Causal reasoning is a cornerstone of how humans interpret the world. To model and reason about causality, causal graphs offer a concise yet effective solution. Given the impressive advancements in language models, a crucial question arises: can they really understand causal graphs? To this end, we pioneer an investigation into language models’ understanding of causal graphs. Specifically, we develop a framework to define causal graph understanding, by assessing language models’ behaviors through four practical criteria derived from diverse disciplines (e.g., philosophy and psychology). We then develop CLEAR, a novel benchmark that defines three complexity levels and encompasses 20 causal graph-based tasks across these levels. Finally, based on our framework and benchmark, we conduct extensive experiments on six leading language models and summarize five empirical findings. Our results indicate that while language models demonstrate a preliminary understanding of causal graphs, significant potential for improvement remains.
论文检索
输入标题、作者或关键词，从 2,271 篇学术成果中精准定位
Learning from Relevant Subgoals in Successful Dialogs using Iterative Training for Task-oriented Dialog Systems
PDF ↗Task-oriented Dialog (ToD) systems have to solve multiple subgoals to accomplish user goals, whereas feedback is often obtained only at the end of the dialog. In this work, we propose SUIT (SUbgoal-aware ITerative Training), an iterative training approach for improving ToD systems. We sample dialogs from the model we aim to improve and determine subgoals that contribute to dialog success using distant supervision to obtain high quality training samples. We show how this data improves supervised fine-tuning or, alternatively, preference learning results. Performance improves when applying these steps over several iterations: SUIT reaches new state-of-the-art performance on a popular ToD benchmark.
Mixture of experts (MoE) has become the standard for constructing production-level large language models (LLMs) due to its promise to boost model capacity without causing significant overheads. Nevertheless, existing MoE methods usually enforce a constant top-k routing for all tokens, which is arguably restrictive because various tokens (e.g., "<EOS>” vs. “apple”) may require various numbers of experts for feature abstraction. Lifting such a constraint can help make the most of limited resources and unleash the potential of the model for downstream tasks. In this sense, we introduce **AdaMoE** to realize token-adaptive routing for MoE, where different tokens are permitted to select a various number of experts. AdaMoE makes minimal modifications to the vanilla MoE with top-k routing—it simply introduces a fixed number of *null experts*, which do not consume any FLOPs, to the expert set and increases the value of k. AdaMoE does not force each token to occupy a fixed number of null experts but ensures the average usage of the null experts with a load-balancing loss, leading to an adaptive number of null/true experts used by each token. AdaMoE exhibits a strong resemblance to MoEs with expert choice routing while allowing for trivial auto-regressive modeling. AdaMoE is easy to implement and can be effectively applied to pre-trained (MoE-)LLMs. Extensive studies show that AdaMoE can reduce average expert load (FLOPs) while achieving superior performance. For example, on the ARC-C dataset, applying our method to fine-tuning Mixtral-8x7B can reduce FLOPs by 14.5% while increasing accuracy by 1.69%.Code is available at [this link](https://github.com/CengZihao/AdaMoE).
Large Language Models (LLMs) demonstrate ever-increasing abilities in mathematical and algorithmic tasks, yet their geometric reasoning skills are underexplored. We investigate LLMs’ abilities in constructive geometric problem-solving, – one of the most fundamental steps in developing human mathematical reasoning, revealing notable challenges in this domain. LLMs exhibit biases in variable names, struggle with 2D spatial relationships and planning, and hallucinate object placements. To this end, we introduce a framework that enhances LLMs’ reasoning potential through a multi-agent system conducting internal dialogue. This work underscores LLMs’ limitations in geometric reasoning and improves their capabilities through self-correction, collaboration, and diverse role specializations.
FANTAstic SEquences and Where to Find Them: Faithful and Efficient API Call Generation through State-tracked Constrained Decoding and Reranking
PDF ↗API call generation is the cornerstone of large language models’ tool-using ability that provides access to the larger world. However, existing supervised and in-context learning approaches suffer from high training costs, poor data efficiency, and generated API calls that can be unfaithful to the API documentation and the user’s request. To address these limitations, we propose an output-side optimization approach called FANTASE. Two of the unique contributions of FANTASE are its State-Tracked Constrained Decoding (SCD) and Reranking components. SCD dynamically incorporates appropriate API constraints in the form of Token Search Trie for efficient and guaranteed generation faithfulness with respect to the API documentation. The Reranking component efficiently brings in the supervised signal by leveraging a lightweight model as the discriminator to rerank the beam-searched candidate generations of the large language model. We demonstrate the superior performance of FANTASE in API call generation accuracy, inference efficiency, and context efficiency with DSTC8 and API Bank datasets.
