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.

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.

Clarification questions are an essential dialogue tool to signal misunderstanding, ambiguities, and under-specification in language use. While humans are able to resolve uncertainty by asking questions since childhood, modern dialogue systems struggle to generate effective questions. To make progress in this direction, in this work we take a collaborative dialogue task as a testbed and study how model uncertainty relates to human uncertainty—an as yet under-explored problem. We show that model uncertainty does not mirror human clarification-seeking behavior, which suggests that using human clarification questions as supervision for deciding when to ask may not be the most effective way to resolve model uncertainty. To address this issue, we propose an approach to generating clarification questions based on model uncertainty estimation, compare it to several alternatives, and show that it leads to significant improvements in terms of task success. Our findings highlight the importance of equipping dialogue systems with the ability to assess their own uncertainty and exploit in interaction.

Artificial agents are nowadays challenged to perform embodied AI tasks. To succeed, agents must understand the meaning of verbs and how their corresponding actions transform the surrounding world. In this work, we propose ACT-Thor, a novel controlled benchmark for embodied action understanding. We use the AI2-THOR simulated environment to produce a controlled setup in which an agent, given a before-image and an associated action command, has to determine what the correct after-image is among a set of possible candidates. First, we assess the feasibility of the task via a human evaluation that resulted in 81.4% accuracy, and very high inter-annotator agreement (84.9%). Second, we design both unimodal and multimodal baselines, using state-of-the-art visual feature extractors. Our evaluation and error analysis suggest that only models that have a very structured representation of the actions together with powerful visual features can perform well on the task. However, they still fall behind human performance in a zero-shot scenario where the model is exposed to unseen (action, object) pairs. This paves the way for a systematic way of evaluating embodied AI agents that understand grounded actions.

Pre-trained Vision and Language Transformers achieve high performance on downstream tasks due to their ability to transfer representational knowledge accumulated during pretraining on substantial amounts of data. In this paper, we ask whether it is possible to compete with such models using features based on transferred (pre-trained, frozen) representations combined with a lightweight architecture. We take a multimodal guessing task as our testbed, GuessWhat?!. An ensemble of our lightweight model matches the performance of the finetuned pre-trained transformer (LXMERT). An uncertainty analysis of our ensemble shows that the lightweight transferred representations close the data uncertainty gap with LXMERT, while retaining model diversity leading to ensemble boost. We further demonstrate that LXMERT’s performance gain is due solely to its extra V&L pretraining rather than because of architectural improvements. These results argue for flexible integration of multiple features and lightweight models as a viable alternative to large, cumbersome, pre-trained models.

Generating goal-oriented questions in Visual Dialogue tasks is a challenging and longstanding problem. State-Of-The-Art systems are shown to generate questions that, although grammatically correct, often lack an effective strategy and sound unnatural to humans. Inspired by the cognitive literature on information search and cross-situational word learning, we design Confirm-it, a model based on a beam search re-ranking algorithm that guides an effective goal-oriented strategy by asking questions that confirm the model’s conjecture about the referent. We take the GuessWhat?! game as a case-study. We show that dialogues generated by Confirm-it are more natural and effective than beam search decoding without re-ranking.

Natural language generation systems have witnessed important progress in the last years, but they are shown to generate tokens that are unrelated to the source input. This problem affects computational models in many NLP tasks, and it is particularly unpleasant in multimodal systems. In this work, we assess the rate of object hallucination in multimodal conversational agents playing the GuessWhat?! referential game. Better visual processing has been shown to mitigate this issue in image captioning; hence, we adapt to the GuessWhat?! task the best visual processing models at disposal, and propose two new models to play the Questioner agent. We show that the new models generate few hallucinations compared to other renowned models available in the literature. Moreover, their hallucinations are less severe (affect task-accuracy less) and are more human-like. We also analyse where hallucinations tend to occur more often through the dialogue: hallucinations are less frequent in earlier turns, cause a cascade hallucination effect, and are often preceded by negative answers, which have been shown to be harder to ground.

