LLM-as-a-judge approaches are a practical and effective way of assessing a range of text tasks. However, when using pairwise comparisons to rank a set of candidates, the computational cost scales quadratically with the number of candidates, which has practical limitations. This paper introduces a Product of Expert (PoE) framework for efficient LLM Comparative Assessment. Here individual comparisons are considered experts that provide information on a pair’s score difference. The PoE framework combines the information from these experts to yield an expression that can be maximized with respect to the underlying set of candidates, and is highly flexible where any form of expert can be assumed. When Gaussian experts are used one can derive simple closed-form solutions for the optimal candidate ranking, as well as expressions for selecting which comparisons should be made to maximize the probability of this ranking. Our approach enables efficient comparative assessment, where by using only a small subset of the possible comparisons, one can generate score predictions that correlate well with human judgements. We evaluate the approach on multiple NLG tasks and demonstrate that our framework can yield considerable computational savings when performing pairwise comparative assessment. With many candidate texts, using as few as 2% of comparisons the PoE solution can achieve similar performance to when all comparisons are used.

Large Language Models (LLMs) are powerful zero-shot assessors used in real-world situations such as assessing written exams and benchmarking systems. Despite these critical applications, no existing work has analyzed the vulnerability of judge-LLMs to adversarial manipulation. This work presents the first study on the adversarial robustness of assessment LLMs, where we demonstrate that short universal adversarial phrases can be concatenated to deceive judge LLMs to predict inflated scores. Since adversaries may not know or have access to the judge-LLMs, we propose a simple surrogate attack where a surrogate model is first attacked, and the learned attack phrase then transferred to unknown judge-LLMs. We propose a practical algorithm to determine the short universal attack phrases and demonstrate that when transferred to unseen models, scores can be drastically inflated such that irrespective of the assessed text, maximum scores are predicted. It is found that judge-LLMs are significantly more susceptible to these adversarial attacks when used for absolute scoring, as opposed to comparative assessment. Our findings raise concerns on the reliability of LLM-as-a-judge methods, and emphasize the importance of addressing vulnerabilities in LLM assessment methods before deployment in high-stakes real-world scenarios.

With the recent emergence of powerful instruction-tuned large language models (LLMs), various helpful conversational Artificial Intelligence (AI) systems have been deployed across many applications. When prompted by users, these AI systems successfully perform a wide range of tasks as part of a conversation. To provide some sort of memory and context, such approaches typically condition their output on the entire conversational history. Although this sensitivity to the conversational history can often lead to improved performance on subsequent tasks, we find that performance can in fact also be negatively impacted, if there is a _task-switch_. To the best of our knowledge, our work makes the first attempt to formalize the study of such vulnerabilities and interference of tasks in conversational LLMs caused by task-switches in the conversational history. Our experiments across 5 datasets with 15 task switches using popular LLMs reveal that many of the task-switches can lead to significant performance degradation.

Watermarking generative-AI systems, such as LLMs, has gained considerable interest, driven by their enhanced capabilities across a wide range of tasks. Although current approaches have demonstrated that small, context-dependent shifts in the word distributions can be used to apply and detect watermarks, there has been little work in analyzing the impact that these perturbations have on the quality of generated texts. Balancing high detectability with minimal performance degradation is crucial in terms of selecting the appropriate watermarking setting; therefore this paper proposes a simple analysis framework where comparative assessment, a flexible NLG evaluation framework, is used to assess the quality degradation caused by a particular watermark setting. We demonstrate that our framework provides easy visualization of the quality-detection trade-off of watermark settings, enabling a simple solution to find an LLM watermark operating point that provides a well-balanced performance. This approach is applied to two different summarization systems and a translation system, enabling cross-model analysis for a task, and cross-task analysis.

Large Language Models (LLMs) have demonstrated impressive zero-shot capabilities and versatility in NLP tasks, however they sometimes fail to maintain crucial invariances for specific tasks. One example is permutation sensitivity, where LLMs’ outputs may significantly vary depending on the order of the input options. While debiasing techniques can mitigate these issues, and yield better performance and reliability, they often come with a high computational cost at inference. This paper addresses this inefficiency at inference time. The aim is to distill the capabilities of a computationally intensive, debiased, teacher model into a more compact student model. We explore two variants of student models: one based on pure distillation, and the other on an error-correction approach for more complex tasks, where the student corrects a single biased decision from the teacher to achieve a debiased output. Our approach is general and can be applied to both black-box and white-box LLMs. Furthermore, we demonstrate that our compact, encoder-only student models can outperform their larger, biased teacher counterparts, achieving better results with significantly fewer parameters.