SCA: Selective Compression Attention for Efficiently Extending the Context Window of Large Language Models
PDF ↗Large language models (LLMs) have achieved impressive performance across various domains, but the limited context window and the expensive computational cost of processing long texts restrict their more comprehensive application. In this paper, we propose Selective Compression Attention (SCA), a general and effective method to expand the context window and reduce memory footprint by compressing the KV cache of LLMs. Specifically, through preliminary experiments, we found that the KV cache contains many similar vectors, resulting in information redundancy, which can be compressed by retaining representative vectors and discarding others. Therefore, SCA continuously selects the most distinctive vectors to keep through a greedy algorithm, reducing information loss during compression. Extensive experiments on various tasks verify the effectiveness of our method. Compared with existing methods, SCA can significantly reduce the impact on model performance under the same compression ratio. Furthermore, the context window of LLMs can be efficiently expanded using SCA without any training, which can even achieve better performance than specially fine-tuned long context models.
A Deep Analysis of the Impact of Multiword Expressions and Named Entities on Chinese-English Machine Translations
PDF ↗In this paper, we present a study on the impact of so-called multiword expressions (MWEs) and multiword named entities (NEs) on the performance of Chinese-English machine translation (MT) systems. Built on an extended version of the data from the WMT22 Metrics Shared Task (with extra labels of 9 types of Chinese MWEs, and 19 types of Chinese multiword NEs) which includes scores and error annotations provided by human experts, we make further extraction of MWE- and NE-related translation errors. By investigating the human evaluation scores and the error rates on each category of MWEs and NEs, we find that: 1) MT systems tend to perform significantly worse on Chinese sentences with most kinds of MWEs and NEs; 2) MWEs and NEs which make up of about twenty percent of tokens, i.e. characters in Chinese, result in one-third of translation errors; 3) for 13 categories of MWEs and NEs, the error rates exceed 50% with the highest to be 84.8%. Based on the results, we emphasize that MWEs and NEs are still a bottleneck issue for MT and special attention to MWEs and NEs should be paid to further improving the performance of MT systems.
Code-switching in speech, particularly between languages that use different scripts, can potentially be correctly transcribed in various forms, including different ways of transliteration of the embedded language into the matrix language script. Traditional methods for measuring accuracy, such as Word Error Rate (WER), are too strict to address this challenge. In this paper, we introduce PolyWER, a proposed framework for evaluating speech recognition systems to handle language-mixing. PolyWER accepts transcriptions of code-mixed segments in different forms, including transliterations and translations. We demonstrate the algorithms use cases through detailed examples, and evaluate it against human judgement. To enable the use of this metric, we appended the annotations of a publicly available Arabic-English code-switched dataset with transliterations and translations of code-mixed speech. We also utilize these additional annotations for fine-tuning ASR models and compare their performance using PolyWER. In addition to our main finding on PolyWER’s effectiveness, our experiments show that alternative annotations could be more effective for fine-tuning monolingual ASR models.
In the digital era, table structure recognition technology is a critical tool for processing and analyzing large volumes of tabular data. Previous methods primarily focus on visual aspects of table structure recovery but often fail to effectively comprehend the textual semantics within tables, particularly for descriptive textual cells. In this paper, we introduce UniTabNet, a novel framework for table structure parsing based on the image-to-text model. UniTabNet employs a “divide-and-conquer” strategy, utilizing an image-to-text model to decouple table cells and integrating both physical and logical decoders to reconstruct the complete table structure. We further enhance our framework with the Vision Guider, which directs the model’s focus towards pertinent areas, thereby boosting prediction accuracy. Additionally, we introduce the Language Guider to refine the model’s capability to understand textual semantics in table images. Evaluated on prominent table structure datasets such as PubTabNet, PubTables1M, WTW, and iFLYTAB, UniTabNet achieves a new state-of-the-art performance, demonstrating the efficacy of our approach. The code will also be made publicly available.