In this paper, we define and evaluate a methodology for extracting history-dependent spatial questions from visual dialogues. We say that a question is history-dependent if it requires (parts of) its dialogue history to be interpreted. We argue that some kinds of visual questions define a context upon which a follow-up spatial question relies. We call the question that restricts the context: trigger, and we call the spatial question that requires the trigger question to be answered: zoomer. We automatically extract different trigger and zoomer pairs based on the visual property that the questions rely on (e.g. color, number). We manually annotate the automatically extracted trigger and zoomer pairs to verify which zoomers require their trigger. We implement a simple baseline architecture based on a SOTA multimodal encoder. Our results reveal that there is much room for improvement for answering history-dependent questions.

When training a model on referential dialogue guessing games, the best model is usually chosen based on its task success. We show that in the popular end-to-end approach, this choice prevents the model from learning to generate linguistically richer dialogues, since the acquisition of language proficiency takes longer than learning the guessing task. By comparing models playing different games (GuessWhat, GuessWhich, and Mutual Friends), we show that this discrepancy is model- and task-agnostic. We investigate whether and when better language quality could lead to higher task success. We show that in GuessWhat, models could increase their accuracy if they learn to ground, encode, and decode also words that do not occur frequently in the training set.

An interesting challenge for situated dialogue systems is referential visual dialog: by asking questions, the system has to identify the referent to which the user refers to. Task success is the standard metric used to evaluate these systems. However, it does not consider how effective each question is, that is how much each question contributes to the goal. We propose a new metric, that measures question effectiveness. As a preliminary study, we report the new metric for state of the art publicly available models on GuessWhat?!. Surprisingly, successful dialogues do not have a higher percentage of effective questions than failed dialogues. This suggests that a system with high task success is not necessarily one that generates good questions.

In this paper, we study the grounding skills required to answer spatial questions asked by humans while playing the GuessWhat?! game. We propose a classification for spatial questions dividing them into absolute, relational, and group questions. We build a new answerer model based on the LXMERT multimodal transformer and we compare a baseline with and without visual features of the scene. We are interested in studying how the attention mechanisms of LXMERT are used to answer spatial questions since they require putting attention on more than one region simultaneously and spotting the relation holding among them. We show that our proposed model outperforms the baseline by a large extent (9.70% on spatial questions and 6.27% overall). By analyzing LXMERT errors and its attention mechanisms, we find that our classification helps to gain a better understanding of the skills required to answer different spatial questions.

Task success is the standard metric used to evaluate referential visual dialogue systems. In this paper we propose two new metrics that evaluate how each question contributes to the goal. First, we measure how effective each question is by evaluating whether the question discards objects that are not the referent. Second, we define referring questions as those that univocally identify one object in the image. We report the new metrics for human dialogues and for state of the art publicly available models on GuessWhat?!. Regarding our first metric, we find that successful dialogues do not have a higher percentage of effective questions for most models. With respect to the second metric, humans make questions at the end of the dialogue that are referring, confirming their guess before guessing. Human dialogues that use this strategy have a higher task success but models do not seem to learn it.

Inspired by the literature on multisensory integration, we develop a computational model to ground quantifiers in perception. The model learns to pick, out of nine quantifiers (‘few’, ‘many’, ‘all’, etc.), the one that is more likely to describe the percent of animals in a visual-auditory input containing both animals and artifacts. We show that relying on concurrent sensory inputs increases model performance on the quantification task. Moreover, we evaluate the model in a situation in which only the auditory modality is given, while the visual one is ‘hallucinanted’ either from the auditory input itself or from a linguistic caption describing the quantity of entities in the auditory input. This way, the model exploits prior associations between modalities. We show that the model profits from the prior knowledge and outperforms the auditory-only setting.

Capturing semantic relations between sentences, such as entailment, is a long-standing challenge for computational semantics. Logic-based models analyse entailment in terms of possible worlds (interpretations, or situations) where a premise P entails a hypothesis H iff in all worlds where P is true, H is also true. Statistical models view this relationship probabilistically, addressing it in terms of whether a human would likely infer H from P. In this paper, we wish to bridge these two perspectives, by arguing for a visually-grounded version of the Textual Entailment task. Specifically, we ask whether models can perform better if, in addition to P and H, there is also an image (corresponding to the relevant “world” or “situation”). We use a multimodal version of the SNLI dataset (Bowman et al., 2015) and we compare “blind” and visually-augmented models of textual entailment. We show that visual information is beneficial, but we also conduct an in-depth error analysis that reveals that current multimodal models are not performing “grounding” in an optimal fashion.