Text and vision foundation models can perform many tasks in a zero-shot setting, a desirable property that enables these systems to be applied in general and low-resource settings. There has been far less work, however, on the zero-shot abilities of ASR foundation models, with these systems typically fine-tuned to specific tasks or constrained to applications that match their training criterion and data annotation. In this work we investigate the ability of Whisper and MMS, ASR foundation models trained primarily for speech recognition, to perform zero-shot audio classification. We use simple template-based text prompts at the decoder and use the resulting decoding probabilities to generate zero-shot predictions. Without training the model on extra data or adding any new parameters, we demonstrate that Whisper shows promising zero-shot classification performance on a range of 8 audio-classification datasets, outperforming the accuracy of existing state-of-the-art zero-shot baselines by an average of 9%. One important step to unlock the emergent ability is debiasing, where a simple unsupervised reweighting method of the class probabilities yields consistent significant performance gains. We further show that performance increases with model size, implying that as ASR foundation models scale up, they may exhibit improved zero-shot performance.

Encoder-decoder foundation models have displayed state-of-the-art performance on a range of autoregressive sequence tasks. This paper proposes a simple and lightweight modification to such systems to control the behaviour according to a specific attribute of interest. This paper proposes a novel inference-efficient approach to modifying the behaviour of an encoder-decoder system according to a specific attribute of interest. Specifically, we show that a small proxy network can be used to find a sample-by-sample perturbation of the encoder output of a frozen foundation model to trigger the decoder to generate improved decodings. This work explores a specific realization of this framework focused on improving the COMET performance of Flan-T5 on Machine Translation and the WER of Whisper foundation models on Speech Recognition. Results display consistent improvements in performance evaluated through COMET and WER respectively. Furthermore, experiments also show that the proxies are robust to the exact nature of the data used to train them and can extend to other domains.

Enterprise retrieval augmented generation (RAG) offers a highly flexible framework for combining powerful large language models (LLMs) with internal, possibly temporally changing, documents. In RAG, documents are first chunked. Relevant chunks are then retrieved for a user query, which are passed as context to a synthesizer LLM to generate the query response. However, the retrieval step can limit performance, as incorrect chunks can lead the synthesizer LLM to generate a false response. This work applies a zero-shot adaptation of standard dense retrieval steps for more accurate chunk recall. Specifically, a chunk is first decomposed into atomic statements. A set of synthetic questions are then generated on these atoms (with the chunk as the context). Dense retrieval involves finding the closest set of synthetic questions, and associated chunks, to the user query. It is found that retrieval with the atoms leads to higher recall than retrieval with chunks. Further performance gain is observed with retrieval using the synthetic questions generated over the atoms. Higher recall at the retrieval step enables higher performance of the enterprise LLM using the RAG pipeline.

Current developments in large language models (LLMs) have enabled impressive zero-shot capabilities across various natural language tasks. An interesting application of these systems is in the automated assessment of natural language generation (NLG), a highly challenging area with great practical benefit. In this paper, we explore two options for exploiting the emergent abilities of LLMs for zero-shot NLG assessment: absolute score prediction, and comparative assessment which uses relative comparisons between pairs of candidates. Though comparative assessment has not been extensively studied in NLG assessment, we note that humans often find it more intuitive to compare two options rather than scoring each one independently. This work examines comparative assessment from multiple perspectives: performance compared to absolute grading; positional biases in the prompt; and efficient ranking in terms of the number of comparisons. We illustrate that LLM comparative assessment is a simple, general and effective approach for NLG assessment. For moderate-sized open-source LLMs, such as FlanT5 and Llama2-chat, comparative assessment is superior to prompt scoring, and in many cases can achieve performance competitive with state-of-the-art methods. Additionally, we demonstrate that LLMs often exhibit strong positional biases when making pairwise comparisons, and we propose debiasing methods that can further improve performance.

