Attention as Grounding: Exploring Textual and Cross-Modal Attention on Entities and Relations in Language-and-Vision Transformer

Findings of the Association for Computational Linguistics: ACL 2022, pages 4062 - 4073 May 22-27, 2022 c⃝2022 Association for Computational Linguistics Attention as Grounding: Exploring Textual and Cross-Modal Attention on Entities and Relations in Language-and-Vision Transformer Nikolai Ilinykh and Simon Dobnik Centre for Linguistic Theory and Studies in Probability (CLASP) Department of Philosophy, Linguistics and Theory of Science (FLoV) University of Gothenburg, Sweden {nikolai.ilinykh,simon.dobnik}@gu.se Abstract We explore how a multi-modal transformer trained for generation of longer image descriptions learns syntactic and semantic representations about entities and relations grounded in objects at the level of masked self-attention (text generation) and cross- modal attention (information fusion). We observe that cross-attention learns the visual grounding of noun phrases into objects and high-level semantic information about spatial relations, while text-to-text attention captures low-level syntactic knowledge between words. This concludes that language models in a multi-modal task learn different semantic information about objects and relations cross-modally and uni-modally (text-only). Our code is available here: https://github.com/GU-CLASP/ attention-as-grounding. 1 Introduction In this paper, we examine what kind of knowledge is encoded in the multi-modal transformer. Exist- ing work has mostly looked at the knowledge cap- tured in models that operate with a single modality (text). For instance, previous research has shown that the attention weights in large-scale models, e.g. BERT (Devlin et al., 2019), implicitly encode knowledge of sentence structure (Raganato and Tiedemann, 2018; Ravishankar et al., 2021), part- of-speech tags, syntactic dependencies (Clark et al., 2019; Vig and Belinkov, 2019), subject-verb agree- ment between words (Goldberg, 2019), and even in- formation about textual co-reference (Tenney et al., 2019). Only a few papers have inspected what is captured by multi-modal architectures. Cao et al. (2020) demonstrate that the attention heads in image-and-text transformers effectively encode linguistic and cross-modal knowledge. Ilinykh and Dobnik (2021) provide the analysis of how lan- guage representations are indirectly affected by visual information in language-and-vision model. Here we inspect what the model learns about two types of words in the multi-modal setting: (i) words denoting objects in the scene (e.g. “a red chair”), (ii) words depicting spatial relations between ob- jects (e.g. “a chair next to the table”). While it is relatively simple to associate nouns with specific image regions, words describing relations are much harder to ground (Lu et al., 2017), possibly because visual representations are typically designed to cap- ture objects without any explicit knowledge of re- lations. Secondly, grounded relations depend on knowledge from both vision and language modali- ties which contains information about the objects and their mode of interaction (what) as well as their physical location (where) (Ghanimifard and Dob- nik, 2019). Ideally, each relation (and also other types of words) should be grounded in both modali- ties, but to a different degree.1 However, studies of language-and-vision models indicate that they are frequently biased towards one modality, most often to language (Goyal et al., 2017). Therefore, the main research challenge is to develop architectures that learn to utilise an appropriate ratio of visual and language knowledge for generation …

Content truncated to 500 words