Toeing the Party Line: Election Manifestos as a Key to Understand Political Discourse on Twitter
PDF ↗Political discourse on Twitter is a moving target: politicians continuously make statements about their positions. It is therefore crucial to track their discourse on social media to understand their ideological positions and goals. However, Twitter data is also challenging to work with since it is ambiguous and often dependent on social context, and consequently, recent work on political positioning has tended to focus strongly on manifestos (parties’ electoral programs) rather than social media.In this paper, we extend recently proposed methods to predict pairwise positional similarities between parties from the manifesto case to the Twitter case, using hashtags as a signal to fine-tune text representations, without the need for manual annotation. We verify the efficacy of fine-tuning and conduct a series of experiments that assess the robustness of our method for low-resource scenarios. We find that our method yields stable positionings reflective of manifesto positionings, both in scenarios with all tweets of candidates across years available and when only smaller subsets from shorter time periods are available. This indicates that it is possible to reliably analyze the relative positioning of actors without the need for manual annotation, even in the noisier context of social media.
In the pursuit of developing Large Language Models (LLMs) that adhere to societal standards, it is imperative to detect the toxicity in the generated text. The majority of existing toxicity metrics rely on encoder models trained on specific toxicity datasets, which are susceptible to out-of-distribution (OOD) problems and depend on the dataset’s definition of toxicity. In this paper, we introduce a robust metric grounded on LLMs to flexibly measure toxicity according to the given definition. We first analyze the toxicity factors, followed by an examination of the intrinsic toxic attributes of LLMs to ascertain their suitability as evaluators. Finally, we evaluate the performance of our metric with detailed analysis. Our empirical results demonstrate outstanding performance in measuring toxicity within verified factors, improving on conventional metrics by 12 points in the F1 score. Our findings also indicate that upstream toxicity significantly influences downstream metrics, suggesting that LLMs are unsuitable for toxicity evaluations within unverified factors.
Cognitive dynamics, which refer to the evolution in human cognitive processes, are pivotal to advance human understanding of the world. Recent advancements in large language models (LLMs) highlight their potential for cognitive simulation. However, these LLM-based cognitive studies primarily focus on replicating human cognition in specific contexts, overlooking the inherently dynamic nature of cognition. To bridge this gap, we explore the cognitive dynamics of LLMs and present a corresponding task inspired by longitudinal studies. Toward the task, we develop CogBench, a novel benchmark to assess the cognitive dynamics of LLMs and validate it through participant surveys. We also design two evaluation metrics for CogBench, including Authenticity and Rationality. Recognizing the inherent static nature of LLMs, we further introduce CogGPT for the task, which features an innovative iterative cognitive mechanism to develop lifelong cognitive dynamics. Empirical results demonstrate the superiority of CogGPT over several existing methods, particularly in its ability to facilitate role-specific cognitive dynamics under continuous information flows. We will release the code and data to enable further research.
Machine translation (MT) in the medical domain plays a pivotal role in enhancing healthcare quality and disseminating medical knowledge. Despite advancements in English-Thai MT technology, common MT approaches often underperform in the medical field due to their inability to precisely translate medical terminologies. Our research prioritizes not merely improving translation accuracy but also maintaining medical terminology in English within the translated text through code-switched (CS) translation. We developed a method to produce CS medical translation data, fine-tuned a CS translation model with this data, and evaluated its performance against strong baselines, such as Google Neural Machine Translation (NMT) and GPT-3.5/GPT-4. Our model demonstrated competitive performance in automatic metrics and was highly favored in human preference evaluations. Our evaluation result also shows that medical professionals significantly prefer CS translations that maintain critical English terms accurately, even if it slightly compromises fluency. Our code and test set are publicly available https://github.com/preceptorai-org/NLLB_CS_EM_NLP2024.
Distilling Instruction-following Abilities of Large Language Models with Task-aware Curriculum Planning
PDF ↗Instruction tuning aims to align large language models (LLMs) with open-domain instructions and human-preferred responses. While several studies have explored autonomous approaches to distilling and annotating instructions from powerful proprietary LLMs, such as ChatGPT, they often neglect the impact of the distributions and characteristics of tasks, together with the varying difficulty of instructions in training sets. This oversight can lead to imbalanced knowledge capabilities and poor generalization powers of student LLMs. To address these challenges, we introduce Task-Aware Curriculum Planning for Instruction Refinement (TAPIR), a multi-round distillation framework that utilizes an oracle LLM to select instructions that are difficult for a student LLM to follow. To balance the student’s capabilities, task distributions in training sets are adjusted with responses automatically refined according to their corresponding tasks. In addition, by incorporating curriculum planning, our approach systematically escalates the difficulty levels of tasks, progressively enhancing the student LLM’s capabilities. We rigorously evaluate TAPIR using several widely recognized benchmarks (such as AlpacaEval 2.0, MT-Bench, etc.) and multiple student LLMs. Empirical results demonstrate that student LLMs, trained with our method and less training data, outperform larger instruction-tuned models and strong distillation baselines.
