ACL 2024 论文列表

Image retrieval from contextual descriptions (IRCD) aims to identify an image within a set of minimally contrastive candidates based on linguistically complex text. Despite the success of VLMs, they still significantly lag behind human performance in IRCD. The main challenges lie in aligning key contextual cues in two modalities, where these subtle cues are concealed in tiny areas of multiple contrastive images and within the complex linguistics of textual descriptions. This motivates us to propose ContextBLIP, a simple yet effective method that relies on a doubly contextual alignment scheme for challenging IRCD. Specifically, 1) our model comprises a multi-scale adapter, a matching loss, and a text-guided masking loss. The adapter learns to capture fine-grained visual cues. The two losses enable iterative supervision for the adapter, gradually highlighting the focal patches of a single image to the key textual cues. We term such a way as intra-contextual alignment. 2) Then, ContextBLIP further employs an inter-context encoder to learn dependencies among candidates, facilitating alignment between the text to multiple images. We term this step as inter-contextual alignment. Consequently, the nuanced cues concealed in each modality can be effectively aligned. Experiments on two benchmarks show the superiority of our method. We observe that ContextBLIP can yield comparable results with GPT-4V, despite involving about 7,500 times fewer parameters.

Multi-change captioning aims to describe complex and coupled changes within an image pair in natural language. Compared with single-change captioning, this task requires the model to have higher-level cognition ability to reason an arbitrary number of changes. In this paper, we propose a novel context-aware difference distilling (CARD) network to capture all genuine changes for yielding sentences. Given an image pair, CARD first decouples context features that aggregate all similar/dissimilar semantics, termed common/difference context features. Then, the consistency and independence constraints are designed to guarantee the alignment/discrepancy of common/difference context features. Further, the common context features guide the model to mine locally unchanged features, which are subtracted from the pair to distill locally difference features. Next, the difference context features augment the locally difference features to ensure that all changes are distilled. In this way, we obtain an omni-representation of all changes, which is translated into linguistic sentences by a transformer decoder. Extensive experiments on three public datasets show CARD performs favourably against state-of-the-art methods. The code is available at https://github.com/tuyunbin/CARD.

Incomplete utterance rewriting (IUR) aims to reconstruct the utterance with omitted information and pronouns to be standalone and complete based on the context. The existing works predominantly focus on simple ellipsis and coreference problems in brief multi-turn dialogues. But in actual scenarios: 1) the context of the dialogues frequently comprises multiple similar candidates for ellipsis and coreference resolution, pouring to confuse. 2) the number of turns tends to be more extensive, while the content with various topics also grows more complex. This paper proposes a novel method called CaT to address these issues. In particular, we first devise a tacker model, distilled from GPT4-turbo, to adopt Context Tracking that dynamically updates a list of key phrases turn by turn, as accurate candidates for ellipsis and coreference resolution. Second, we further present the Dynamic Context Introduction mechanism to filter irrelevant preceding contexts that are not relied on by any element within the key phrase list to condense extended dialogues. Comprehensive experiments indicate that our solution provides a significant improvement over the existing baselines, and achieves state-of-the-art on three benchmarks.

To answer a question, language models often need to integrate prior knowledge learned during pretraining and new information presented in context. We hypothesize that models perform this integration in a predictable way across different questions and contexts: models will rely more on prior knowledge for questions about entities (e.g., persons, places, etc.) that they are more familiar with due to higher exposure in the training corpus, and be more easily persuaded by some contexts than others. To formalize this problem, we propose two mutual information-based metrics to measure a model’s dependency on a context and on its prior about an entity: first, the persuasion score of a given context represents how much a model depends on the context in its decision, and second, the susceptibility score of a given entity represents how much the model can be swayed away from its original answer distribution about an entity. We empirically test our metrics for their validity and reliability. Finally, we explore and find a relationship between the scores and the model’s expected familiarity with an entity, and provide two use cases to illustrate their benefits.

