ACL 2024 论文列表

Stochastic sampling strategies such as top-k and top-p have been widely used in dialogue generation task. However, as an open-domain chatting system, there will be two different conversation scenarios, i.e. chit-chat and knowledge-based question answering. In the former situation, responses diversity is essential due to the one-to-many nature in dialogue. The latter, on the other hand, requires less randomness given that stochastic decoding strategy entails the risk of generating incorrect information. As a result, an adaptive and flexible decoding strategy is needed to cope with these two scenarios simultaneously. To this end, we propose the dynamic decoding strategy (DDS), which can adjust the decoding space w.r.t. different contexts. In DDS, both sequence-level and token-level adaptive search can be achieved to adjust the decoding process in a unified framework. Besides, our adaptive algorithm can not only be used during model inference, but it can also be applied during the model training stage to further enhance the performance. Comprehensive experiments indicate that the proposed decoding strategy can consistently improve the performance of pre-trained dialogue models when coupled with four well-used stochastic decoding algorithms.

We consider two popular approaches to KnowledgeGraph Completion (KGC): textual modelsthat rely on textual entity descriptions, andstructure-based models that exploit the connectivitystructure of the Knowledge Graph(KG). Preliminary experiments show that theseapproaches have complementary strengths:structure-based models perform exceptionallywell when the gold answer is easily reachablefrom the query head in the KG, while textualmodels exploit descriptions to give goodperformance even when the gold answer isnot easily reachable. In response, we proposeDynaSemble, a novel method for learningquery-dependent ensemble weights to combinethese approaches by using the distributions ofscores assigned by the models in the ensembleto all candidate entities. DynaSemble achievesstate-of-the-art results on three standard KGCdatasets, with up to 6.8 pt MRR and 8.3 ptHits@1 gains over the best baseline model forthe WN18RR dataset.

This study investigates the existence of positional biases in Transformer-based language models for text representation learning, particularly in the context of web document retrieval. We build on previous research that demonstrated loss of information in the middle of input sequences for causal language models, extending it to the domain of embedding learning. We examine positional biases at multiple stages of the training pipeline for an encoder-decoder neural retrieval model, namely language model pre-training, contrastive pre-training, and contrastive fine-tuning. Experiments with the MS-MARCO document collection reveal that after contrastive pre-training the model already generates embeddings that better capture the beginning of the input content, with fine-tuning further aggravating this effect.

As large language models (LLM) are increasingly used for text generation tasks, it is critical to audit their usages, govern their applications, and mitigate their potential harms. Existing watermark techniques are shown effective in embedding single human-imperceptible and machine-detectable patterns without significantly affecting generated text quality and semantics. However, the efficiency in detecting watermarks, i.e., the minimum number of tokens required to assert detection with significance and robustness against post-editing, is still debatable. In this paper, we propose, Duwak, to fundamentally enhance the efficiency and quality of watermarking by embedding dual secret patterns in both token probability distribution and sampling schemes. To mitigate expression degradation caused by biasing toward certain tokens, we design a contrastive search to watermark the sampling scheme, which minimizes the token repetition and enhances the diversity. We theoretically explain the interdependency of the two watermarks within Duwak. We evaluate Duwak extensively on Llama2 and Vicuna under various post-editing attacks, against four state-of-the-art watermarking techniques and combinations of them. Our results show that Duwak marked text achieves the highest watermarked text quality at the lowest required token count for detection, up to 70% tokens less than existing approaches, especially under post paraphrasing.

Syntactically Controlled Paraphrase Generation (SCPG), which aims at generating sentences having syntactic structures resembling given exemplars, is attracting more research efforts in recent years. We took an empirical survey on previous SCPG datasets and methods and found three tacitly approved while seldom mentioned intrinsic shortcomings/trade-offs in terms of data obtaining, task formulation, and pre-training strategies. As a mitigation to these shortcomings, we proposed a novel Dual-Stage Multi-Task (DSMT) pre-training scheme, involving a series of structure-oriented and syntax-oriented tasks, which, in our opinion, gives sequential text models the ability of com-prehending intrinsically non-sequential structures like Linearized Constituency Trees (LCTs), understanding the underlying syntactics, and even generating them by parsing sentences. We performed further pre-training of the popular T5 model on these novel tasks and fine-tuned the trained model on every possible variant of SCPG task in literature, finding that our models significantly outperformed (up to 10+ BLEU-4) previous state-of-the-art methods. Finally, we carried out ablation studies which demonstrated the effectiveness of our DSMT methods and emphasized on the SCPG performance gains compared to vanilla T5 models, especially on hard samples or under few-shot settings.

