Denis Paperno

Bart Evelo, Meaghan Fowlie, Denis Paperno

Do neural models, such as Large Language Models, genuinely acquire compositional abilities for interpretation of natural language? When we talk about semantic interpretation, we can distinguish two complementary aspects: establishing what an expression refers to in the world (which we call the Extensional task) and representing its sense in a structured way (which we call the Intensional task). We evaluate LLMs and humans on both tasks in the setting of the Personal Relation Task (Paperno 2022) in which, given a universe of people and their relationships with each other, one is asked to interpret a noun phrase such as "Amber's parent's friend". Here, for the Intensional task, the answer is the formula "friend(parent(amber))", and for the Extensional task, the person. We find that humans and LLMs show opposite strengths: humans perform better on Extensional than Intensional tasks, and LLMs vice versa. Our methodology brings greater nuance to the understanding of compositional abilities in modern machine learning models. Our results support the notion that the lack of referential grounding in LLM training is a crucial missing component in mimicking human-like language understanding.

Timothee Mickus, Kees van Deemter, Mathieu Constant et al.

Word embeddings have advanced the state of the art in NLP across numerous tasks. Understanding the contents of dense neural representations is of utmost interest to the computational semantics community. We propose to focus on relating these opaque word vectors with human-readable definitions, as found in dictionaries. This problem naturally divides into two subtasks: converting definitions into embeddings, and converting embeddings into definitions. This task was conducted in a multilingual setting, using comparable sets of embeddings trained homogeneously.

Timothee Mickus, Denis Paperno, Mathieu Constant

Pretrained embeddings based on the Transformer architecture have taken the NLP community by storm. We show that they can mathematically be reframed as a sum of vector factors and showcase how to use this reframing to study the impact of each component. We provide evidence that multi-head attentions and feed-forwards are not equally useful in all downstream applications, as well as a quantitative overview of the effects of finetuning on the overall embedding space. This approach allows us to draw connections to a wide range of previous studies, from vector space anisotropy to attention weights.

Sofia Nikiforova, Tejaswini Deoskar, Denis Paperno et al.

Accurately reporting what objects are depicted in an image is largely a solved problem in automatic caption generation. The next big challenge on the way to truly humanlike captioning is being able to incorporate the context of the image and related real world knowledge. We tackle this challenge by creating an end-to-end caption generation system that makes extensive use of image-specific encyclopedic data. Our approach includes a novel way of using image location to identify relevant open-domain facts in an external knowledge base, with their subsequent integration into the captioning pipeline at both the encoding and decoding stages. Our system is trained and tested on a new dataset with naturally produced knowledge-rich captions, and achieves significant improvements over multiple baselines. We empirically demonstrate that our approach is effective for generating contextualized captions with encyclopedic knowledge that is both factually accurate and relevant to the image.

Claudia Tagliaferri, Sofia Axioti, Albert Gatt et al.

Derivationally related words, such as "runner" and "running", exhibit semantic differences which also elicit different visual scenarios. In this paper, we ask whether Vision and Language (V\&L) models capture such distinctions at the morphological level, using a a new methodology and dataset. We compare the results from V\&L models to human judgements and find that models' predictions differ from those of human participants, in particular displaying a grammatical bias. We further investigate whether the human-model misalignment is related to model architecture. Our methodology, developed on one specific morphological contrast, can be further extended for testing models on capturing other nuanced language features.

Aleksey Tikhonov, Lisa Bylinina, Denis Paperno

Multimodal embeddings aim to enrich the semantic information in neural representations of language compared to text-only models. While different embeddings exhibit different applicability and performance on downstream tasks, little is known about the systematic representation differences attributed to the visual modality. Our paper compares word embeddings from three vision-and-language models (CLIP, OpenCLIP and Multilingual CLIP) and three text-only models, with static (FastText) as well as contextual representations (multilingual BERT; XLM-RoBERTa). This is the first large-scale study of the effect of visual grounding on language representations, including 46 semantic parameters. We identify meaning properties and relations that characterize words whose embeddings are most affected by the inclusion of visual modality in the training data; that is, points where visual grounding turns out most important. We find that the effect of visual modality correlates most with denotational semantic properties related to concreteness, but is also detected for several specific semantic classes, as well as for valence, a sentiment-related connotational property of linguistic expressions.

Timothee Mickus, Elaine Zosa, Denis Paperno

Grounding has been argued to be a crucial component towards the development of more complete and truly semantically competent artificial intelligence systems. Literature has divided into two camps: While some argue that grounding allows for qualitatively different generalizations, others believe it can be compensated by mono-modal data quantity. Limited empirical evidence has emerged for or against either position, which we argue is due to the methodological challenges that come with studying grounding and its effects on NLP systems. In this paper, we establish a methodological framework for studying what the effects are - if any - of providing models with richer input sources than text-only. The crux of it lies in the construction of comparable samples of populations of models trained on different input modalities, so that we can tease apart the qualitative effects of different input sources from quantifiable model performances. Experiments using this framework reveal qualitative differences in model behavior between cross-modally grounded, cross-lingually grounded, and ungrounded models, which we measure both at a global dataset level as well as for specific word representations, depending on how concrete their semantics is.

