a novel search space for intermediate exits and employ the idea of differentiable neural architecture search (DNAS) to design proper exit architectures for different intermediate layers automatically. Second, we propose a simple-yet-effective comparison-based early exiting mechanism (COBEE), which can help PTMs achieve better performance and speedup tradeoffs. Extensive experiments show that our LECO achieves the SOTA performances for multi-exit BERT training and dynamic early exiting.

Authorship attribution aims to identify the author of an anonymous text. The task becomes even more worthwhile when it comes to literary works. For example, pen names were commonly used by female authors in the 19th century resulting in some literary works being incorrectly attributed or claimed. With this motivation, we collated a dataset of late 19th century novels in English. Due to the imbalance in the dataset and the unavailability of enough data per author, we employed the GANBERT model along with data sampling strategies to fine-tune a transformer-based model for authorship attribution. Differently from the earlier studies on the GAN-BERT model, we conducted transfer learning on comparatively smaller author subsets to train more focused author-specific models yielding performance over 0.88 accuracy and F1 scores. Furthermore, we observed that increasing the sample size has a negative impact on the model’s performance. Our research mainly contributes to the ongoing authorship attribution research using GAN-BERT architecture, especially in attributing disputed novelists in the late 19th century.

Instructional texts for specific target groups should ideally take into account the prior knowledge and needs of the readers in order to guide them efficiently to their desired goals. However, targeting specific groups also carries the risk of reflecting disparate social norms and subtle stereotypes. In this paper, we investigate the extent to which how-to guides from one particular platform, wikiHow, differ in practice depending on the intended audience. We conduct two case studies in which we examine qualitative features of texts written for specific audiences. In a generalization study, we investigate which differences can also be systematically demonstrated using computational methods. The results of our studies show that guides from wikiHow, like other text genres, are subject to subtle biases. We aim to raise awareness of these inequalities as a first step to addressing them in future work.

Sarcasm is a form of figurative language that serves as a humorous tool for mockery and ridicule. We present a novel architecture for sarcasm generation with emoji from a non-sarcastic input sentence in English. We divide the generation task into two sub tasks: one for generating textual sarcasm and another for collecting emojis associated with those sarcastic sentences. Two key elements of sarcasm are incorporated into the textual sarcasm generation task: valence reversal and semantic incongruity with context, where the context may involve shared commonsense or general knowledge between the speaker and their audience. The majority of existing sarcasm generation works have focused on this textual form. However, in the real world, when written texts fall short of effectively capturing the emotional cues of spoken and face-to-face communication, people often opt for emojis to accurately express their emotions. Due to the wide range of applications of emojis, incorporating appropriate emojis to generate textual sarcastic sentences helps advance sarcasm generation. We conclude our study by evaluating the generated sarcastic sentences using human judgement. All the codes and data used in this study has been made publicly available.

With the fast-growing popularity of current large pre-trained language models (LLMs), it is necessary to dedicate efforts to making them more reliable. In this thesis proposal, we aim to improve the reliability of natural language generation systems (NLG) by researching the semantic accuracy of their outputs. We look at this problem from the outside (evaluation) and from the inside (interpretability). We propose a novel method for evaluating semantic accuracy and discuss the importance of working towards a unified and objective benchmark for NLG metrics. We also review interpretability approaches which could help us pinpoint the sources of inaccuracies within the models and explore potential mitigation strategies.

The art of mathematical reasoning stands as a fundamental pillar of intellectual progress and is a central catalyst in cultivating human ingenuity. Researchers have recently published a plethora of works centered around the task of solving Math Word Problems (MWP) — a crucial stride towards general AI. These existing models are susceptible to dependency on shallow heuristics and spurious correlations to derive the solution expressions. In order to ameliorate this issue, in this paper, we propose a framework for MWP solvers based on the generation of linguistic variants of the problem text. The approach involves solving each of the variant problems and electing the predicted expression with the majority of the votes. We use DeBERTa (Decoding-enhanced BERT with disentangled attention) as the encoder to leverage its rich textual representations and enhanced mask decoder to construct the solution expressions. Furthermore, we introduce a challenging dataset, ParaMAWPS, consisting of paraphrased, adversarial, and inverse variants of selectively sampled MWPs from the benchmark Mawps dataset. We extensively experiment on this dataset along with other benchmark datasets using some baseline MWP solver models. We show that training on linguistic variants of problem statements and voting on candidate predictions improve the mathematical reasoning and robustness of the model. We make our code and data publicly available.