Sequence-to-sequence models often require an expensive autoregressive decoding process. However, for some downstream tasks such as out-of-distribution (OOD) detection and resource allocation, the actual decoding output is not needed, just a scalar attribute of this sequence. In such scenarios, where knowing the quality of a system’s output to predict poor performance prevails over knowing the output itself, is it possible to bypass the autoregressive decoding? We propose Non-Autoregressive Proxy (NAP) models that can efficiently predict scalar-valued sequence-level attributes. Importantly, NAPs predict these metrics directly from the encodings, avoiding the expensive decoding stage. We consider two sequence tasks: Machine Translation (MT) and Automatic Speech Recognition (ASR). In OOD for MT, NAPs outperform ensembles while being significantly faster. NAPs are also proven capable of predicting metrics such as BERTScore (MT) or word error rate (ASR). For downstream tasks, such as data filtering and resource optimization, NAPs generate performance predictions that outperform predictive uncertainty while being highly inference efficient.

Automated essay scoring (AES) to evaluate second language (L2) proficiency has been a firmly established technology used in educational contexts for decades. Although holistic scoring has seen advancements in AES that match or even exceed human performance, analytic scoring still encounters issues as it inherits flaws and shortcomings from the human scoring process. The recent introduction of large language models presents new opportunities for automating the evaluation of specific aspects of L2 writing proficiency. In this paper, we perform a series of experiments using GPT-4 in a zero-shot fashion on a publicly available dataset annotated with holistic scores based on the Common European Framework of Reference and aim to extract detailed information about their underlying analytic components. We observe significant correlations between the automatically predicted analytic scores and multiple features associated with the individual proficiency components.

Understanding the contribution of the inputs on the output is useful across many tasks. This work provides an information-theoretic framework to analyse the influence of inputs for text classification tasks. Natural language processing (NLP) tasks take either a single or multiple text elements to predict an output variable. Each text element has two components: the semantic meaning and a linguistic realization. Multiple-choice reading comprehension (MCRC) and sentiment classification (SC) are selected to showcase the framework. For MCRC, it is found that the relative context influence on the output reduces on more challenging datasets. In particular, more challenging contexts allows greater variation in the question complexity. Hence, test creators need to carefully consider the choice of the context when designing multiple-choice questions for assessment. For SC, it is found the semantic meaning of the input dominates compared to its linguistic realization when determining the sentiment. The framework is made available at: https://github.com/WangLuran/nlp-element-influence.

Text simplification is a common task where the text is adapted to make it easier to understand. Similarly, text elaboration can make a passage more sophisticated, offering a method to control the complexity of reading comprehension tests. However, text simplification and elaboration tasks are limited to only relatively alter the readability of texts. It is useful to directly modify the readability of any text to an absolute target readability level to cater to a diverse audience. Ideally, the readability of readability-controlled generated text should be independent of the source text. Therefore, we propose a novel readability-controlled text modification task. The task requires the generation of 8 versions at various target readability levels for each input text. We introduce novel readability-controlled text modification metrics. The baselines for this task use ChatGPT and Llama-2, with an extension approach introducing a two-step process (generating paraphrases by passing through the language model twice). The zero-shot approaches are able to push the readability of the paraphrases in the desired direction but the final readability remains correlated with the original text’s readability. We also find greater drops in semantic and lexical similarity between the source and target texts with greater shifts in the readability.

In this paper, we consider the challenge of summarizing patients medical progress notes in a limited data setting. For the Problem List Summarization (shared task 1A) at the BioNLP Workshop 2023, we demonstrate that ClinicalT5 fine-tuned to 765 medical clinic notes outperforms other extractive, abstractive and zero-shot baselines, yielding reasonable baseline systems for medical note summarization. Further, we introduce Hierarchical Ensemble of Summarization Models (HESM), consisting of token-level ensembles of diverse fine-tuned ClinicalT5 models, followed by Minimum Bayes Risk (MBR) decoding. Our HESM approach lead to a considerable summarization performance boost, and when evaluated on held-out challenge data achieved a ROUGE-L of 32.77, which was the best-performing system at the top of the shared task leaderboard.

