EMNLP 2023 Papers

Scientific literature understanding tasks have gained significant attention due to their potential to accelerate scientific discovery. Pre-trained language models (LMs) have shown effectiveness in these tasks, especially when tuned via contrastive learning. However, jointly utilizing pre-training data across multiple heterogeneous tasks (e.g., extreme multi-label paper classification, citation prediction, and literature search) remains largely unexplored. To bridge this gap, we propose a multi-task contrastive learning framework, SciMult, with a focus on facilitating common knowledge sharing across different scientific literature understanding tasks while preventing task-specific skills from interfering with each other. To be specific, we explore two techniques -- task-aware specialization and instruction tuning. The former adopts a Mixture-of-Experts Transformer architecture with task-aware sub-layers; the latter prepends task-specific instructions to the input text so as to produce task-aware outputs. Extensive experiments on a comprehensive collection of benchmark datasets verify the effectiveness of our task-aware specialization strategy, where we outperform state-of-the-art scientific pre-trained LMs. Code, datasets, and pre-trained models can be found at https://scimult.github.io/.

Large Language Models (LLMs) are popular for their impressive abilities, but the need for model-specific fine-tuning or task-specific prompt engineering can hinder their generalization. We propose UPRISE (Universal Prompt Retrieval for Improving zero-Shot Evaluation), which tunes a lightweight and versatile retriever that automatically retrieves prompts for a given zero-shot task input. Specifically, we demonstrate universality in a cross-task and cross-model scenario: the retriever is tuned on diverse tasks, but tested on unseen task types; we use a small frozen LLM, GPT-Neo-2.7B, for tuning the retriever, but test the retriever on different LLMs of much larger scales, such as BLOOM-7.1B, OPT-66B and GPT3-175B. Additionally, we show that UPRISE mitigates the hallucination problem in our experiments with ChatGPT, suggesting its potential to improve even the strongest LLMs. Our model and code are available at https://github.com/microsoft/LMOps.

In task-oriented dialogs (TOD), reinforcement learning (RL) algorithms train a model to directly optimize response for task-related metrics. However, RL often needs to perform exploration, which can be time-consuming due to the slow auto-regressive sequence generation process. We investigate an approach to create a more efficient RL-based algorithm to improve TOD performance in an offline setting. First, we use a faster generation procedure that samples from independent next-word distributions after training the language model (LM) with supervised learning. We then introduce a fine-grained reward function to help the model focus on learning key information in a dialog, by measuring the importance and semantic closeness of each generated token. Experiments on the MultiWoZ dataset show our new training algorithm, Keywords Reinforcement Learning with Next-word Sampling (KRLS), achieves state-of-the-art performance on the end-to-end response generation task, with a 15% training time reduction compared to a standard RL algorithm using auto-regressive generation.

Although large language models (LLMs) are often pre-trained on large-scale multilingual texts, their reasoning abilities and real-world knowledge are mainly evaluated based on English datasets. Assessing LLM capabilities beyond English is increasingly vital but hindered due to the lack of suitable datasets. In this work, we introduce IndoMMLU, the first multi-task language understanding benchmark for Indonesian culture and languages, which consists of questions from primary school to university entrance exams in Indonesia. By employing professional teachers, we obtain 14,981 questions across 64 tasks and education levels, with 46% of the questions focusing on assessing proficiency in the Indonesian language and knowledge of nine local languages and cultures in Indonesia. Our empirical evaluations show that GPT-3.5 only manages to pass the Indonesian primary school level, with limited knowledge of local Indonesian languages and culture. Other smaller models such as BLOOMZ and Falcon perform at even lower levels.

Current NLP models appear to be achieving performance comparable to human capabilities on well-established benchmarks. New benchmarks are now necessary to test deeper layers of understanding of natural languages by these models. Blackbird's Language Matrices are a recently developed framework that draws inspiration from tests of human analytic intelligence. The BLM task has revealed that successful performances in previously studied linguistic problems do not yet stem from a deep understanding of the generative factors that define these problems. In this study, we define a new BLM task for predicate-argument structure, and develop a structured dataset for its investigation, concentrating on the spray-load verb alternations in English, as a case study. The context sentences include one alternant from the spray-load alternation and the target sentence is the other alternant, to be chosen among a minimally contrastive and adversarial set of answers. We describe the generation process of the dataset and the reasoning behind the generating rules. The dataset aims to facilitate investigations into how verb information is encoded in sentence embeddings and how models generalize to the complex properties of argument structures. Benchmarking experiments conducted on the dataset and qualitative error analysis on the answer set reveal the inherent challenges associated with the problem even for current high-performing representations.

