Many recent language model (LM) interpretability studies have adopted the circuits framework, which aims to find the minimal computational subgraph, or circuit, that explains LM behavior on a given task. Most studies determine which edges belong in a LM's circuit by performing causal interventions on each edge independently, but this scales poorly with model size. Edge attribution patching (EAP), gradient-based approximation to interventions, has emerged as a scalable but imperfect solution to this problem. In this paper, we introduce a new method - EAP with integrated gradients (EAP-IG) - that aims to better maintain a core property of circuits: faithfulness. A circuit is faithful if all model edges outside the circuit can be ablated without changing the model's performance on the task; faithfulness is what justifies studying circuits, rather than the full model. Our experiments demonstrate that circuits found using EAP are less faithful than those found using EAP-IG, even though both have high node overlap with circuits found previously using causal interventions. We conclude more generally that when using circuits to compare the mechanisms models use to solve tasks, faithfulness, not overlap, is what should be measured.
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, humans show clear production preferences 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, with all models failing in tasks that require high-level reasoning such as assigning quantifiers.
In everyday language use, speakers frequently utter and interpret sentences that are semantically underspecified, namely, whose content is insufficient to fully convey their message or interpret them univocally. For example, to interpret the underspecified sentence "Don't spend too much", which leaves implicit what (not) to spend, additional linguistic context or outside knowledge is needed. In this work, we propose a novel Dataset of semantically Underspecified Sentences grouped by Type (DUST) and use it to study whether pre-trained language models (LMs) correctly identify and interpret underspecified sentences. We find that newer LMs are reasonably able to identify underspecified sentences when explicitly prompted. However, interpreting them correctly is much harder for any LMs. Our experiments show that when interpreting underspecified sentences, LMs exhibit little uncertainty, contrary to what theoretical accounts of underspecification would predict. Overall, our study reveals limitations in current models' processing of sentence semantics and highlights the importance of using naturalistic data and communicative scenarios when evaluating LMs' language capabilities.
There is an intricate relation between the properties of an image and how humans behave while describing the image. This behavior shows ample variation, as manifested in human signals such as eye movements and when humans start to describe the image. Despite the value of such signals of visuo-linguistic variation, they are virtually disregarded in the training of current pretrained models, which motivates further investigation. Using a corpus of Dutch image descriptions with concurrently collected eye-tracking data, we explore the nature of the variation in visuo-linguistic signals, and find that they correlate with each other. Given this result, we hypothesize that variation stems partly from the properties of the images, and explore whether image representations encoded by pretrained vision encoders can capture such variation. Our results indicate that pretrained models do so to a weak-to-moderate degree, suggesting that the models lack biases about what makes a stimulus complex for humans and what leads to variations in human outputs.
A proper evaluation of stories generated for a sequence of images -- the task commonly referred to as visual storytelling -- must consider multiple aspects, such as coherence, grammatical correctness, and visual grounding. In this work, we focus on evaluating the degree of grounding, that is, the extent to which a story is about the entities shown in the images. We analyze current metrics, both designed for this purpose and for general vision-text alignment. Given their observed shortcomings, we propose a novel evaluation tool, GROOViST, that accounts for cross-modal dependencies, temporal misalignments (the fact that the order in which entities appear in the story and the image sequence may not match), and human intuitions on visual grounding. An additional advantage of GROOViST is its modular design, where the contribution of each component can be assessed and interpreted individually.
Despite the impressive performance achieved by pre-trained language-and-vision models in downstream tasks, it remains an open question whether this reflects a proper understanding of image-text interaction. In this work, we explore to what extent they handle basic linguistic constructions -- active-passive voice, coordination, and relative clauses -- that even preschool children can typically master. We present BLA, a novel, automatically constructed benchmark to evaluate multimodal models on these Basic Language Abilities. We show that different types of Transformer-based systems, such as CLIP, ViLBERT, and BLIP2, generally struggle with BLA in a zero-shot setting, in line with previous findings. Our experiments, in particular, show that most of the tested models only marginally benefit when fine-tuned or prompted with construction-specific samples. Yet, the generative BLIP2 shows promising trends, especially in an in-context learning setting. This opens the door to using BLA not only as an evaluation benchmark but also to improve models' basic language abilities.
