ABSTRACT
Neural networks have proven to be very successful in automatically capturing the composition of language and different structures across a range of multi-modal tasks. Thus, an important question to investigate is how neural networks learn and organise such structures. Numerous studies have examined the knowledge captured by language models (LSTMs, transformers) and vision architectures (CNNs, vision transformers) for respective uni-modal tasks. However, very few have explored what structures are acquired by multi-modal transformers where linguistic and visual features are combined. It is critical to understand the representations learned by each modality, their respective interplay, and the task's effect on these representations in large-scale architectures. In this paper, we take a multi-modal transformer trained for image captioning and examine the structure of the self-attention patterns extracted from the visual stream. Our results indicate that the information about different relations between objects in the visual stream is hierarchical and varies from local to a global object-level understanding of the image. In particular, while visual representations in the first layers encode the knowledge of relations between semantically similar object detections, often constituting neighbouring objects, deeper layers expand their attention across more distant objects and learn global relations between them. We also show that globally attended objects in deeper layers can be linked with entities described in image descriptions, indicating a critical finding - the indirect effect of language on visual representations. In addition, we highlight how object-based input representations affect the structure of learned visual knowledge and guide the model towards more accurate image descriptions. A parallel question that we investigate is whether the insights from cognitive science echo the structure of representations that the current neural architecture learns. The proposed analysis of the inner workings of multi-modal transformers can be used to better understand and improve on such problems as pre-training of large-scale multi-modal architectures, multi-modal information fusion and probing of attention weights. In general, we contribute to the explainable multi-modal natural language processing and currently shallow understanding of how the input representations and the structure of the multi-modal transformer affect visual representations.
ABSTRACT
MOTIVATION: Scholarly biomedical publications report on the findings of a research investigation. Scientists use a well-established discourse structure to relate their work to the state of the art, express their own motivation and hypotheses and report on their methods, results and conclusions. In previous work, we have proposed ways to explicitly annotate the structure of scientific investigations in scholarly publications. Here we present the means to facilitate automatic access to the scientific discourse of articles by automating the recognition of 11 categories at the sentence level, which we call Core Scientific Concepts (CoreSCs). These include: Hypothesis, Motivation, Goal, Object, Background, Method, Experiment, Model, Observation, Result and Conclusion. CoreSCs provide the structure and context to all statements and relations within an article and their automatic recognition can greatly facilitate biomedical information extraction by characterizing the different types of facts, hypotheses and evidence available in a scientific publication. RESULTS: We have trained and compared machine learning classifiers (support vector machines and conditional random fields) on a corpus of 265 full articles in biochemistry and chemistry to automatically recognize CoreSCs. We have evaluated our automatic classifications against a manually annotated gold standard, and have achieved promising accuracies with 'Experiment', 'Background' and 'Model' being the categories with the highest F1-scores (76%, 62% and 53%, respectively). We have analysed the task of CoreSC annotation both from a sentence classification as well as sequence labelling perspective and we present a detailed feature evaluation. The most discriminative features are local sentence features such as unigrams, bigrams and grammatical dependencies while features encoding the document structure, such as section headings, also play an important role for some of the categories. We discuss the usefulness of automatically generated CoreSCs in two biomedical applications as well as work in progress. AVAILABILITY: A web-based tool for the automatic annotation of articles with CoreSCs and corresponding documentation is available online at http://www.sapientaproject.com/software http://www.sapientaproject.com also contains detailed information pertaining to CoreSC annotation and links to annotation guidelines as well as a corpus of manually annotated articles, which served as our training data. CONTACT: liakata@ebi.ac.uk SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.
