Recent scholarship on reasoning in LLMs has supplied evidence of impressive performance and flexible adaptation to machine generated or human critique. Nonmonotonic reasoning, crucial to human cognition for navigating the real world, remains a challenging, yet understudied task. In this work, we study nonmonotonic reasoning capabilities of seven state-of-the-art LLMs in one abstract and one commonsense reasoning task featuring generics, such as ‘Birds fly’, and exceptions, ‘Penguins don’t fly’ (see Fig. 1). While LLMs exhibit reasoning patterns in accordance with human nonmonotonic reasoning abilities, they fail to maintain stable beliefs on truth conditions of generics at the addition of supporting examples (‘Owls fly’) or unrelated information (‘Lions have manes’).Our findings highlight pitfalls in attributing human reasoning behaviours to LLMs as long as consistent reasoning remains elusive.
论文检索
输入标题、作者或关键词，从 1,915 篇学术成果中精准定位
While human speakers use a variety of different expressions when describing the same object in an image, giving rise to a distribution of plausible labels driven by pragmatic constraints, the extent to which current Vision & Language Large Language Models (VLLMs) can mimic this crucial feature of language use is an open question. This applies to common, everyday objects, but it is particularly interesting for uncommon or novel objects for which a category label may be lacking or fuzzy. Furthermore, similar patterns of variation are observed among human speakers for highly context-sensitive expressions, such as the quantifiers ‘few’ or ‘most’. In our work, we evaluate VLLMs (FROMAGe, BLIP-2, LLaVA) on three categories (nouns, attributes, and quantifiers) where humans show great subjective variability concerning the distribution over plausible labels, using datasets and resources mostly under-explored in previous work. Our results reveal mixed evidence on the ability of VLLMs to capture human naming preferences at generation time: while some models are good at mimicking human distributions for nouns and attributes, all of them fail to assign quantifiers, a task that requires more accurate, high-level reasoning.
The Probabilities Also Matter: A More Faithful Metric for Faithfulness of Free-Text Explanations in Large Language Models
PDF ↗In order to oversee advanced AI systems, it is important to understand their reasons for generating a given output. When prompted, large language models (LLMs) can provide natural language explanations or reasoning traces that sound plausible and receive high ratings from human annotators. However, it is unclear to what extent these explanations are truly capturing the factors responsible for the model’s predictions: the most “human-like” explanation may be different from the one that is most faithful to the model’s true decision making process. In this work, we introduce the correlational counterfactual test (CCT), a faithfulness metric based on counterfactual input edits that takes into account not just the binary label change, but the total shift in the model’s predicted label distribution. We evaluate the faithfulness of free-text explanations generated by few-shot-prompted LLMs from the Llama-2 family on three NLP tasks. We find that these explanations are indeed more likely to mention factors when they are impactful to the model’s prediction, with the degree of association increasing with model size but varying significantly by task.
We analysed a sample of NLP research papers archived in ACL Anthology as an attempt to quantify the degree of openness and the benefit of such an open culture in the NLP community. We observe that papers published in different NLP venues show different patterns related to artefact reuse. We also note that more than 30% of the papers we analysed do not release their artefacts publicly. Further, we observe a wide language-wise disparity in publicly available NLP-related artefacts.
Most existing rationalization approaches are susceptible to degeneration accumulation due to a lack of effective control over the learning direction of the model during training. To address this issue, we propose a novel approach AGR (Agent-Guided Rationalization), guiding the next action of the model based on its current training state. Specifically, we introduce causal intervention calculus to quantify the causal effects inherent during rationale training, and utilize reinforcement learning process to refine the learning bias of them. Furthermore, we pretrain an agent within this reinforced causal environment to guide the next step of the model. We theoretically demonstrate that a good model needs the desired guidance, and empirically show the effectiveness of our approach, outperforming existing state-of-the-art methods on BeerAdvocate and HotelReview datasets.
In this paper, we introduce Neural Information Retrieval resources for 11 widely spoken Indian Languages (Assamese, Bengali, Gujarati, Hindi, Kannada, Malayalam, Marathi, Oriya, Punjabi, Tamil, and Telugu) from two major Indian language families (Indo-Aryan and Dravidian). These resources include (a) INDIC-MARCO, a multilingual version of the MS MARCO dataset in 11 Indian Languages created using Machine Translation, and (b) Indic-ColBERT, a collection of 11 distinct Monolingual Neural Information Retrieval models, each trained on one of the 11 languages in the INDIC-MARCO dataset. To the best of our knowledge, IndicIRSuite is the first attempt at building large-scale Neural Information Retrieval resources for a large number of Indian languages, and we hope that it will help accelerate research in Neural IR for Indian Languages. Experiments demonstrate that Indic-ColBERT achieves 47.47% improvement in the MRR@10 score averaged over the INDIC-MARCO baselines for all 11 Indian languages except Oriya, 12.26% improvement in the NDCG@10 score averaged over the MIRACL Bengali and Hindi Language baselines, and 20% improvement in the MRR@100 Score over the Mr. Tydi Bengali Language baseline.