Recently it has been shown that without any access to the contextual passage, multiple choice reading comprehension (MCRC) systems are able to answer questions significantly better than random on average. These systems use their accumulated “world knowledge” to directly answer questions, rather than using information from the passage. This paper examines the possibility of exploiting this observation as a tool for test designers to ensure that the form of “world knowledge” is acceptable for a particular set of questions. We propose information-theory based metrics that enable the level of “world knowledge” exploited by systems to be assessed. Two metrics are described: the expected number of options, which measures whether a passage-free system can identify the answer a question using world knowledge; and the contextual mutual information, which measures the importance of context for a given question. We demonstrate that questions with low expected number of options, and hence answerable by the shortcut system, are often similarly answerable by humans without context. This highlights that the general knowledge ‘shortcuts’ could be equally used by exam candidates, and that our proposed metrics may be helpful for future test designers to monitor the quality of questions.

Generative Large Language Models (LLMs) such as GPT-3 are capable of generating highly fluent responses to a wide variety of user prompts. However, LLMs are known to hallucinate facts and make non-factual statements which can undermine trust in their output. Existing fact-checking approaches either require access to the output probability distribution (which may not be available for systems such as ChatGPT) or external databases that are interfaced via separate, often complex, modules. In this work, we propose “SelfCheckGPT”, a simple sampling-based approach that can be used to fact-check the responses of black-box models in a zero-resource fashion, i.e. without an external database. SelfCheckGPT leverages the simple idea that if an LLM has knowledge of a given concept, sampled responses are likely to be similar and contain consistent facts. However, for hallucinated facts, stochastically sampled responses are likely to diverge and contradict one another. We investigate this approach by using GPT-3 to generate passages about individuals from the WikiBio dataset, and manually annotate the factuality of the generated passages. We demonstrate that SelfCheckGPT can: i) detect non-factual and factual sentences; and ii) rank passages in terms of factuality. We compare our approach to several baselines and show that our approach has considerably higher AUC-PR scores in sentence-level hallucination detection and higher correlation scores in passage-level factuality assessment compared to grey-box methods.

Adversarial attack research in natural language processing (NLP) has made significant progress in designing powerful attack methods and defence approaches. However, few efforts have sought to identify which source samples are the most attackable or robust, i.e. can we determine for an unseen target model, which samples are the most vulnerable to an adversarial attack. This work formally extends the definition of sample attackability/robustness for NLP attacks. Experiments on two popular NLP datasets, four state of the art models and four different NLP adversarial attack methods, demonstrate that sample uncertainty is insufficient for describing characteristics of attackable/robust samples and hence a deep learning based detector can perform much better at identifying the most attackable and robust samples for an unseen target model. Nevertheless, further analysis finds that there is little agreement in which samples are considered the most attackable/robust across different NLP attack methods, explaining a lack of portability of attackability detection methods across attack methods.

Multiple-choice tests are a common approach for assessing candidates’ comprehension skills. Standard multiple-choice reading comprehension exams require candidates to select the correct answer option from a discrete set based on a question in relation to a contextual passage. For appropriate assessment, the distractor answer options must by definition be incorrect but plausible and diverse. However, generating good quality distractors satisfying these criteria is a challenging task for content creators. We propose automated assessment metrics for the quality of distractors in multiple-choice reading comprehension tests. Specifically, we define quality in terms of the incorrectness, plausibility and diversity of the distractor options. We assess incorrectness using the classification ability of a binary multiple-choice reading comprehension system. Plausibility is assessed by considering the distractor confidence - the probability mass associated with the distractor options for a standard multi-class multiple-choice reading comprehension system. Diversity is assessed by pairwise comparison of an embedding-based equivalence metric between the distractors of a question. To further validate the plausibility metric we compare against candidate distributions over multiple-choice questions and agreement with a ChatGPT model’s interpretation of distractor plausibility and diversity.