Hierarchical text classification aims at categorizing texts into a multi-tiered tree-structured hierarchy of labels. Existing methods pay more attention to capture hierarchy-aware text feature by exploiting explicit parent-child relationships, while interactions between peer labels are rarely taken into account, resulting in severe label confusion within each layer. In this work, we propose a novel Dual Prompt Tuning (DPT) method, which emphasizes identifying discrimination among peer labels by performing contrastive learning on each hierarchical layer. We design an innovative hand-crafted prompt containing slots for both positive and negative label predictions to cooperate with contrastive learning. In addition, we introduce a label hierarchy self-sensing auxiliary task to ensure cross-layer label consistency. Extensive experiments demonstrate that DPT achieves significant improvements and outperforms the current state-of-the-art methods on BGC and RCV1-V2 benchmark datasets.

Drug safety research is crucial for maintaining public health, often requiring comprehensive data support. However, the resources currently available to the public are limited and fail to provide a comprehensive understanding of the relationship between drugs and their side effects. This paper introduces “DrugWatch”, an easy-to-use and interactive multi-source information visualisation platform for drug safety study. It allows users to understand common side effects of drugs and their statistical information, flexibly retrieve relevant medical reports, or annotate their own medical texts with our automated annotation tool. Supported by NLP technology and enriched with interactive visual components, we are committed to providing researchers and practitioners with a one-stop information analysis, retrieval, and annotation service. The demonstration video is available at https://www.youtube.com/watch?v=RTqDgxzETjw. We also deployed an online demonstration system at https://drugwatch.net/.

We present a novel inference scheme, self-speculative decoding, for accelerating Large Language Models (LLMs) without the need for an auxiliary model. This approach is characterized by a two-stage process: drafting and verification. The drafting stage generates draft tokens at a slightly lower quality but more quickly, which is achieved by selectively skipping certain intermediate layers during drafting. Subsequently, the verification stage employs the original LLM to validate those draft output tokens in one forward pass. This process ensures the final output remains identical to that produced by the unaltered LLM. Moreover, the proposed method requires no additional neural network training and no extra memory footprint, making it a plug-and-play and cost-effective solution for inference acceleration. Benchmarks with LLaMA-2 and its variants demonstrated a speedup up to 1.99×.

Teachers are important to imparting knowledge and guiding learners, and the role of large language models (LLMs) as potential educators is emerging as an important area of study. Recognizing LLMs’ capability to generate educational content can lead to advances in automated and personalized learning. While LLMs have been tested for their comprehension and problem-solving skills, their capability in teaching remains largely unexplored.In teaching, questioning is a key skill that guides students to analyze, evaluate, and synthesize core concepts and principles.Therefore, our research introduces a benchmark to evaluate the questioning capability in education as a teacher of LLMs through evaluating their generated educational questions, utilizing Anderson and Krathwohl’s taxonomy across general, monodisciplinary, and interdisciplinary domains. We shift the focus from LLMs as learners to LLMs as educators, assessing their teaching capability through guiding them to generate questions. We apply four metrics, including relevance, coverage, representativeness, and consistency, to evaluate the educational quality of LLMs’ outputs. Our results indicate that GPT-4 demonstrates significant potential in teaching general, humanities, and science courses; Claude2 appears more apt as an interdisciplinary teacher. Furthermore, the automatic scores align with human perspectives.