We introduce a new, extensive multidimensional quality metrics (MQM) annotated dataset covering 11 language pairs in the biomedical domain. We use this dataset to investigate whether machine translation (MT) metrics which are fine-tuned on human-generated MT quality judgements are robust to domain shifts between training and inference. We find that fine-tuned metrics exhibit a substantial performance drop in the unseen domain scenario relative to both metrics that rely on the surface form and pre-trained metrics that are not fine-tuned on MT quality judgments.
In this study, we investigate the capabilities and inherent biases of advanced large language models (LLMs) such as GPT-3.5 and GPT-4 in the context of debate evaluation. We discover that LLM’s performance exceeds humans and surpasses the performance of state-of-the-art methods fine-tuned on extensive datasets. We additionally explore and analyze biases present in LLMs, including positional bias, lexical bias, order bias, which may affect their evaluative judgments. Our findings reveal a consistent bias in both GPT-3.5 and GPT-4 towards the second candidate response presented, attributed to prompt design. We also uncover a lexical bias in both GPT-3.5 and GPT-4, especially when label sets carry connotations such as numerical or sequential, highlighting the critical need for careful label verbalizer selection in prompt design. Additionally, our analysis indicates a tendency of both models to favor the debate’s concluding side as the winner, suggesting an end-of-discussion bias.
We present MaskLID, a simple, yet effective, code-switching (CS) language identification (LID) method. MaskLID does not require any training and is designed to complement current high-performance sentence-level LIDs. Sentence-level LIDs are classifiers trained on monolingual texts to provide single labels, typically using a softmax layer to turn scores into probabilities. However, in cases where a sentence is composed in both L1 and L2 languages, the LID classifier often only returns the dominant label L1. To address this limitation, MaskLID employs a strategy to mask text features associated with L1, allowing the LID to classify the text as L2 in the next round. This method uses the LID itself to identify the features that require masking and does not rely on any external resource. In this work, we explore the use of MaskLID for two open-source LIDs (GlotLID and OpenLID), that are both based on the FastText architecture. Code and demo are available at https://github.com/cisnlp/MaskLID.
For large language models (LLMs) to be effective in the financial domain – where each decision can have a significant impact – it is necessary to investigate realistic tasks and data. Financial professionals often interact with documents spanning hundreds of pages, but most financial research datasets only deal with short excerpts from these documents. To address this, we introduce a long-document financial QA task. We augment 7,437 questions from the existing FinQA dataset with full-document context, extending the average context length from under 700 words in FinQA to 123k words in DocFinQA. We conduct extensive experiments over retrieval-based QA pipelines and long-context language models. Based on our experiments, DocFinQA proves a significant challenge for even state-of-the-art systems. We also provide a case study on a subset of the longest documents in DocFinQA and find that models particularly struggle with these documents. Addressing these challenges may have a wide-reaching impact across applications where specificity and long-range contexts are critical, like gene sequences and legal document contract analysis. DocFinQA dataset is publicly accessible.
Topic modeling is a widely used technique to analyze large document corpora. With the ever-growing emergence of scientific contributions in the field, non-technical users may often use the simplest available software module, independent of whether there are potentially better models available. We present a Simplified Topic Retrieval, Exploration, and Analysis Module (STREAM) for user-friendly topic modelling and especially subsequent interactive topic visualization and analysis. For better topic analysis, we implement multiple intruder-word based topic evaluation metrics. Additionally, we publicize multiple new datasets that can extend the so far very limited number of publicly available benchmark datasets in topic modeling. We integrate downstream interpretable analysis modules to enable users to easily analyse the created topics in downstream tasks together with additional tabular information.The code is available at the following link: https://github.com/AnFreTh/STREAM
State-of-the-art sign language translation (SLT) systems facilitate the learning process through gloss annotations, either in an end2end manner or by involving an intermediate step. Unfortunately, gloss labelled sign language data is usually not available at scale and, when available, gloss annotations widely differ from dataset to dataset. We present a novel approach using sentence embeddings of the target sentences at training time that take the role of glosses. The new kind of supervision does not need any manual annotation but it is learned on raw textual data. As our approach easily facilitates multilinguality, we evaluate it on datasets covering German (PHOENIX-2014T) and American (How2Sign) sign languages and experiment with mono- and multilingual sentence embeddings and translation systems. Our approach significantly outperforms other gloss-free approaches, setting the new state-of-the-art for data sets where glosses are not available and when no additional SLT datasets are used for pretraining, diminishing the gap between gloss-free and gloss-dependent systems.