Deep learning based systems are susceptible to adversarial attacks, where a small, imperceptible change at the input alters the model prediction. However, to date the majority of the approaches to detect these attacks have been designed for image processing systems. Many popular image adversarial detection approaches are able to identify adversarial examples from embedding feature spaces, whilst in the NLP domain existing state of the art detection approaches solely focus on input text features, without consideration of model embedding spaces. This work examines what differences result when porting these image designed strategies to Natural Language Processing (NLP) tasks - these detectors are found to not port over well. This is expected as NLP systems have a very different form of input: discrete and sequential in nature, rather than the continuous and fixed size inputs for images. As an equivalent model-focused NLP detection approach, this work proposes a simple sentence-embedding “residue” based detector to identify adversarial examples. On many tasks, it out-performs ported image domain detectors and recent state of the art NLP specific detectors.

Machine reading comprehension (MRC) has drawn a lot of attention as an approach for assessing the ability of systems to understand natural language. Usually systems focus on selecting the correct answer to a question given a contextual paragraph. However, for many applications of multiple-choice MRC systems there are two additional considerations. For multiple-choice exams there is often a negative marking scheme; there is a penalty for an incorrect answer. In terms of an MRC system this means that the system is required to have an idea of the uncertainty in the predicted answer. The second consideration is that many multiple-choice questions have the option of none-of-the-above (NOA) indicating that none of the answers is applicable, rather than there always being the correct answer in the list of choices. This paper investigates both of these issues by making use of predictive uncertainty. Whether the system should propose an answer is a direct application of answer uncertainty. There are two possibilities when considering the NOA option. The simplest is to explicitly build a system on data that includes this option. Alternatively uncertainty can be applied to detect whether the other options include the correct answer. If the system is not sufficiently confident it will select NOA. As there is no standard corpus available to investigate these topics, the ReClor corpus is modified by removing the correct answer from a subset of possible answers. A high-performance MRC system is used to evaluate whether answer uncertainty can be applied in these situations. It is shown that uncertainty does allow questions that the system is not confident about to be detected. Additionally it is shown that uncertainty outperforms a system explicitly built with an NOA option.

Grammatical error correction (GEC) systems are a useful tool for assessing a learner’s writing ability. These systems allow the grammatical proficiency of a candidate’s text to be assessed without requiring an examiner or teacher to read the text. A simple summary of a candidate’s ability can be measured by the total number of edits between the input text and the GEC system output: the fewer the edits the better the candidate. With advances in deep learning, GEC systems have become increasingly powerful and accurate. However, deep learning systems are susceptible to adversarial attacks, in which a small change at the input can cause large, undesired changes at the output. In the context of GEC for automated assessment, the aim of an attack can be to deceive the system into not correcting (concealing) grammatical errors to create the perception of higher language ability. An interesting aspect of adversarial attacks in this scenario is that the attack needs to be simple as it must be applied by, for example, a learner of English. The form of realistic attack examined in this work is appending the same phrase to each input sentence: a concatenative universal attack. The candidate only needs to learn a single attack phrase. State-of-the-art GEC systems are found to be susceptible to this form of simple attack, which transfers to different test sets as well as system architectures,

Machine learning systems have shown impressive performance across a range of natural language tasks. However, it has been hypothesized that these systems are prone to learning spurious correlations that may be present in the training data. Though these correlations will not impact in-domain performance, they are unlikely to generalize well to out-of-domain data, limiting the applicability of systems. This work examines this phenomenon on text classification tasks. Rather than artificially injecting features into the data, we demonstrate that real spurious correlations can be exploited by current state-of-the-art deep-learning systems. Specifically, we show that even when only ‘stop’ words are available at the input stage, it is possible to predict the class significantly better than random. Though it is shown that these stop words are not required for good in-domain performance, they can degrade the ability of the system to generalize well to out-of-domain data.

Spoken ‘grammatical error correction’ (SGEC) is an important process to provide feedback for second language learning. Due to a lack of end-to-end training data, SGEC is often implemented as a cascaded, modular system, consisting of speech recognition, disfluency removal, and grammatical error correction (GEC). This cascaded structure enables efficient use of training data for each module. It is, however, difficult to compare and evaluate the performance of individual modules as preceeding modules may introduce errors. For example the GEC module input depends on the output of non-native speech recognition and disfluency detection, both challenging tasks for learner data. This paper focuses on the assessment and development of SGEC systems. We first discuss metrics for evaluating SGEC, both individual modules and the overall system. The system-level metrics enable tuning for optimal system performance. A known issue in cascaded systems is error propagation between modules. To mitigate this problem semi-supervised approaches and self-distillation are investigated. Lastly, when SGEC system gets deployed it is important to give accurate feedback to users. Thus, we apply filtering to remove edits with low-confidence, aiming to improve overall feedback precision. The performance metrics are examined on a Linguaskill multi-level data set, which includes the original non-native speech, manual transcriptions and reference grammatical error corrections, to enable system analysis and development.

