NAACL 2024 Papers

Off-the-shelf pre-trained language models have become the de facto standard in NLP pipelines for a multitude of downstream tasks. However, the inability of these models to properly encode numerals limits their performance on tasks requiring numeric comprehension. We introduce strategies to semantically prime numerals in any corpus by generating anchors governed by the distribution of numerals in said corpus, thereby enabling mathematically grounded representations of these numeral tokens. We establish the superiority of our proposed techniques through evaluation on a range of numeracy tasks for both in-domain (seen) and out-domain (unseen) numerals. Further, we expand our empirical evaluations to numerals ranging from 1 to 10 billion, a significantly broader range compared to previous studies of the same nature, and we demonstrate significant improvements in the mathematical grounding of our learned embeddings.

In the current user-server interaction paradigm of prompted generation with large language models (LLMs) on cloud, the server fully controls the generation process, which leaves zero options for users who want to keep the generated text private to themselves. For privacy-aware text generation on cloud, we propose LatticeGen, a cooperative protocol in which the server still handles most of the computation while the client controls the sampling operation. The key idea is that the true generated sequence is mixed with noise tokens by the client and hidden in a noised lattice. Only the client knows which tokens are the true ones. Considering potential attacks from a hypothetically malicious server and how the client can defend against it, we propose the repeated beam-search attack and the mixing noise scheme. In our experiments we apply LatticeGen to protect both prompt and generation. It is shown that while the noised lattice degrades generation quality, LatticeGen successfully protects the true generation to a remarkable degree under strong attacks (more than 50% of the semantic remains hidden as measured by BERTScore).

Recently, large language models (LLMs) have emerged as a groundbreaking technology and their unparalleled text generation capabilities have sparked interest in their application to the fundamental sentence representation learning task. Existing methods have explored utilizing LLMs as data annotators to generate synthesized data for training contrastive learning based sentence embedding models such as SimCSE. However, since contrastive learning models are sensitive to the quality of sentence pairs, the effectiveness of these methods is largely influenced by the content generated from LLMs, highlighting the need for more refined generation in the context of sentence representation learning. Building upon this premise, we propose MultiCSR, a multi-level contrastive sentence representation learning framework that decomposes the process of prompting LLMs to generate a corpus for training base sentence embedding models into three stages (i.e., sentence generation, sentence pair construction, in-batch training) and refines the generated content at these three distinct stages, ensuring only high-quality sentence pairs are utilized to train a base contrastive learning model. Our extensive experiments reveal that MultiCSR enables a less advanced LLM to surpass the performance of ChatGPT, while applying it to ChatGPT achieves better state-of-the-art results. Comprehensive analyses further underscore the potential of our framework in various application scenarios and achieving better sentence representation learning with LLMs.

Ranking documents using Large Language Models (LLMs) by directly feeding the query and candidate documents into the prompt is an interesting and practical problem. However, researchers have found it difficult to outperform fine-tuned baseline rankers on benchmark datasets.We analyze pointwise and listwise ranking prompts used by existing methods and argue that off-the-shelf LLMs do not fully understand these challenging ranking formulations. In this paper, we propose to significantly reduce the burden on LLMs by using a new technique called Pairwise Ranking Prompting (PRP).Our results are the first in the literature to achieve state-of-the-art ranking performance on standard benchmarks using moderate-sized open-sourced LLMs. On TREC-DL 2019&2020, PRP based on the Flan-UL2 model with 20B parameters performs favorably with the previous best approach in the literature, which is based on the blackbox commercial GPT-4 that has 50x (estimated) model size, while outperforming other LLM-based solutions, such as InstructGPT which has 175B parameters, by over 10% for all ranking metrics. By using the same prompt template on seven BEIR tasks, PRP outperforms supervised baselines and outperforms the blackbox commercial ChatGPT solution by 4.2% and pointwise LLM-based solutions by more than 10% on average NDCG@10.Furthermore, we propose several variants of PRP to improve efficiency and show that it is possible to achieve competitive results even with linear complexity.