A popular approach for improving the correctness of output from large language models (LLMs) is Self-Consistency - poll the LLM multiple times and output the most frequent solution. Existing Self-Consistency techniques always generate a constant number of samples per question, where a better approach will be to non-uniformly distribute the available budget based on the amount of agreement in the samples generated so far. In response, we introduce Adaptive-Consistency, a cost-efficient, model-agnostic technique that dynamically adjusts the number of samples per question using a lightweight stopping criterion. Our experiments over 17 reasoning and code generation datasets and three LLMs demonstrate that Adaptive-Consistency reduces sample budget by up to 7.9 times with an average accuracy drop of less than 0.1%

Enhancing the zero-shot performance of instruction-following models requires heavy computation, either by scaling the total number of training datasets or the model size. In this work, we explore how retrieval of soft prompts obtained through prompt tuning can efficiently assist hard prompts in zero-shot task generalization. Specifically, we train soft prompt embeddings for each prompt through prompt tuning, store the samples of the training instances mapped with the prompt embeddings, and retrieve the corresponding prompt embedding of the training instance closest to the query instance during inference. While only adding 0.007% additional parameters, retrieval of soft prompt enhances the performance of T0 on unseen tasks by outperforming it on 10 out of 11 datasets as well as improving the mean accuracy of T0 on BIG-bench benchmark by 2.39% points. Also, we report an interesting finding that retrieving source embeddings trained on similar answer choice formats is more important than those on similar task types.

For a language model (LM) to faithfully model human language, it must compress vast, potentially infinite information into relatively few dimensions. We propose analyzing compression in (pre-trained) LMs from two points of view: geometric and information-theoretic. We demonstrate that the two views are highly correlated, such that the intrinsic geometric dimension of linguistic data predicts their coding length under the LM. We then show that, in turn, high compression of a linguistic dataset predicts rapid adaptation to that dataset, confirming that being able to compress linguistic information is an important part of successful LM performance. As a practical byproduct of our analysis, we evaluate a battery of intrinsic dimension estimators for the first time on linguistic data, showing that only some encapsulate the relationship between information-theoretic compression, geometric compression, and ease-of-adaptation.

Comparative reasoning is a process of comparing objects, concepts, or entities to draw conclusions, which constitutes a fundamental cognitive ability. In this paper, we propose a novel framework to pre-train language models for enhancing their abilities of comparative reasoning over texts. While there have been approaches for NLP tasks that require comparative reasoning, they suffer from costly manual data labeling and limited generalizability to different tasks. Our approach introduces a novel method of collecting scalable data for text-based entity comparison, which leverages both structured and unstructured data. Moreover, we present a framework of pre-training language models via three novel objectives on comparative reasoning. Evaluation on downstream tasks including comparative question answering, question generation, and summarization shows that our pre-training framework significantly improves the comparative reasoning abilities of language models, especially under low-resource conditions. This work also releases the first integrated benchmark for comparative reasoning.

Machine translation (MT) quality estimation (QE) is a crucial task to estimate the quality of MT outputs when reference translations are unavailable. Many studies focus on generating pseudo data using large parallel corpus and achieve remarkable success in the supervised setting. However, pseudo data solutions are less satisfying in unsupervised scenarios because the pseudo labels are inaccurate or the pseudo translations differ from the real ones. To address these problems, we propose to generate pseudo data using the MT model with constrained beam search (CBSQE). CBSQE preserves the reference parts with high MT probabilities as correct translations, while the rest parts as the wrong ones for MT generation. Therefore, CBSQE can reduce the false negative labels caused by synonyms. Overall, beam search will prefer a more real hypothesis with a higher MT generation likelihood. Extensive experiments demonstrate that CBSQE outperforms strong baselines in both supervised and unsupervised settings. Analyses further show the superiority of CBSQE. The code is available at https://github.com/NJUNLP/njuqe.

Large language models (LLMs) encode vast amounts of world knowledge. However, since these models are trained on large swaths of internet data, they are at risk of inordinately capturing information about dominant groups. This imbalance can propagate into generated language. In this work, we study and operationalise a form of geographical erasure wherein language models underpredict certain countries. We demonstrate consistent instances of erasure across a range of LLMs. We discover that erasure strongly correlates with low frequencies of country mentions in the training corpus. Lastly, we mitigate erasure by finetuning using a custom objective.

Despite tremendous advances in AI, it remains a significant challenge to develop interactive task guidance systems that can offer situated, personalized guidance and assist humans in various tasks. These systems need to have a sophisticated understanding of the user as well as the environment, and make timely accurate decisions on when and what to say. To address this issue, we created a new multimodal benchmark dataset, Watch, Talk and Guide (WTaG) based on natural interaction between a human user and a human instructor. We further proposed two tasks: User and Environment Understanding, and Instructor Decision Making. We leveraged several foundation models to study to what extent these models can be quickly adapted to perceptually enabled task guidance. Our quantitative, qualitative, and human evaluation results show that these models can demonstrate fair performances in some cases with no task-specific training, but a fast and reliable adaptation remains a significant challenge. Our benchmark and baselines will provide a stepping stone for future work on situated task guidance.

How much private information do text embeddings reveal about the original text? We investigate the problem of embedding \textit{inversion}, reconstructing the full text represented in dense text embeddings. We frame the problem as controlled generation: generating text that, when reembedded, is close to a fixed point in latent space. We find that although a naive model conditioned on the embedding performs poorly, a multi-step method that iteratively corrects and re-embeds text is able to recover 92% of 32-token text inputs exactly. We train our model to decode text embeddings from two state-of-the-art embedding models, and also show that our model can recover important personal information (full names) from a dataset of clinical notes.

