NAACL 2024 Papers

This paper introduces Q-tuning, a novel approach for continual prompt tuning that enables the lifelong learning of a pre-trained language model. When learning a new task, Q-tuning trains a task-specific prompt by adding it to a prompt queue consisting of the prompts from older tasks. To better transfer the knowledge of old tasks, we design an adaptive knowledge aggregation technique that reweighs previous prompts in the queue with a learnable low-rank matrix. Once the prompt queue reaches its maximum capacity, we leverage a PCA-based eviction rule to reduce the queue’s size, allowing the newly trained prompt to be added while preserving the primary knowledge of old tasks. In order to mitigate the accumulation of information loss caused by the eviction, we additionally propose a globally shared prefix prompt and a memory retention regularization based on information theory. Extensive experiments demonstrate that our approach outperforms the state-of-the-art methods substantially on continual prompt tuning benchmarks. Moreover, our approach enables lifelong learning on linearly growing task sequences while requiring constant complexity for training and inference.

Instruction tuning aligns the response of large language models (LLMs) with human preferences.Despite such efforts in human–LLM alignment, we find that instruction tuning does not always make LLMs human-like from a cognitive modeling perspective. More specifically, next-word probabilities estimated by instruction-tuned LLMs are often worse at simulating human reading behavior than those estimated by base LLMs.In addition, we explore prompting methodologies for simulating human reading behavior with LLMs. Our results show that prompts reflecting a particular linguistic hypothesis improve psychometric predictive power, but are still inferior to small base models.These findings highlight that recent advancements in LLMs, i.e., instruction tuning and prompting, do not offer better estimates than direct probability measurements from base LLMs in cognitive modeling. In other words, pure next-word probability remains a strong predictor for human reading behavior, even in the age of LLMs.

Large Language Models (LLMs) have exhibited remarkable proficiency across a wide array of NLP tasks. However, the escalation in model size also engenders substantial deployment costs. While few efforts have explored model pruning techniques to reduce the size of LLMs, they mainly center on general or task-specific weights. This leads to suboptimal performance due to lacking specificity on the target domain or generality on different tasks when applied to domain-specific challenges. This work introduces an innovative unstructured dual-pruning methodology, D-Pruner, for domain-specific compression on LLM. It extracts a compressed, domain-specific, and task- agnostic LLM by identifying LLM weights that are pivotal for general capabilities, like linguistic capability and multi-task solving, and domain-specific knowledge. More specifically, we first assess general weight importance by quantifying the error incurred upon their removal with the help of an open-domain calibration dataset. Then, we utilize this general weight importance to refine the training loss, so that it preserves generality when fitting into a specific domain. Moreover, by efficiently approximating weight importance with the refined training loss on a domain-specific calibration dataset, we obtain a pruned model emphasizing generality and specificity. Our comprehensive experiments across various tasks in healthcare and legal domains show the effectiveness of D-Pruner in domain-specific compression. Our code is available at https://github.com/psunlpgroup/D-Pruner.

Prompt-based methods have been used extensively across NLP to build zero- and few-shot label predictors. Many NLP tasks are naturally structured: that is, their outputs consist of multiple labels which constrain each other. Annotating data for such tasks can be cumbersome. Can the promise of the prompt-based paradigm be extended to such structured outputs? In this paper, we present a framework for constructing zero- and few-shot linguistic structure predictors. Our key insight is that we can use structural constraints—and combinatorial inference derived from them—to filter out inconsistent structures predicted by large language models. We instantiated this framework on two structured prediction tasks, and five datasets. Across all cases, our results show that enforcing consistency not only constructs structurally valid outputs, but also improves performance over the unconstrained variants.

Referring Expression Generation (REG) is the task of generating a description that unambiguously identifies a given target in the scene. Different from Image Captioning (IC), REG requires learning fine-grained characteristics of not only the scene objects but also their surrounding context. Referring expressions are usually not singular; an object can often be uniquely referenced in numerous ways, for instance, by color, by location, or by relationship with other objects. Most prior works, however, have not explored this ‘aspect-based multiplicity’ of referring expressions. Hence, in this work, we focus on the Aspect-Controlled REG task, which requires generating a referring expression conditioned on the input aspect(s), where an aspect captures a style of reference. By changing the input aspect such as color, location, action etc., one can generate multiple distinct expressions per target region. To solve this new task, we first modify BLIP for aligning image-regions and text-expressions. We achieve this through a novel approach for feeding the input by drawing a bounding box around the target image-region and prompting the model to generate the referring expression. Our base REG model already beats all prior works in CIDEr score. To tackle Aspect-Controlled REG, we append ‘aspect tokens’ to the prompt and show that distinct expressions can be generated by just changing the prompt. Finally, to prove the high-quality and diversity of the data generated by our proposed aspect-controlled REG model, we also perform data-augmentation-based evaluation on the downstream Referring Expression Comprehension (REC) task. With just half of the real data augmented with the generated synthetic data, we achieve performance comparable to training with 100% of real data, using a SOTA REC model.