Transformer-based models have achieved state-of-the-art results in a wide range of natural language processing (NLP) tasks including document summarization. Typically these systems are trained by fine-tuning a large pre-trained model to the target task. One issue with these transformer-based models is that they do not scale well in terms of memory and compute requirements as the input length grows. Thus, for long document summarization, it can be challenging to train or fine-tune these models. In this work, we exploit large pre-trained transformer-based models and address long-span dependencies in abstractive summarization using two methods: local self-attention; and explicit content selection. These approaches are compared on a range of network configurations. Experiments are carried out on standard long-span summarization tasks, including Spotify Podcast, arXiv, and PubMed datasets. We demonstrate that by combining these methods, we can achieve state-of-the-art results on all three tasks in the ROUGE scores. Moreover, without a large-scale GPU card, our approach can achieve comparable or better results than existing approaches.

Transformer models have achieved state-of-the-art results in a wide range of NLP tasks including summarization. Training and inference using large transformer models can be computationally expensive. Previous work has focused on one important bottleneck, the quadratic self-attention mechanism in the encoder. Modified encoder architectures such as LED or LoBART use local attention patterns to address this problem for summarization. In contrast, this work focuses on the transformer’s encoder-decoder attention mechanism. The cost of this attention becomes more significant in inference or training approaches that require model-generated histories. First, we examine the complexity of the encoder-decoder attention. We demonstrate empirically that there is a sparse sentence structure in document summarization that can be exploited by constraining the attention mechanism to a subset of input sentences, whilst maintaining system performance. Second, we propose a modified architecture that selects the subset of sentences to constrain the encoder-decoder attention. Experiments are carried out on abstractive summarization tasks, including CNN/DailyMail, XSum, Spotify Podcast, and arXiv.

Increased demand to learn English for business and education has led to growing interest in automatic spoken language assessment and teaching systems. With this shift to automated approaches it is important that systems reliably assess all aspects of a candidate’s responses. This paper examines one form of spoken language assessment; whether the response from the candidate is relevant to the prompt provided. This will be referred to as off-topic spoken response detection. Two forms of previously proposed approaches are examined in this work: the hierarchical attention-based topic model (HATM); and the similarity grid model (SGM). The work focuses on the scenario when the prompt, and associated responses, have not been seen in the training data, enabling the system to be applied to new test scripts without the need to collect data or retrain the model. To improve the performance of the systems for unseen prompts, data augmentation based on easy data augmentation (EDA) and translation based approaches are applied. Additionally for the HATM, a form of prompt dropout is described. The systems were evaluated on both seen and unseen prompts from Linguaskill Business and General English tests. For unseen data the performance of the HATM was improved using data augmentation, in contrast to the SGM where no gains were obtained. The two approaches were found to be complementary to one another, yielding a combined F0.5 score of 0.814 for off-topic response detection where the prompts have not been seen in training.

There is a growing demand for automatic assessment of spoken English proficiency. These systems need to handle large variations in input data owing to the wide range of candidate skill levels and L1s, and errors from ASR. Some candidates will be a poor match to the training data set, undermining the validity of the predicted grade. For high stakes tests it is essential for such systems not only to grade well, but also to provide a measure of their uncertainty in their predictions, enabling rejection to human graders. Previous work examined Gaussian Process (GP) graders which, though successful, do not scale well with large data sets. Deep Neural Network (DNN) may also be used to provide uncertainty using Monte-Carlo Dropout (MCD). This paper proposes a novel method to yield uncertainty and compares it to GPs and DNNs with MCD. The proposed approach explicitly teaches a DNN to have low uncertainty on training data and high uncertainty on generated artificial data. On experiments conducted on data from the Business Language Testing Service (BULATS), the proposed approach is found to outperform GPs and DNNs with MCD in uncertainty-based rejection whilst achieving comparable grading performance.