Large Language Models (LLMs) have demonstrated remarkable capabilities in various NLP tasks. However, previous works have shown these models are sensitive towards prompt wording, and few-shot demonstrations and their order, posing challenges to fair assessment of these models. As these models become more powerful, it becomes imperative to understand and address these limitations. In this paper, we focus on LLMs robustness on the task of multiple-choice questions—commonly adopted task to study reasoning and fact-retrieving capability of LLMs. Investigating the sensitivity of LLMs towards the order of options in multiple-choice questions, we demonstrate a considerable performance gap of approximately 13% to 85% in LLMs on different benchmarks, when answer options are reordered, even when using demonstrations in a few-shot setting. Through a detailed analysis, we conjecture that this sensitivity arises when LLMs are uncertain about the prediction between the top-2/3 choices, and specific options placements may favor certain prediction between those top choices depending on the question caused by positional bias. We also identify patterns in top-2 choices that amplify or mitigate the model’s bias toward option placement. We found that for amplifying bias, the optimal strategy involves positioning the top two choices as the first and last options. Conversely, to mitigate bias, we recommend placing these choices among the adjacent options. To validate our conjecture, we conduct various experiments and adopt two approaches to calibrate LLMs’ predictions, leading to up to 8 percentage points improvement across different models and benchmarks.

Large Language Models (LLMs) are increasingly used for accessing information on the web. Their truthfulness and factuality are thus of great interest. To help users make the right decisions about the information they get, LLMs should not only provide information but also help users fact-check it. We conduct human experiments with 80 crowdworkers to compare language models with search engines (information retrieval systems) at facilitating fact-checking. We prompt LLMs to validate a given claim and provide corresponding explanations. Users reading LLM explanations are significantly more efficient than those using search engines while achieving similar accuracy. However, they over-rely on the LLMs when the explanation is wrong. To reduce over-reliance on LLMs, we ask LLMs to provide contrastive information—explain both why the claim is true and false, and then we present both sides of the explanation to users. This contrastive explanation mitigates users’ over-reliance on LLMs, but cannot significantly outperform search engines. Further, showing both search engine results and LLM explanations offers no complementary benefits compared to search engines alone. Taken together, our study highlights that natural language explanations by LLMs may not be a reliable replacement for reading the retrieved passages, especially in high-stakes settings where over-relying on wrong AI explanations could lead to critical consequences.

Healthcare tasks such as predicting clinical outcomes across medical and surgical populations, disease prediction, predicting patient health journeys, are typically approached with supervised learning on task-specific datasets. We demonstrate that language models begin to learn these tasks without any explicit supervision when trained on a new dataset of billions of administrative claims, which essentially encapsulates the practice of medicine, offering a unique perspective on patient care and treatment patterns. Our model, MediClaimGPT, a 125M parameter Transformer demonstrates strong zero-shot predictive capabilities, accurately forecasting patient health events across four evaluation datasets, with its capabilities further demonstrated in various downstream tasks. A significant application of MediClaimGPT is in generating high-quality, clinically plausible synthetic claims data, enhancing healthcare data utility while preserving patient privacy. This research underscores the potential of language models in handling complex datasets and their strategic application in healthcare and related fields.

As research in human-centered NLP advances, there is a growing recognition of the importance of incorporating human and social factors into NLP models. At the same time, our NLP systems have become heavily reliant on LLMs, most of which do not model authors. To build NLP systems that can truly understand human language, we must better integrate human contexts into LLMs. This brings to the fore a range of design considerations and challenges in terms of what human aspects to capture, how to represent them, and what modeling strategies to pursue. To address these, we advocate for three positions toward creating large human language models (LHLMs) using concepts from psychological and behavioral sciences: First, LM training should include the human context. Second, LHLMs should recognize that people are more than their group(s). Third, LHLMs should be able to account for the dynamic and temporally-dependent nature of the human context. We refer to relevant advances and present open challenges that need to be addressed and their possible solutions in realizing these goals.

Recent works have demonstrated success in controlling sentence attributes (e.g., sentiment) and structure (e.g., syntactic structure) based on the diffusion language model. A key component that drives theimpressive performance for generating high-quality samples from noise is iteratively denoise for thousands of steps. While beneficial, the complexity of starting from the noise and the learning steps has limited its implementation to many NLP real-world applications. This paper proposes Language Rectified Flow (LF).Our method is based on the reformulation of the standard probabilistic flow models.Language rectified flow learns (neural) ordinary differentialequation models to transport between the source distribution and the target distribution, henceproviding a unified and effective solution to generative modeling and domain transfer.From the source distribution, our language rectified flow yields fast simulation and effectively decreases the inference time. Experiments on three challenging fine-grained control tasks and multiple high-quality text editing show that our method consistently outperforms its baselines. Extensive experiments and ablation studies demonstrate that our method can be general, effective, and beneficial for many NLP tasks.