Given several documents, multi-hop question generation (MQG) is a task aims to generate complicated questions that require reasoning over multiple pieces of these documents to find the answer. To perform this task, existing studies focus on designing advanced architectures to locate essential keywords or sentences in multiple documents and then generate questions accordingly, where they normally do not note that question types could provide crucial hints for extracting key information from the documents for MQG. In general, supervised approaches are used that rely on large annotated data, which is not available in many low-resource scenarios and thus makes MQG hard in these domains. Consider the recent success of large language models (LLMs) on natural language processing tasks using limited labeled data under few-shot settings, in this paper, we propose an approach named type-aware semantics extraction-based chain-of-thought method (TASE-CoT) for few-shot MQG. Specifically, our approach firstly extracts question types and essential semantic phrases from the given documents and the answer. Then, we design a three-step CoT template to leverage the extracted question type and semantic phrases to predict multi-hop questions. Extensive experiments and the results demonstrate the effectiveness of our approach and the proposed modules.

Pre-trained language models (PLMs) have attracted enormous attention over the past few years with their unparalleled performances. Meanwhile, the soaring cost to train PLMs as well as their amazing generalizability have jointly contributed to few-shot fine-tuning and prompting as the most popular training paradigms for natural language processing (NLP) models. Nevertheless, existing studies have shown that these NLP models can be backdoored such that model behavior is manipulated when trigger tokens are presented.In this paper, we propose PromptFix, a novel backdoor mitigation strategy for NLP models via adversarial prompt-tuning in few-shot settings.Unlike existing NLP backdoor removal methods, which rely on accurate trigger inversion and subsequent model fine-tuning, PromptFix keeps the model parameters intact and only utilizes two extra sets of soft tokens which approximate the trigger and counteract it respectively. The use of soft tokens and adversarial optimization eliminates the need to enumerate possible backdoor configurations and enables an adaptive balance between trigger finding and preservation of performance.Experiments with various backdoor attacks validate the effectiveness of the proposed method and the performances when domain shift is present further shows PromptFix’s applicability to models pretrained on unknown data source which is the common case in prompt tuning scenarios.

Recent singing-voice-synthesis (SVS) methods have achieved remarkable audio quality and naturalness, yet they lack the capability to control the style attributes of the synthesized singing explicitly. We propose Prompt-Singer, the first SVS method that enables attribute controlling on singer gender, vocal range and volume with natural language. We adopt a model architecture based on a decoder-only transformer with a multi-scale hierarchy, and design a range-melody decoupled pitch representation that enables text-conditioned vocal range control while keeping melodic accuracy. Furthermore, we explore various experiment settings, including different types of text representations, text encoder fine-tuning, and introducing speech data to alleviate data scarcity, aiming to facilitate further research. Experiments show that our model achieves favorable controlling ability and audio quality. Audio samples are available at http://prompt-singer.github.io .

We introduce Prompt Tuned Embedding Classification (PTEC) for classifying companies within an investment firm’s proprietary industry taxonomy, supporting their thematic investment strategy. PTEC assigns companies to the sectors they primarily operate in, conceptualizing this process as a multi-label text classification task. Prompt Tuning, usually deployed as a text-to-text (T2T) classification approach, ensures low computational cost while maintaining high task performance. However, T2T classification has limitations on multi-label tasks due to the generation of non-existing labels, permutation invariance of the label sequence, and a lack of confidence scores. PTEC addresses these limitations by utilizing a classification head in place of the Large Language Models (LLMs) language head. PTEC surpasses both baselines and human performance while lowering computational demands. This indicates the continuing need to adapt state-of-the-art methods to domain-specific tasks, even in the era of LLMs with strong generalization abilities.