Existing approaches to relation extraction obtain relation embeddings by concatenating embeddings of the head and tail entities. Despite the popularity of this approach, we find that such representations mostly capture the types of the entities involved, leading to false positives and confusion between relations that involve entities of the same type. Another possibility is to use a prompt with a [MASK] token to directly learn relation embeddings, but this approach tends to perform poorly. We show that this underperformance comes from the fact that information about entity types is insufficiently captured by the [MASK] embeddings. We therefore propose a simple model, which combines such [MASK] embeddings with entity embeddings. Despite its simplicity, our model consistently outperforms the state-of-the-art across several benchmarks, even when the entity embeddings are obtained from a pre-trained entity typing model. We also experiment with a self-supervised pre-training strategy which further improves the results.
As a manner to augment pretrained large language models (LLM), knowledge injection is critical to develop vertical domain large models and has been widely studied. While most current approaches, including parameter-efficient fine-tuning (PEFT) and block expansion methods, uniformly apply knowledge across all LLM layers, it raises the question: are all layers equally crucial for knowledge injection? We embark upon evaluating the importance of each layer to locate the optimal layer range for knowledge injection. Intuitively, more important layers should play more critical roles in knowledge injection and deserve denser injection. We observe performance dips in question-answering benchmarks after the removal or expansion of the shallow layers, and the degradation shrinks as the layer gets deeper, indicating that the shallow layers hold the key to knowledge injection. This insight leads us to propose the S strategy, a post-pretraining strategy of selectively enhancing shallow layers while pruning the less effective deep ones. Based on this strategy, we introduce Llama Slayer 8B. We experimented on the corpus of code & math and demonstrated the effectiveness of our strategy. Further experiments across different LLM, Mistral-7B, and a legal corpus confirmed the approach’s general applicability, underscoring its wide-ranging efficacy.
SecureSQL: Evaluating Data Leakage of Large Language Models as Natural Language Interfaces to Databases
PDF ↗With the widespread application of Large Language Models (LLMs) in Natural Language Interfaces to Databases (NLIDBs), concerns about security issues in NLIDBs have been increasing gradually. However, research on sensitive data leakage in NLIDBs is relatively limited. Therefore, we propose a benchmark to assess the potential of language models to leak sensitive data when generating SQL queries. This benchmark covers 932 samples from 34 different domains, including medical, legal, financial, and political aspects. We evaluate 15 models from six LLM families, and the results show that the model with the best performance has an accuracy of 61.7%, whereas humans achieve an accuracy of 94%. Most models perform close to or even below the level of random selection. We also evaluate two common attack methods, namely prompt injection and inference attacks, as well as a defense method based on chain-of-thoughts (COT) prompting. Experimental results show that both attack methods significantly impact the model, while the defense method based on COT prompting dose not significantly improve accuracy, further highlighting the severity of sensitive data leakage issues in NLIDBs. We hope this research will draw more attention and further study from the researchers on this issue.
The recently proposed ToolkenGPT tool learning paradigm demonstrates promising performance but suffers from two major issues: first, it cannot benefit from tool documentation, and second, it often makes mistakes in whether to use a tool at all. We introduce Toolken+ that mitigates the first problem by reranking top-k tools selected by ToolkenGPT and the second problem with a special REJECT option such that the model will generate a vocabulary token if REJECT is ranked first. We demonstrate the effectiveness of Toolken+ on multistep numerical reasoning and tool selection tasks.
Giving Control Back to Models: Enabling Offensive Language Detection Models to Autonomously Identify and Mitigate Biases
PDF ↗The rapid development of social media has led to an increase in online harassment and offensive speech, posing significant challenges for effective content moderation. Existing automated detection models often exhibit a bias towards predicting offensive speech based on specific vocabulary, which not only compromises model fairness but also potentially exacerbates biases against vulnerable and minority groups. Addressing these issues, this paper proposes a bias self-awareness and data self-iteration framework for mitigating model biases. This framework aims to “giving control back to models: enabling offensive language detection models to autonomously identify and mitigate biases” through bias self-awareness algorithms and self-iterative data augmentation method. Experimental results demonstrate that the proposed framework effectively reduces the false positive rate of models in both in-distribution and out-of-distribution tests, enhances model accuracy and fairness, and shows promising performance improvements in detecting offensive speech on larger-scale datasets.