Animacy - whether an entity is alive and sentient - is fundamental to cognitive processing, impacting areas such as memory, vision, and language. However, animacy is not always expressed directly in language: in English it often manifests indirectly, in the form of selectional constraints on verbs and adjectives. This poses a potential issue for transformer language models (LMs): they often train only on text, and thus lack access to extralinguistic information from which humans learn about animacy. We ask: how does this impact LMs' animacy processing - do they still behave as humans do? We answer this question using open-source LMs. Like previous studies, we find that LMs behave much like humans when presented with entities whose animacy is typical. However, we also show that even when presented with stories about atypically animate entities, such as a peanut in love, LMs adapt: they treat these entities as animate, though they do not adapt as well as humans. Even when the context indicating atypical animacy is very short, LMs pick up on subtle clues and change their behavior. We conclude that despite the limited signal through which LMs can learn about animacy, they are indeed sensitive to the relevant lexical semantic nuances available in English.
Feature attribution methods have become a staple method to disentangle the complex behavior of black box models. Despite their success, some scholars have argued that such methods suffer from a serious flaw: they do not allow a reliable interpretation in terms of human concepts. Simply put, visualizing an array of feature contributions is not enough for humans to conclude something about a model's internal representations, and confirmation bias can trick users into false beliefs about model behavior. We argue that a structured approach is required to test whether our hypotheses on the model are confirmed by the feature attributions. This is what we call the "semantic match" between human concepts and (sub-symbolic) explanations. Building on the conceptual framework put forward in Cin\`a et al. [2023], we propose a structured approach to evaluate semantic match in practice. We showcase the procedure in a suite of experiments spanning tabular and image data, and show how the assessment of semantic match can give insight into both desirable (e.g., focusing on an object relevant for prediction) and undesirable model behaviors (e.g., focusing on a spurious correlation). We couple our experimental results with an analysis on the metrics to measure semantic match, and argue that this approach constitutes the first step towards resolving the issue of confirmation bias in XAI.
Intelligent systems that aim at mastering language as humans do must deal with its semantic underspecification, namely, the possibility for a linguistic signal to convey only part of the information needed for communication to succeed. Consider the usages of the pronoun they, which can leave the gender and number of its referent(s) underspecified. Semantic underspecification is not a bug but a crucial language feature that boosts its storage and processing efficiency. Indeed, human speakers can quickly and effortlessly integrate semantically-underspecified linguistic signals with a wide range of non-linguistic information, e.g., the multimodal context, social or cultural conventions, and shared knowledge. Standard NLP models have, in principle, no or limited access to such extra information, while multimodal systems grounding language into other modalities, such as vision, are naturally equipped to account for this phenomenon. However, we show that they struggle with it, which could negatively affect their performance and lead to harmful consequences when used for applications. In this position paper, we argue that our community should be aware of semantic underspecification if it aims to develop language technology that can successfully interact with human users. We discuss some applications where mastering it is crucial and outline a few directions toward achieving this goal.
Dialogue participants may have varying levels of knowledge about the topic under discussion. In such cases, it is essential for speakers to adapt their utterances by taking their audience into account. Yet, it is an open question how such adaptation can be modelled in computational agents. In this paper, we model a visually grounded referential game between a knowledgeable speaker and a listener with more limited visual and linguistic experience. Inspired by psycholinguistic theories, we endow our speaker with the ability to adapt its referring expressions via a simulation module that monitors the effectiveness of planned utterances from the listener's perspective. We propose an adaptation mechanism building on plug-and-play approaches to controlled language generation, where utterance generation is steered on the fly by the simulator without finetuning the speaker's underlying language model. Our results and analyses show that our approach is effective: the speaker's utterances become closer to the listener's domain of expertise, which leads to higher communicative success.