Prompt engineering is an essential technique for enhancing the abilities of large language models (LLMs) by providing explicit and specific instructions. It enables LLMs to excel in various tasks, such as arithmetic reasoning, question answering, summarization, relation extraction, machine translation, and sentiment analysis. Researchers have been actively exploring different prompt engineering strategies, such as Chain of Thought (CoT), Zero-CoT, and In-context learning. However, an unresolved problem arises from the fact that current approaches lack a solid mathematical solution for determining optimal prompts. To address this issue in prompt engineering, we propose a new and effective approach called Prompt Space. Our methodology utilizes text embeddings to obtain basis vectors by matrix decomposition, and then constructs a space for representing all prompts. Prompt Space significantly outperforms state-of-the-art prompt paradigms on ten public reasoning benchmarks. Notably, without the help of the CoT method and the prompt “Let’s think step by step”, Prompt Space shows superior performance over the few-shot method. Overall, our approach provides a robust and effective mathematical framework for selecting simple and effective prompts. This advancement marks a significant step towards improving prompt engineering for a wide variety of applications in LLMs. Our code is publicly available at https://github.com/YouBLEI/Prompt-Space

Prior work shows that program-aided reasoning, in which large language models (LLMs) are combined with programs written in programming languages such as Python, can significantly improve accuracy on various reasoning tasks. However, while accuracy is essential, it is also important for such reasoners to “know what they know”, which can be quantified through the calibration of the model. In this paper, we compare the calibration of Program Aided Language Models (PAL) and text-based Chain-of-thought (COT) prompting techniques over 5 datasets and 2 model types - LLaMA models and OpenAI models. Our results indicate that PAL leads to improved calibration in 75% of the instances. Our analysis uncovers that prompting styles that produce lesser diversity in generations also have more calibrated results, and thus we also experiment with inducing lower generation diversity using temperature scaling and find that for certain temperatures, PAL is not only more accurate but is also more calibrated than COT. Overall, we demonstrate that, in the majority of cases, program-aided reasoners better know what they know than text-based counterparts.

In e-commerce, opinion summarization is the process of summarizing the consensus opinions found in product reviews. However, the potential of additional sources such as product description and question-answers (QA) has been considered less often. Moreover, the absence of any supervised training data makes this task challenging. To address this, we propose a novel synthetic dataset creation (SDC) strategy that leverages information from reviews as well as additional sources for selecting one of the reviews as a pseudo-summary to enable supervised training. Our Multi-Encoder Decoder framework for Opinion Summarization (MEDOS) employs a separate encoder for each source, enabling effective selection of information while generating the summary. For evaluation, due to the unavailability of test sets with additional sources, we extend the Amazon, Oposum+, and Flipkart test sets and leverage ChatGPT to annotate summaries. Experiments across nine test sets demonstrate that the combination of our SDC approach and MEDOS model achieves on average a 14.5% improvement in ROUGE-1 F1 over the SOTA. Moreover, comparative analysis underlines the significance of incorporating additional sources for generating more informative summaries. Human evaluations further indicate that MEDOS scores relatively higher in coherence and fluency with 0.41 and 0.5 (−1 to 1) respectively, compared to existing models. To the best of our knowledge, we are the first to generate opinion summaries leveraging additional sources in a self-supervised setting.

We investigate how pretrained language models (PLM) encode the grammatical category of verbal aspect in Russian. Encoding of aspect in transformer LMs has not been studied previously in any language. A particular challenge is posed by ”alternative contexts”: where either the perfective or the imperfective aspect is suitable grammatically and semantically. We perform probing using BERT and RoBERTa on alternative and non-alternative contexts. First, we assess the models’ performance on aspect prediction, via behavioral probing. Next, we examine the models’ performance when their contextual representations are substituted with counterfactual representations, via causal probing. These counterfactuals alter the value of the “boundedness” feature—a semantic feature, which characterizes the action in the context. Experiments show that BERT and RoBERTa do encode aspect—mostly in their final layers. The counterfactual interventions affect perfective and imperfective in opposite ways, which is consistent with grammar: perfective is positively affected by adding the meaning of boundedness, and vice versa. The practical implications of our probing results are that fine-tuning only the last layers of BERT on predicting aspect is faster and more effective than fine-tuning the whole model. The model has high predictive uncertainty about aspect in alternative contexts, which tend to lack explicit hints about the boundedness of the described action.

The NLP community typically relies on performance of a model on a held-out test set to assess generalization. Performance drops observed in datasets outside of official test sets are generally attributed to “out-of-distribution” effects. Here, we explore the foundations of generalizability and study the factors that affect it, articulating lessons from clinical studies. In clinical research, generalizability is an act of reasoning that depends on (a) *internal validity* of experiments to ensure controlled measurement of cause and effect, and (b) *external validity* or transportability of the results to the wider population. We demonstrate how learning spurious correlations, such as the distance between entities in relation extraction tasks, can affect a model’s internal validity and in turn adversely impact generalization. We, therefore, present the need to ensure internal validity when building machine learning models in NLP. Our recommendations also apply to generative large language models, as they are known to be sensitive to even minor semantic preserving alterations. We also propose adapting the idea of *matching* in randomized controlled trials and observational studies to NLP evaluation to measure causation.