Simultaneous Machine Translation (SiMT) aims to yield a real-time partial translation with a monotonically growing source-side context.However, there is a counterintuitive phenomenon about the context usage between training and inference: *e.g.*, in wait-k inference, model consistently trained with wait-k is much worse than that model inconsistently trained with wait-k' (k'≠ k) in terms of translation quality. To this end, we first investigate the underlying reasons behind this phenomenon and uncover the following two factors: 1) the limited correlation between translation quality and training loss; 2) exposure bias between training and inference. Based on both reasons, we then propose an effective training approach called context consistency training accordingly, which encourages consistent context usage between training and inference by optimizing translation quality and latency as bi-objectives and exposing the predictions to the model during the training. The experiments on three language pairs demonstrate that our SiMT system encouraging context consistency outperforms existing SiMT systems with context inconsistency for the first time.

Large language models (LLMs) are highly capable at a variety of tasks given the right prompt, but writing one is still a difficult and tedious process. In this work, we introduce ConstitutionalExperts, a method for learning a prompt consisting of constitutional principles (i.e. rules), given a training dataset. Unlike prior methods that optimize the prompt as a single entity, our method incrementally improves the prompt by surgically editing individual principles. We also show that we can improve overall performance by learning unique prompts for different semantic regions of the training data and using a mixture-of-experts (MoE) architecture to route inputs at inference time. We compare our method to other state of the art prompt-optimization techniques across six benchmark datasets. We also investigate whether MoE improves these other techniques. Our results suggest that ConstitutionalExperts outperforms other prompt optimization techniques by 10.9% (F1) and that mixture-of-experts improves all techniques, suggesting its broad applicability.

Efficiently modeling historical information is a critical component in addressing user queries within a conversational question-answering (QA) context, as historical context plays a vital role in clarifying the user’s questions. However, irrelevant history induces noise in the reasoning process, especially for those questions with a considerable historical context. In our novel model-agnostic approach, referred to as **CoTaH** (**Co**nsistency-**T**rained **a**ugmented **H**istory), we augment the historical information with synthetic questions and subsequently employ consistency training to train a model that utilizes both real and augmented historical data to implicitly make the reasoning robust to irrelevant history. To the best of our knowledge, this is the first instance of research using synthetic question generation as a form of data augmentation to model conversational QA settings. By citing a common modeling error prevalent in previous research, we introduce a new baseline and compare our model’s performance against it, demonstrating an improvement in results, particularly in later turns of the conversation, when dealing with questions that include a large historical context.

Recently, Temporal Knowledge Graph Forecasting (TKGF) has emerged as a pivotal domain for forecasting future events. Unlike black-box neural network methods, rule-based approaches are lauded for their efficiency and interpretability. For this line of work, it is crucial to correctly estimate the predictive effectiveness of the rules, i.e., the confidence. However, the existing literature lacks in-depth investigation into how confidence evolves with time. Moreover, inaccurate and heuristic confidence estimation limits the performance of rule-based methods. To alleviate such issues, we propose a framework named TempValid to explicitly model the temporal validity of rules for TKGF. Specifically, we design a time function to model the interaction between temporal information with confidence. TempValid conceptualizes confidence and other coefficients as learnable parameters to avoid inaccurate estimation and combinatorial explosion. Furthermore, we introduce a rule-adversarial negative sampling and a time-aware negative sampling strategies to facilitate TempValid learning. Extensive experiments show that TempValid significantly outperforms previous state-of-the-art (SOTA) rule-based methods on six TKGF datasets. Moreover, it exhibits substantial advancements in cross-domain and resource-constrained rule learning scenarios.

As the use of Large Language Models (LLMs) becomes more widespread, understanding their self-evaluation of confidence in generated responses becomes increasingly important as it is integral to the reliability of the output of these models. We introduce the concept of Confidence-Probability Alignment, that connects an LLM’s internal confidence, quantified by token probabilities, to the confidence conveyed in the model’s response when explicitly asked about its certainty. Using various datasets and prompting techniques that encourage model introspection, we probe the alignment between models’ internal and expressed confidence. These techniques encompass using structured evaluation scales to rate confidence, including answer options when prompting, and eliciting the model’s confidence level for outputs it does not recognize as its own. Notably, among the models analyzed, OpenAI’s GPT-4 showed the strongest confidence-probability alignment, with an average Spearman’s  ̂𝜌 of 0.42, across a wide range of tasks. Our work contributes to the ongoing efforts to facilitate risk assessment in the application of LLMs and to further our understanding of model trustworthiness.