Neural machine translation systems estimate probabilities of target sentences given source sentences, yet these estimates may not align with human preferences. This work introduces QE-fusion, a method that synthesizes translations using a quality estimation metric (QE), which correlates better with human judgments. QE-fusion leverages a pool of candidates sampled from a model, combining spans from different candidates using a QE metric such as CometKiwi. We compare QE-fusion against beam search and recent reranking techniques, such as Minimum Bayes Risk decoding or QE-reranking. Our method consistently improves translation quality in terms of COMET and BLEURT scores when applied to large language models (LLMs) used for translation (PolyLM, XGLM, Llama2, Mistral, ALMA, and Tower) and to multilingual translation models (NLLB), over five language pairs. Notably, QE-fusion exhibits larger improvements for LLMs due to their ability to generate diverse outputs. We demonstrate that our approach generates novel translations in over half of the cases and consistently outperforms other methods across varying numbers of candidates (5–200). Furthermore, we empirically establish that QE-fusion scales linearly with the number of candidates in the pool.

Despite efforts to expand the knowledge of large language models (LLMs), knowledge gaps—missing or outdated information in LLMs—might always persist given the evolving nature of knowledge. In this work, we study approaches to identify LLM knowledge gaps and abstain from answering questions when knowledge gaps are present. We first adapt existing approaches to model calibration or adaptation through fine-tuning/prompting and analyze their ability to abstain from generating low-confidence outputs. Motivated by their failures in self-reflection and over-reliance on held-out sets, we propose two novel approaches that are based on model collaboration, i.e., LLMs probing other LLMs for knowledge gaps, either cooperatively or competitively. Extensive experiments with three LLMs on four QA tasks featuring diverse knowledge domains demonstrate that both cooperative and competitive approaches to unveiling LLM knowledge gaps achieve up to 19.3% improvements on abstain accuracy against the strongest baseline. Further analysis reveals that our abstention methods pinpoint failure cases in retrieval augmentation and knowledge gaps in multi-hop reasoning.

The fairness and trustworthiness of Large Language Models (LLMs) are receiving increasing attention. Implicit hate speech, which employs indirect language to convey hateful intentions, occupies a significant portion of practice. However, the extent to which LLMs effectively address this issue remains insufficiently examined. This paper delves into the capability of LLMs to detect implicit hate speech and express confidence in their responses. Our evaluation meticulously considers various prompt patterns and mainstream uncertainty estimation methods. Our findings highlight that LLMs exhibit two extremes: (1) LLMs display excessive sensitivity towards groups or topics that may cause fairness issues, resulting in misclassifying benign statements as hate speech. (2) LLMs’ confidence scores for each method excessively concentrate on a fixed range, remaining unchanged regardless of the dataset’s complexity. Consequently, the calibration performance is heavily reliant on primary classification accuracy. These discoveries unveil new limitations of LLMs, underscoring the need for caution when optimizing models to ensure they do not veer towards extremes. This serves as a reminder to carefully consider sensitivity and confidence in the pursuit of model fairness.

Image-based advertisements are complex multimodal stimuli that often contain unusual visual elements and figurative language. Previous research on automatic ad understanding has reported impressive zero-shot accuracy of contrastive vision-and-language models (VLMs) on an ad-explanation retrieval task. Here, we examine the original task setup and show that contrastive VLMs can solve it by exploiting grounding heuristics. To control for this confound, we introduce TRADE, a new evaluation test set with adversarial grounded explanations. While these explanations look implausible to humans, we show that they “fool” four different contrastive VLMs. Our findings highlight the need for an improved operationalisation of automatic ad understanding that truly evaluates VLMs’ multimodal reasoning abilities. We make our code and TRADE available at https://github.com/dmg-illc/trade.