Questions posed by information-seeking users often contain implicit false or potentially harmful assumptions. In a high-risk domain such as maternal and infant health, a question-answering system must recognize these pragmatic constraints and go beyond simply answering user questions, examining them in context to respond helpfully. To achieve this, we study assumptions and implications, or pragmatic inferences, made when mothers ask questions about pregnancy and infant care by collecting a dataset of 2,727 inferences from 500 questions across three diverse sources. We study how health experts naturally address these inferences when writing answers, and illustrate that informing existing QA pipelines with pragmatic inferences produces responses that are more complete, mitigating the propagation of harmful beliefs.

Entity Resolution (ER) is an essential task in data integration and its goal is to find records that represent the same entity in a dataset. Deep learning models, especially large pre-trained language models, have achieved state-of-the-art results on this task. A typical ER pipeline consists of Entity Blocking and Entity Matching: Entity Blocking finds candidate record pairs that potentially match and Entity Matching determines if the pairs match. The goal of the entity blocking step is to include as many matching pairs as possible while including as few non-matching pairs as possible. On the other hand, the blocking task can also be considered as an Information Retrieval (IR) task. However, state-of-the-art neural IR models that are based on large language models have not been evaluated on the ER task. What’s more, the generalization ability of state-of-the-art methods for entity blocking is not well-studied but an import aspect in real-world applications. In this work, we evaluate state-of-the-art models for Entity Blocking along with neural IR models on a wide range of real-world datasets, and also study their in-distribution and out-of-distribution generalization abilities.

Opinion summarization is automatically generating summaries from a variety of subjective information, such as product reviews or political opinions. The challenge of opinions summarization lies in presenting divergent or even conflicting opinions. We conduct an analysis of previous summarization models, which reveals their inclination to amplify the polarity bias, emphasizing the majority opinions while ignoring the minority opinions. To address this issue and make the summarizer express both sides of opinions, we introduce the concept of polarity calibration, which aims to align the polarity of output summary with that of input text. Specifically, we develop a reinforcement training approach for polarity calibration. This approach feeds the polarity distance between output summary and input text as reward into the summarizer, and also balance polarity calibration with content preservation and language naturality. We evaluate our Polarity Calibration model (PoCa) on two types of opinions summarization tasks: summarizing product reviews and political opinions articles. Automatic and human evaluation demonstrate that our approach can mitigate the polarity mismatch between output summary and input text, as well as maintain the content semantic and language quality.

Large-scale pre-trained language models have displayed unrivaled capacity in generating text that closely resembles human-written text. Nevertheless, generating texts adhering to specific conditions without fine-tuning or adding new parameters can be challenging. Contemporary approaches commonly rely on either prompts or auxiliary models to avoid modifying the language models. These auxiliary models are designed to assess whether a generated token contributes to meeting the desired requirements. These approaches adjust the distribution of the next token during the inference phase by leveraging the prediction score of the desired attribute to calculate gradients. However, these auxiliary models typically require the language model’s latent states. This prerequisite challenges integrating various existing black box attribute models or tools. We present the Plug-in Language Model (PiLM) as a solution to address the limitations. PiLM leverages reinforcement learning to utilize black box tools directly, adjusting the latent state to control text generation. However, performing backpropagation during the inference phase is time-consuming for PiLM. By replacing backpropagation with a simple regression model, PiLM can achieve an inference time comparable to that of the original LLM. Experiment results show that our approaches in this paper outperform existing state-of-the-art methods that rely on gradient-based, weighted decoding, or prompt-based methodologies.

Recent advancements in integrating external tools with Large Language Models (LLMs) have opened new frontiers, with applications in mathematical reasoning, code generators, and smart assistants. However, existing methods, relying on simple one-time retrieval strategies, fall short on effectively and accurately shortlisting relevant tools. This paper introduces a novel PLUTO (Planning, Learning, and Understanding for TOols) approach, encompassing “Plan-and-Retrieve (P&R)” and “Edit-and-Ground (E&G)” paradigms. The P&R paradigm consists of a neural retrieval module for shortlisting relevant tools and an LLM-based query planner that decomposes complex queries into actionable tasks, enhancing the effectiveness of tool utilization. The E&G paradigm utilizes LLMs to enrich tool descriptions based on user scenarios, bridging the gap between user queries and tool functionalities. Experiment results demonstrate that these paradigms significantly improve the recall and NDCG in tool retrieval tasks, significantly surpassing current state-of-the-art models.