This paper presents a systematic defense of large language model (LLM) hallucinations or ‘confabulations’ as a potential resource instead of a categorically negative pitfall. The standard view is that confabulations are inherently problematic and AI research should eliminate this flaw. In this paper, we argue and empirically demonstrate that measurable semantic characteristics of LLM confabulations mirror a human propensity to utilize increased narrativity as a cognitive resource for sense-making and communication. In other words, it has potential value. Specifically, we analyze popular hallucination benchmarks and reveal that hallucinated outputs display increased levels of narrativity and semantic coherence relative to veridical outputs. This finding reveals a tension in our usually dismissive understandings of confabulation. It suggests, counter-intuitively, that the tendency for LLMs to confabulate may be intimately associated with a positive capacity for coherent narrative-text generation.

Through reading the documentation in the context, tool-using language models can dynamically extend their capability using external tools. The cost is that we have to input lengthy documentation every time the model needs to use the tool, occupying the input window as well as slowing down the decoding process.Given the progress in general-purpose compression, soft context compression is a suitable approach to alleviate the problem. However, when compressing tool documentation, existing methods suffer from the weaknesses of key information loss (specifically, tool/parameter name errors) and difficulty in adjusting the length of compressed sequences based on documentation lengths.To address these problems, we propose two strategies for compressing tool documentation into concise and precise summary sequences for tool-using language models. 1) Selective compression strategy mitigates key information loss by deliberately retaining key information as raw text tokens. 2) Block compression strategy involves dividing tool documentation into short chunks and then employing a fixed-length compression model to achieve variable-length compression. This strategy facilitates the flexible adjustment of the compression ratio.Results on API-Bank and APIBench show that our approach reaches a performance comparable to the upper-bound baseline under up to 16x compression ratio.

This paper introduces ConceptMath, a bilingual (English and Chinese), fine-grained benchmark that evaluates concept-wise mathematical reasoning of Large Language Models (LLMs). Unlike traditional benchmarks that evaluate general mathematical reasoning with an average accuracy, ConceptMath systemically organizes math problems under a hierarchy of math concepts, so that mathematical reasoning can be evaluated at different granularity with concept-wise accuracies. Based on our ConcepthMath, we then evaluate a broad range of LLMs, and we observe existing LLMs, though achieving high average accuracies on traditional benchmarks, exhibit significant performance variations across different math concepts and may even fail catastrophically on the most basic ones. Besides, we also introduce an efficient fine-tuning strategy to enhance the weaknesses of existing LLMs. Finally, we hope ConceptMath could guide the developers to understand the fine-grained mathematical abilities of their models and facilitate the growth of foundation models. Code is available at https://github.com/conceptmath/conceptmath.

Many recent language models (LMs) are capable of in-context learning (ICL), manifested in the LMs’ ability to perform a new task solely from natural-language instruction. Previous work curating in-context learners assumes that ICL emerges from a vast over-parametrization or the scale of multi-task training. However, recent theoretical work attributes the ICL ability to concept-dependent training data and creates functional in-context learners even in small-scale, synthetic settings.In this work, we practically explore this newly identified axis of ICL quality. We propose Concept-aware Training (CoAT), a framework for constructing training scenarios that make it beneficial for the LM to learn to utilize the analogical reasoning concepts from demonstrations. We find that by using CoAT, pre-trained transformers can learn to better utilise new latent concepts from demonstrations and that such ability makes ICL more robust to the functional deficiencies of the previous models. Finally, we show that concept-aware in-context learners are much more effective in in-context learning a majority of unseen tasks compared to traditional instruction tuning, and fare comparably also to previous in-context learners trained in large-scale multitask learning requiring magnitudes of more training data.

Artificial agents that learn to communicate in order to accomplish a given task acquire communication protocols that are typically opaque to a human. A large body of work has attempted to evaluate the emergent communication via various evaluation measures, with **compositionality** featuring as a prominent desired trait. However, current evaluation procedures do not directly expose the compositionality of the emergent communication. We propose a procedure to assess the compositionality of emergent communication by finding the best-match between emerged words and natural language concepts.The best-match algorithm provides both a global score and a translation-map from emergent words to natural language concepts. To the best of our knowledge, it is the first time that such direct and interpretable mapping between emergent words and human concepts is provided.