Research on abusive language detection and content moderation is crucial to combat online harm. However, current limitations set by regulatory bodies and social media platforms can make it difficult to share collected data. We address this challenge by exploring the possibility to replace existing datasets in English for abusive language detection with synthetic data obtained by rewriting original texts with an instruction-based generative model.We show that such data can be effectively used to train a classifier whose performance is in line, and sometimes better, than a classifier trained on original data. Training with synthetic data also seems to improve robustness in a cross-dataset setting. A manual inspection of the generated data confirms that rewriting makes it impossible to retrieve the original texts online.

kNN-MT has utilized neighborhood knowledge for auxiliary decoding, significantly improving translation performance. Subsequently, kNN-KD transitions the use of neighborhood knowledge from the decoding phase to the training phase, to address the temporal and spatial inefficiencies inherent in kNN-MT. However, kNN-KD transfers all the kNN knowledge arbitrarily, which has the potential to restrict the learning of student models. In this paper, we propose a novel domain-aware kNN-KD method, which filters out domain-relevant neighborhood knowledge for learning in the distillation process. Notably, this entire process exclusively utilizes the neighborhood knowledge of the original model, eliminating the need for establishing any additional datastores. Experiments on four domain translation tasks demonstrate that our method achieves state-of-the-art performance, realizing an average gain of 1.55 COMET and 1.42 BLEU scores, by further enhancing the translation of rare words. Source code can be accessed at https://github.com/wangzx1219/Dk-KD.

This paper focuses on answering subjective questions about products. Different from the factoid question with a single answer span, this subjective one involves multiple viewpoints. For example, the question of ‘how the phone’s battery is?’ not only involves facts of battery capacity but also contains users’ opinions on the battery’s pros and cons. A good answer should be able to integrate these heterogeneous and even inconsistent viewpoints, which is formalized as a subjective induction QA task. For this task, the data distributions are often imbalanced across different product domains. It is hard for traditional methods to work well without considering the shift of domain patterns. To address this problem, we propose a novel domain-adaptive model. Concretely, for each sample in the source and target domain, we first retrieve answer-related knowledge and represent them independently. To facilitate knowledge transferring, we then disentangle the representations into domain-invariant and domain-specific latent factors. Moreover, we develop an adversarial discriminator with contrastive learning to reduce the impact of out-of-domain bias. Based on learned latent vectors in a target domain, we yield multi-perspective summaries as inductive answers. Experiments on popular datasets show the effectiveness of our method.

Code Large Language Models (Code LLMs) have demonstrated outstanding performance in code-related tasks. Various instruction finetuning approaches have been proposed to boost the code generation performance of pre-trained Code LLMs. In this paper, we introduce a diverse instruction model DolphCoder with self-evaluating for code generation. It learns diverse instruction targets and combines a code evaluation objective to enhance its code generation ability. Our model achieves superior performance on the HumanEval and MBPP benchmarks, demonstrating new insights for future code instruction tuning work. Our key findings are: (1) Augmenting more diverse responses with more distinct reasoning paths increases the code capability of LLMs. (2) Improving one’s ability to evaluate the correctness of code also enhances their ability to create it.

Information about pretraining corpora used to train the current best-performing language models is seldom discussed: commercial models rarely detail their data, and even open models are often released without accompanying training data or recipes to reproduce them. As a result, it is challenging to conduct and advance scientific research on language modeling, such as understanding how training data impacts model capabilities and limitations. To facilitate scientific research on language model pretraining, we curate and release Dolma, a three-trillion-token English corpus, built from a diverse mixture of web content, scientific papers, code, public-domain books, social media, and encyclopedic materials. We extensively document Dolma, including its design principles, details about its construction, and a summary of its contents. We present analyses and experimental results on intermediate states of Dolma to share what we have learned about important data curation practices. Finally, we open-source our data curation toolkit to enable reproduction of our work as well as support further research in large-scale data curation.

The burgeoning generative capabilities of large language models (LLMs) have raised growing concerns about abuse, demanding automatic machine-generated text detectors. DetectGPT, a zero-shot metric-based detector, first introduces perturbation and shows great performance improvement. However, in DetectGPT, the random perturbation strategy could introduce noise, and logit regression depends on the threshold, harming the generalizability and applicability of individual or small-batch inputs. Hence, we propose a novel fine-tuned detector, PECOLA, bridging metric-based and fine-tuned methods by contrastive learning on selective perturbation. Selective strategy retains important tokens during perturbation and weights for multi-pair contrastive learning. The experiments show that PECOLA outperforms the state-of-the-art (SOTA) by 1.20% in accuracy on average on four public datasets. And we further analyze the effectiveness, robustness, and generalization of the method.