Shaomu Tan, Denis Paperno

In many real-world scenarios, the absence of external knowledge source like Wikipedia restricts question answering systems to rely on latent internal knowledge in limited dialogue data. In addition, humans often seek answers by asking several questions for more comprehensive information. As the dialog becomes more extensive, machines are challenged to refer to previous conversation rounds to answer questions. In this work, we propose to leverage latent knowledge in existing conversation logs via a neural Retrieval-Reading system, enhanced with a TFIDF-based text summarizer refining lengthy conversational history to alleviate the long context issue. Our experiments show that our Retrieval-Reading system can exploit retrieved background knowledge to generate significantly better answers. The results also indicate that our context summarizer significantly helps both the retriever and the reader by introducing more concise and less noisy contextual information.

Yingjin Song, Yupei Du, Denis Paperno et al.

This paper introduces the TempVS benchmark, which focuses on temporal grounding and reasoning capabilities of Multimodal Large Language Models (MLLMs) in image sequences. TempVS consists of three main tests (i.e., event relation inference, sentence ordering and image ordering), each accompanied with a basic grounding test. TempVS requires MLLMs to rely on both visual and linguistic modalities to understand the temporal order of events. We evaluate 38 state-of-the-art MLLMs, demonstrating that models struggle to solve TempVS, with a substantial performance gap compared to human capabilities. We also provide fine-grained insights that suggest promising directions for future research. Our TempVS benchmark data and code are available at https://github.com/yjsong22/TempVS.

Jim Maar, Denis Paperno, Callum Stuart McDougall et al.

Prior work suggests that language models, while trained on next token prediction, show implicit planning behavior: they may select the next token in preparation to a predicted future token, such as a likely rhyming word, as supported by a prior qualitative study of Claude 3.5 Haiku using a cross-layer transcoder. We propose much simpler techniques for assessing implicit planning in language models. With case studies on rhyme poetry generation and question answering, we demonstrate that our methodology easily scales to many models. Across models, we find that the generated rhyme (e.g. "-ight") or answer to a question ("whale") can be manipulated by steering at the end of the preceding line with a vector, affecting the generation of intermediate tokens leading up to the rhyme or answer word. We show that implicit planning is a universal mechanism, present in smaller models than previously thought, starting from 1B parameters. Our methodology offers a widely applicable direct way to study implicit planning abilities of LLMs. More broadly, understanding planning abilities of language models can inform decisions in AI safety and control.

Yingjin Song, Denis Paperno, Albert Gatt

Visual storytelling systems generate multi-sentence stories from image sequences. In this task, capturing contextual information and bridging visual variation bring additional challenges. We propose a simple yet effective framework that leverages the generalization capabilities of pretrained foundation models, only training a lightweight vision-language mapping network to connect modalities, while incorporating context to enhance coherence. We introduce a multimodal contrastive objective that also improves visual relevance and story informativeness. Extensive experimental results, across both automatic metrics and human evaluations, demonstrate that the stories generated by our framework are diverse, coherent, informative, and interesting.

Daniil Ignatev, Ayman Santeer, Albert Gatt et al.

We propose a zero-shot method for Natural Language Inference (NLI) that leverages multimodal representations by grounding language in visual contexts. Our approach generates visual representations of premises using text-to-image models and performs inference by comparing these representations with textual hypotheses. We evaluate two inference techniques: cosine similarity and visual question answering. Our method achieves high accuracy without task-specific fine-tuning, demonstrating robustness against textual biases and surface heuristics. Additionally, we design a controlled adversarial dataset to validate the robustness of our approach. Our findings suggest that leveraging visual modality as a meaning representation provides a promising direction for robust natural language understanding.

Daniil Ignatev, Denis Paperno, Massimo Poesio

The task of perspective-aware classification introduces a bottleneck in terms of parametric efficiency that did not get enough recognition in existing studies. In this article, we aim to address this issue by applying an existing architecture, the hypernetwork+adapters combination, to perspectivist classification. Ultimately, we arrive at a solution that can compete with specialized models in adopting user perspectives on hate speech and toxicity detection, while also making use of considerably fewer parameters. Our solution is architecture-agnostic and can be applied to a wide range of base models out of the box.

Timothee Mickus, Mathieu Constant, Denis Paperno

Can language models learn grounded representations from text distribution alone? This question is both central and recurrent in natural language processing; authors generally agree that grounding requires more than textual distribution. We propose to experimentally test this claim: if any two words have different meanings and yet cannot be distinguished from distribution alone, then grounding is out of the reach of text-based models. To that end, we present early work on an online game for the collection of human judgments on the distributional similarity of word pairs in five languages. We further report early results of our data collection campaign.