Finetuning pretrained models on downstream generation tasks often leads to catastrophic forgetting in zero-shot conditions. In this work, we focus on summarization and tackle the problem through the lens of language-independent representations. After training on monolingual summarization, we perform zero-shot transfer to new languages or language pairs. We first show naively finetuned models are highly language-specific in both output behavior and internal representations, resulting in poor zero-shot performance. Next, we propose query-key (QK) finetuning to decouple task-specific knowledge from the pretrained language generation abilities. Then, after showing downsides of the standard adversarial language classifier, we propose a balanced variant that more directly enforces language-agnostic representations. Moreover, our qualitative analyses show removing source language identity correlates to zero-shot summarization performance. Our code is openly available.

Logical reasoning is a fundamental aspect of human intelligence and a key component of tasks like problem-solving and decision-making. Recent advancements have enabled Large Language Models (LLMs) to potentially exhibit reasoning capabilities, but complex logical reasoning remains a challenge. The state-of-the-art, solver-augmented language models, use LLMs to parse natural language logical questions into symbolic representations first and then adopt external logical solvers to take in the symbolic representations and output the answers. Despite their impressive performance, any parsing errors will inevitably result in the failure of the execution of external logical solvers and no answer to the logical questions. In this paper, we introduce LoGiPT, a novel language model that directly internalizes and emulates the reasoning processes of logical solvers and avoids parsing errors by learning strict adherence to solver syntax and grammar. LoGiPT is fine-tuned on a newly constructed instruction-tuning dataset derived from revealing and refining the invisible reasoning process of deductive solvers. Experimental results on two public deductive reasoning benchmarks show that LoGiPT outperforms state-of-the-art solver-augmented LMs and few-shot prompting methods on competitive LLMs like GPT-4. This project is available in https://github.com/Cyril-JZ/LoGiPT.

In difficult decision-making scenarios, it is common to have conflicting opinions among expert human decision-makers as there may not be a single right answer. Such decisions may be guided by different attributes that can be used to characterize an individual’s decision. We introduce a novel dataset for medical triage decision-making, labeled with a set of decision-maker attributes (DMAs). This dataset consists of 62 scenarios, covering six different DMAs, including ethical principles such as fairness and moral desert. We present a novel software framework for human-aligned decision-making by utilizing these DMAs, paving the way for trustworthy AI with better guardrails. Specifically, we demonstrate how large language models (LLMs) can serve as ethical decision-makers, and how their decisions can be aligned to different DMAs using zero-shot prompting. Our experiments focus on different open-source models with varying sizes and training techniques, such as Falcon, Mistral, and Llama 2. Finally, we also introduce a new form of weighted self-consistency that improves the overall quantified performance. Our results provide new research directions in the use of LLMs as alignable decision-makers. The dataset and open-source software are publicly available at: https://github.com/ITM-Kitware/llm-alignable-dm.

A primary criticism towards language models (LMs) is their inscrutability. This paper presents evidence that, despite their size and complexity, LMs sometimes exploit a simple vector arithmetic style mechanism to solve some relational tasks using regularities encoded in the hidden space of the model (e.g., Poland:Warsaw::China:Beijing). We investigate a range of language model sizes (from 124M parameters to 176B parameters) in an in-context learning setting, and find that for a variety of tasks (involving capital cities, uppercasing, and past-tensing) a key part of the mechanism reduces to a simple additive update typically applied by the feedforward (FFN) networks. We further show that this mechanism is specific to tasks that require retrieval from pretraining memory, rather than retrieval from local context. Our results contribute to a growing body of work on the interpretability of LMs, and offer reason to be optimistic that, despite the massive and non-linear nature of the models, the strategies they ultimately use to solve tasks can sometimes reduce to familiar and even intuitive algorithms.

Recent progress in natural language processing (NLP) owes much to remarkable advances in large language models (LLMs). Nevertheless, LLMs frequently “hallucinate,” resulting in non-factual outputs. Our carefully-designed human evaluation substantiates the serious hallucination issue, revealing that even GPT-3.5 produces factual outputs less than 25% of the time. This underscores the importance of fact verifiers in order to measure and incentivize progress. Our systematic investigation affirms that LLMs can be repurposed as effective fact verifiers with strong correlations with human judgments. Surprisingly, FLAN-T5-11B , the least factual generator in our study, performs the best as a fact verifier, even outperforming more capable LLMs like GPT3.5 and ChatGPT. Delving deeper, we analyze the reliance of these LLMs on high-quality evidence, as well as their deficiencies in robustness and generalization ability. Our study presents insights for developing trustworthy generation models.