Born Differently Makes a Difference: Counterfactual Study of Bias in Biography Generation from a Data-to-Text Perspective
PDF ↗How do personal attributes affect biography generation? Addressing this question requires an identical pair of biographies where only the personal attributes of interest are different. However, it is rare in the real world. To address this, we propose a counterfactual methodology from a data-to-text perspective, manipulating the personal attributes of interest while keeping the co-occurring attributes unchanged. We first validate that the fine-tuned Flan-T5 model generates the biographies based on the given attributes. This work expands the analysis of gender-centered bias in text generation. Our results confirm the well-known bias in gender and also show the bias in regions, in both individual and its related co-occurring attributes in semantic machining and sentiment.
Content moderators play a key role in keeping the conversation on social media healthy. While the high volume of content they need to judge represents a bottleneck to the moderation pipeline, no studies have explored how models could support them to make faster decisions. There is, by now, a vast body of research into detecting hate speech, sometimes explicitly motivated by a desire to help improve content moderation, but published research using real content moderators is scarce. In this work we investigate the effect of explanations on the speed of real-world moderators. Our experiments show that while generic explanations do not affect their speed and are often ignored, structured explanations lower moderators’ decision making time by 7.4%.
Do Large Language Models Discriminate in Hiring Decisions on the Basis of Race, Ethnicity, and Gender?
PDF ↗We examine whether large language models (LLMs) exhibit race- and gender-based name discrimination in hiring decisions, similar to classic findings in the social sciences (Bertrand and Mullainathan, 2004). We design a series of templatic prompts to LLMs to write an email to a named job applicant informing them of a hiring decision. By manipulating the applicant’s first name, we measure the effect of perceived race, ethnicity, and gender on the probability that the LLM generates an acceptance or rejection email. We find that the hiring decisions of LLMs in many settings are more likely to favor White applicants over Hispanic applicants. In aggregate, the groups with the highest and lowest acceptance rates respectively are masculine White names and masculine Hispanic names. However, the comparative acceptance rates by group vary under different templatic settings, suggesting that LLMs’ race- and gender-sensitivity may be idiosyncratic and prompt-sensitive.
The Uniform Information Density (UID) hypothesis posits that speakers optimize the communicative properties of their utterances by avoiding spikes in information, thereby maintaining a relatively uniform information profile over time. This paper investigates the impact of UID principles on syntactic reduction, specifically focusing on the optional omission of the connector “that” in English subordinate clauses. Building upon previous research, we extend our investigation to a larger corpus of written English, utilize contemporary large language models (LLMs) and extend the information-uniformity principles by the notion of entropy, to estimate the UID manifestations in the usecase of syntactic reduction choices.
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.
Text-to-SQL, which involves translating natural language into Structured Query Language (SQL), is crucial for enabling broad access to structured databases without expert knowledge. However, designing models for such tasks is challenging due to numerous factors, including the presence of noise, such as ambiguous questions and syntactical errors. This study provides an in-depth analysis of the distribution and types of noise in the widely used BIRD-Bench benchmark and the impact of noise on models. While BIRD-Bench was created to model dirty and noisy database values, it was not created to contain noise and errors in the questions and gold SQL queries. We found that noise in questions and gold queries are prevalent in the dataset, with varying amounts across domains, and with an uneven distribution between noise types. The presence of incorrect gold SQL queries, which then generate incorrect gold answers, has a significant impact on the benchmark’s reliability. Surprisingly, when evaluating models on corrected SQL queries, zero-shot baselines surpassed the performance of state-of-the-art prompting methods. We conclude that informative noise labels and reliable benchmarks are crucial to developing new Text-to-SQL methods that can handle varying types of noise.
EmbSpatial-Bench: Benchmarking Spatial Understanding for Embodied Tasks with Large Vision-Language Models
PDF ↗The recent rapid development of Large Vision-Language Models (LVLMs) has indicated their potential for embodied tasks. However, the critical skill of spatial understanding in embodied environments has not been thoroughly evaluated, leaving the gap between current LVLMs and qualified embodied intelligence unknown. Therefore, we construct EmbSpatial-Bench, a benchmark for evaluating embodied spatial understanding of LVLMs. The benchmark is automatically derived from embodied scenes and covers 6 spatial relationships from an egocentric perspective. Experiments expose the insufficient capacity of current LVLMs (even GPT-4V). We further present EmbSpatial-SFT, an instruction-tuning dataset designed to improve LVLMs’ embodied spatial understanding.
Lay summarisation aims to produce summaries of scientific articles that are comprehensible to non-expert audiences. However, previous work assumes a one-size-fits-all approach, where the content and style of the produced summary are entirely dependent on the data used to train the model. In practice, audiences with different levels of expertise will have specific needs, impacting what content should appear in a lay summary and how it should be presented. Aiming to address this, we propose ATLAS, a novel abstractive summarisation approach that can control various properties that contribute to the overall “layness” of the generated summary using targeted control attributes. We evaluate ATLAS on a combination of biomedical lay summarisation datasets, where it outperforms state-of-the-art baselines using mainstream summarisation metrics.Additional analyses provided on the discriminatory power and emergent influence of our selected controllable attributes further attest to the effectiveness of our approach.