Timothee Mickus, Timothée Bernard, Denis Paperno

Compositionality is a widely discussed property of natural languages, although its exact definition has been elusive. We focus on the proposal that compositionality can be assessed by measuring meaning-form correlation. We analyze meaning-form correlation on three sets of languages: (i) artificial toy languages tailored to be compositional, (ii) a set of English dictionary definitions, and (iii) a set of English sentences drawn from literature. We find that linguistic phenomena such as synonymy and ungrounded stop-words weigh on MFC measurements, and that straightforward methods to mitigate their effects have widely varying results depending on the dataset they are applied to. Data and code are made publicly available.

Timothee Mickus, Denis Paperno, Mathieu Constant et al.

Contextualized word embeddings, i.e. vector representations for words in context, are naturally seen as an extension of previous noncontextual distributional semantic models. In this work, we focus on BERT, a deep neural network that produces contextualized embeddings and has set the state-of-the-art in several semantic tasks, and study the semantic coherence of its embedding space. While showing a tendency towards coherence, BERT does not fully live up to the natural expectations for a semantic vector space. In particular, we find that the position of the sentence in which a word occurs, while having no meaning correlates, leaves a noticeable trace on the word embeddings and disturbs similarity relationships.

Timothee Mickus, Denis Paperno, Mathieu Constant

Defining words in a textual context is a useful task both for practical purposes and for gaining insight into distributed word representations. Building on the distributional hypothesis, we argue here that the most natural formalization of definition modeling is to treat it as a sequence-to-sequence task, rather than a word-to-sequence task: given an input sequence with a highlighted word, generate a contextually appropriate definition for it. We implement this approach in a Transformer-based sequence-to-sequence model. Our proposal allows to train contextualization and definition generation in an end-to-end fashion, which is a conceptual improvement over earlier works. We achieve state-of-the-art results both in contextual and non-contextual definition modeling.

Denis Paperno

Can recurrent neural nets, inspired by human sequential data processing, learn to understand language? We construct simplified datasets reflecting core properties of natural language as modeled in formal syntax and semantics: recursive syntactic structure and compositionality. We find LSTM and GRU networks to generalise to compositional interpretation well, but only in the most favorable learning settings, with a well-paced curriculum, extensive training data, and left-to-right (but not right-to-left) composition.

Alexander Panchenko, Natalia Loukachevitch, Dmitry Ustalov et al.

The paper gives an overview of the Russian Semantic Similarity Evaluation (RUSSE) shared task held in conjunction with the Dialogue 2015 conference. There exist a lot of comparative studies on semantic similarity, yet no analysis of such measures was ever performed for the Russian language. Exploring this problem for the Russian language is even more interesting, because this language has features, such as rich morphology and free word order, which make it significantly different from English, German, and other well-studied languages. We attempt to bridge this gap by proposing a shared task on the semantic similarity of Russian nouns. Our key contribution is an evaluation methodology based on four novel benchmark datasets for the Russian language. Our analysis of the 105 submissions from 19 teams reveals that successful approaches for English, such as distributional and skip-gram models, are directly applicable to Russian as well. On the one hand, the best results in the contest were obtained by sophisticated supervised models that combine evidence from different sources. On the other hand, completely unsupervised approaches, such as a skip-gram model estimated on a large-scale corpus, were able score among the top 5 systems.

Alexander Panchenko, Dmitry Ustalov, Nikolay Arefyev et al.

Semantic relatedness of terms represents similarity of meaning by a numerical score. On the one hand, humans easily make judgments about semantic relatedness. On the other hand, this kind of information is useful in language processing systems. While semantic relatedness has been extensively studied for English using numerous language resources, such as associative norms, human judgments, and datasets generated from lexical databases, no evaluation resources of this kind have been available for Russian to date. Our contribution addresses this problem. We present five language resources of different scale and purpose for Russian semantic relatedness, each being a list of triples (word_i, word_j, relatedness_ij). Four of them are designed for evaluation of systems for computing semantic relatedness, complementing each other in terms of the semantic relation type they represent. These benchmarks were used to organize a shared task on Russian semantic relatedness, which attracted 19 teams. We use one of the best approaches identified in this competition to generate the fifth high-coverage resource, the first open distributional thesaurus of Russian. Multiple evaluations of this thesaurus, including a large-scale crowdsourcing study involving native speakers, indicate its high accuracy.

Denis Paperno, Germán Kruszewski, Angeliki Lazaridou et al.

We introduce LAMBADA, a dataset to evaluate the capabilities of computational models for text understanding by means of a word prediction task. LAMBADA is a collection of narrative passages sharing the characteristic that human subjects are able to guess their last word if they are exposed to the whole passage, but not if they only see the last sentence preceding the target word. To succeed on LAMBADA, computational models cannot simply rely on local context, but must be able to keep track of information in the broader discourse. We show that LAMBADA exemplifies a wide range of linguistic phenomena, and that none of several state-of-the-art language models reaches accuracy above 1% on this novel benchmark. We thus propose LAMBADA as a challenging test set, meant to encourage the development of new models capable of genuine understanding of broad context in natural language text.