Situations and events evoke emotions in humans, but to what extent do they inform the prediction of emotion detection models? This work investigates how well human-annotated emotion triggers correlate with features that models deemed salient in their prediction of emotions. First, we introduce a novel dataset EmoTrigger, consisting of 900 social media posts sourced from three different datasets; these were annotated by experts for emotion triggers with high agreement. Using EmoTrigger, we evaluate the ability of large language models (LLMs) to identify emotion triggers, and conduct a comparative analysis of the features considered important for these tasks between LLMs and fine-tuned models. Our analysis reveals that emotion triggers are largely not considered salient features for emotion prediction models, instead there is intricate interplay between various features and the task of emotion detection.

Previous methods based on Large Language Models (LLM) perform unsupervised dependency parsing by maximizing bi-lexical dependence scores. However, these previous methods adopt dependence scores that are difficult to interpret. These methods cannot incorporate grammatical constraints that previous grammar-based parsing research has shown beneficial to improving parsing performance. In this work, we apply Conditional Mutual Information (CMI), an interpretable metric, to measure the bi-lexical dependence and incorporate grammatical constraints into LLM-based unsupervised parsing. We incorporate Part-Of-Speech information as a grammatical constraint at the CMI estimation stage and integrate two additional grammatical constraints at the subsequent tree decoding stage. We find that the CMI score positively correlates with syntactic dependencies and has a stronger correlation with the syntactic dependencies than baseline scores. Our experiment confirms the benefits and applicability of the proposed grammatical constraints across five languages and eight datasets. The CMI parsing model outperforms state-of-the-art LLM-based models and similarly constrained grammar-based models. Our analysis reveals that the CMI model is strong in retrieving dependency relations with rich lexical interactions but is weak in retrieving relations with sparse lexical interactions, indicating a potential limitation in CMI-based unsupervised parsing methods.

Recently, code language models have achieved notable advancements in addressing a diverse array of essential code comprehension and generation tasks. Yet, the field lacks a comprehensive deep dive and understanding of the code embeddings of multilingual code models. In this paper, we present a comprehensive study on multilingual code embeddings, focusing on the cross-lingual capabilities of these embeddings across different programming languages. Through probing experiments, we demonstrate that code embeddings comprise two distinct components: one deeply tied to the nuances and syntax of a specific language, and the other remaining agnostic to these details, primarily focusing on semantics. Further, we show that when we isolate and eliminate this language-specific component, we witness significant improvements in downstream code retrieval tasks, leading to an absolute increase of up to +17 in the Mean Reciprocal Rank (MRR).

We explore the use of language as a perceptual representation for vision-and-language navigation (VLN), with a focus on low-data settings. Our approach uses off-the-shelf vision systems for image captioning and object detection to convert an agent’s egocentric panoramic view at each time step into natural language descriptions. We then finetune a pretrained language model to select an action, based on the current view and the trajectory history, that would best fulfill the navigation instructions. In contrast to the standard setup which adapts a pretrained language model to work directly with continuous visual features from pretrained vision models, our approach instead uses (discrete) language as the perceptual representation. We explore several use cases of our language-based navigation (LangNav) approach on the R2R VLN benchmark: generating synthetic trajectories from a prompted language model (GPT-4) with which to finetune a smaller language model; domain transfer where we transfer a policy learned on one simulated environment (ALFRED) to another (more realistic) environment (R2R); and combining both vision- and language-based representations for VLN. Our approach is found to improve upon baselines that rely on visual features in settings where only a few expert trajectories (10-100) are available, demonstrating the potential of language as a perceptual representation for navigation.

Diffusion models have exhibited remarkable capabilities in text-to-image generation. However, their performance in image-to-text generation, specifically image captioning, has lagged behind Auto-Regressive (AR) models, casting doubt on their applicability for such tasks. In this work, we revisit diffusion models, highlighting their capacity for holistic context modeling and parallel decoding. With these benefits, diffusion models can alleviate the inherent limitations of AR methods, including their slow inference speed, error propagation, and unidirectional constraints. Furthermore, we identify the prior underperformance of diffusion models stemming from the absence of an effective latent space for image-text alignment, and the discrepancy between continuous diffusion processes and discrete textual data. In response, we introduce a novel architecture, LaDiC, which utilizes a split BERT to create a dedicated latent space for captions and integrates a regularization module to manage varying text lengths. Our framework also includes a diffuser for semantic image-to-text conversion and a Back&Refine technique to enhance token interactivity during inference. LaDiC achieves state-of-the-art performance for diffusion-based methods on the MS COCO dataset with 38.2 BLEU@4 and 126.2 CIDEr, demonstrating exceptional performance without pre-training or ancillary modules. This indicates strong competitiveness with AR models, revealing the previously untapped potential of diffusion models in image-to-text generation.