There have been a lot of interest in the scaling properties of Transformer models. However, not much has been done on the front of investigating the effect of scaling properties of different inductive biases and model architectures. Do model architectures scale differently? If so, how does inductive bias affect scaling behaviour? How does this influence upstream (pretraining) and downstream (transfer)? This paper conducts a systematic study of scaling behaviour of ten diverse model architectures such as Transformers, Switch Transformers, Universal Transformers, Dynamic convolutions, Performers, and recently proposed MLP-Mixers. Via extensive experiments, we show that (1) architecture is an indeed an important consideration when performing scaling and (2) the best performing model can fluctuate at different scales. We believe that the findings outlined in this work has significant implications to how model architectures are currently evaluated in the community.

Clinical trials are critical for drug development. Constructing the appropriate eligibility criteria (i.e., the inclusion/exclusion criteria for patient recruitment) is essential for the trial's success. Proper design of clinical trial protocols should consider similar precedent trials and their eligibility criteria to ensure sufficient patient coverage. In this paper, we present a method named AutoTrial to aid the design of clinical eligibility criteria using language models. It allows (1) controllable generation under instructions via a hybrid of discrete and neural prompting, (2) scalable knowledge incorporation via in-context learning, and (3) explicit reasoning chains to provide rationales for understanding the outputs. Experiments on over 70K clinical trials verify that AutoTrial generates high-quality criteria texts that are fluent and coherent and with high accuracy in capturing the relevant clinical concepts to the target trial. It is noteworthy that our method, with a much smaller parameter size, gains around 60% winning rate against the GPT-3.5 baselines via human evaluations.

Large language models (LLMs) demonstrate impressive multilingual capability, but their performance varies substantially across different languages. In this work, we introduce a simple yet effective method, called cross-lingual-thought prompting (XLT), to systematically improve the multilingual capability of LLMs. Specifically, XLT is a generic template prompt that stimulates cross-lingual and logical reasoning skills to enhance task performance across languages. We conduct comprehensive evaluations on 7 typical benchmarks related to reasoning, understanding, and generation tasks, covering both high-resource and low-resource languages. Experimental results show that XLT not only remarkably enhances the performance of various multilingual tasks but also significantly reduces the gap between the average performance and the best performance of each task in different languages. Notably, XLT brings over 10 points of average improvement in arithmetic reasoning and open-domain question-answering tasks.

Minimum Bayes risk (MBR) decoding outputs the hypothesis with the highest expected utility over the model distribution for some utility function. It has been shown to improve accuracy over beam search in conditional language generation problems and especially neural machine translation, in both human and automatic evaluations. However, the standard sampling-based algorithm for MBR is substantially more computationally expensive than beam search, requiring a large number of samples as well as a quadratic number of calls to the utility function, limiting its applicability. We describe an algorithm for MBR which gradually grows the number of samples used to estimate the utility while pruning hypotheses that are unlikely to have the highest utility according to confidence estimates obtained with bootstrap sampling. Our method requires fewer samples and drastically reduces the number of calls to the utility function compared to standard MBR while being statistically indistinguishable in terms of accuracy. We demonstrate the effectiveness of our approach in experiments on three language pairs, using chrF++ and COMET as utility/evaluation metrics.

With the rapid progress of Large language models (LLMs) and the huge amount of text they generate, it becomes impractical to manually distinguish whether a text is machine-generated. The growing use of LLMs in social media and education, prompts us to develop methods to detect machine-generated text, preventing malicious use such as plagiarism, misinformation, and propaganda. In this paper, we introduce two novel zero-shot methods for detecting machine-generated text by leveraging the Log-Rank information. One is called DetectLLM-LRR, which is fast and efficient, and the other is called DetectLLM-NPR, which is more accurate, but slower due to the need for perturbations. Our experiments on three datasets and seven language models show that our proposed methods improve over the state of the art by 3.9 and 1.75 AUROC points absolute. Moreover, DetectLLM-NPR needs fewer perturbations than previous work to achieve the same level of performance, which makes it more practical for real-world use. We also investigate the efficiency-performance trade-off based on users' preference for these two measures and provide intuition for using them in practice effectively. We release the data and the code of both methods in https://github.com/mbzuai-nlp/DetectLLM.

The recent advent of end-to-end generative retrieval marks a significant shift in document retrieval methods, leveraging differentiable search indexes to directly produce relevant document identifiers (docids) in response to a specific query. Nevertheless, this approach faces two fundamental challenges: (i) a discrepancy between the token-level probabilistic optimization and the broader document-level relevance estimation; (ii) an overemphasis on top-1 results at the expense of overall ranking quality. To tackle these challenges, we propose a generative retrieval model with reinforcement learning from relevance feedback, which aims to align token-level docid generation with document-level relevance estimation. The training process incorporates three stages: supervised fine-tuning, relevance reward model training, and reinforced learning-to-rank from relevance feedback. To train a high-quality reward model, we define "relevance" under three progressive scenarios, which collectively offer a comprehensive evaluation of the document relevance. Experiments conducted on two benchmark datasets demonstrate the effectiveness of our proposed approach.