The task of temporal relation extraction (TRE) involves identifying and extracting temporal relations between events from narratives. We identify two primary issues with TRE systems. First, by formulating TRE as a simple text classification task where every temporal relation is independent, it is hard to enhance the TRE model’s representation of meaning of temporal relations, and its facility with the underlying temporal calculus. We solve the issue by proposing a novel Temporally Contrastive learning model (ConTempo) that increase the model’s awareness of the meaning of temporal relations by leveraging their symmetric or antisymmetric properties. Second, the reusability of innovations has been limited due to incompatibilities in model architectures. Therefore, we propose a unified framework and show that ConTempo is compatible with all three main branches of TRE research. Our results demonstrate that the performance gains of ConTempo are more pronounced, with the total combination achieving state-of-the-art performance on the widely used MATRES and TBD corpora. We furthermore identified and corrected a large number of annotation errors present in the test set of MATRES, after which the performance increase brought by ConTempo becomes more apparent.

In this position paper, we argue that human evaluation of generative large language models (LLMs) should be a multidisciplinary undertaking that draws upon the insights from disciplines such as user experience research and human behavioral psychology to ensure that the experimental design and results are reliable. The conclusions from these evaluations, therefore, must consider factors such as usability, aesthetics and cognitive biases. We highlight how cognitive biases can conflate fluent information and truthfulness, and how cognitive uncertainty affects the reliability of rating scores such as Likert. Furthermore, the evaluation should differentiate the capabilities and weaknesses of increasingly powerful large language models - which requires effective test sets. Scalability of human evaluation is also crucial to wider adoption. Hence, to design an effective human evaluation system in the age of generative NLP we propose the ConSiDERS-The-Human evaluation framework consisting of 6 pillars - Consistency, Scoring Criteria, Differentiating, User Experience, Responsible, and Scalability.

Real-world news comments pose a significant challenge due to their noisy and ambiguous nature, which complicates their modeling for clustering and summarization tasks. Most previous research has predominantly focused on extractive summarization methods within specific constraints. This paper concentrates on Clustering and Abstractive Summarization of online news Comments (CASC). First, we introduce an enhanced fast clustering algorithm that maintains a dynamic similarity threshold to ensure the high density of each comment cluster being built. Moreover, we pioneer the exploration of tuning Large Language Models (LLMs) through a chain-of-thought strategy to generate summaries for each comment cluster. On the other hand, a notable challenge in CASC research is the scarcity of evaluation data. To address this problem, we design an annotation scheme and contribute a manual test suite tailored for CASC. Experimental results on the test suite demonstrate the effectiveness of our improvements to the baseline methods. In addition, the quantitative and qualitative analyses illustrate the adaptability of our approach to real-world news comment scenarios.

Grounded language models use external sources of information, such as knowledge graphs, to meet some of the general challenges associated with pre-training. By extending previous work on compositional generalization in semantic parsing, we allow for a controlled evaluation of the degree to which these models learn and generalize from patterns in knowledge graphs. We develop a procedure for generating natural language questions paired with knowledge graphs that targets different aspects of compositionality and further avoids grounding the language models in information already encoded implicitly in their weights. We evaluate existing methods for combining language models with knowledge graphs and find them to struggle with generalization to sequences of unseen lengths and to novel combinations of seen base components. While our experimental results provide some insight into the expressive power of these models, we hope our work and released datasets motivate future research on how to better combine language models with structured knowledge representations.

Event commonsense reasoning requires the ability to reason about the relationship between events, as well as infer implicit contextunderlying that relationship. However, data scarcity makes it challenging for language models to learn to generate commonsense infer-ences for contexts and questions involving interactions between complex events. To address this demand, we present COM2 (COMplexCOMmonsense), a new dataset created by sampling multi-hop logical queries (e.g., the joint effect or cause of both event A and B, or theeffect of the effect of event C) from an existing commonsense knowledge graph (CSKG), and verbalizing them using handcrafted rules andlarge language models into multiple-choice and text generation questions. Our experiments show that language models trained on COM2 exhibit significant improve ments in complex reasoning ability, resulting in enhanced zero-shot performance in both in-domain and out-of-domain tasks for question answering and generative commonsense reasoning, without expensive human annotations