Learning visually grounded meaning representations
Silberer, Carina Helga
Humans possess a rich semantic knowledge of words and concepts which captures the perceivable physical properties of their real-world referents and their relations. Encoding this knowledge or some of its aspects is the goal of computational models of semantic representation and has been the subject of considerable research in cognitive science, natural language processing, and related areas. Existing models have placed emphasis on different aspects of meaning, depending ultimately on the task at hand. Typically, such models have been used in tasks addressing the simulation of behavioural phenomena, e.g., lexical priming or categorisation, as well as in natural language applications, such as information retrieval, document classification, or semantic role labelling. A major strand of research popular across disciplines focuses on models which induce semantic representations from text corpora. These models are based on the hypothesis that the meaning of words is established by their distributional relation to other words (Harris, 1954). Despite their widespread use, distributional models of word meaning have been criticised as ‘disembodied’ in that they are not grounded in perception and action (Perfetti, 1998; Barsalou, 1999; Glenberg and Kaschak, 2002). This lack of grounding contrasts with many experimental studies suggesting that meaning is acquired not only from exposure to the linguistic environment but also from our interaction with the physical world (Landau et al., 1998; Bornstein et al., 2004). This criticism has led to the emergence of new models aiming at inducing perceptually grounded semantic representations. Essentially, existing approaches learn meaning representations from multiple views corresponding to different modalities, i.e. linguistic and perceptual input. To approximate the perceptual modality, previous work has relied largely on semantic attributes collected from humans (e.g., is round, is sour), or on automatically extracted image features. Semantic attributes have a long-standing tradition in cognitive science and are thought to represent salient psychological aspects of word meaning including multisensory information. However, their elicitation from human subjects limits the scope of computational models to a small number of concepts for which attributes are available. In this thesis, we present an approach which draws inspiration from the successful application of attribute classifiers in image classification, and represent images and the concepts depicted by them by automatically predicted visual attributes. To this end, we create a dataset comprising nearly 700K images and a taxonomy of 636 visual attributes and use it to train attribute classifiers. We show that their predictions can act as a substitute for human-produced attributes without any critical information loss. In line with the attribute-based approximation of the visual modality, we represent the linguistic modality by textual attributes which we obtain with an off-the-shelf distributional model. Having first established this core contribution of a novel modelling framework for grounded meaning representations based on semantic attributes, we show that these can be integrated into existing approaches to perceptually grounded representations. We then introduce a model which is formulated as a stacked autoencoder (a variant of multilayer neural networks), which learns higher-level meaning representations by mapping words and images, represented by attributes, into a common embedding space. In contrast to most previous approaches to multimodal learning using different variants of deep networks and data sources, our model is defined at a finer level of granularity—it computes representations for individual words and is unique in its use of attributes as a means of representing the textual and visual modalities. We evaluate the effectiveness of the representations learnt by our model by assessing its ability to account for human behaviour on three semantic tasks, namely word similarity, concept categorisation, and typicality of category members. With respect to the word similarity task, we focus on the model’s ability to capture similarity in both the meaning and appearance of the words’ referents. Since existing benchmark datasets on word similarity do not distinguish between these two dimensions and often contain abstract words, we create a new dataset in a large-scale experiment where participants are asked to give two ratings per word pair expressing their semantic and visual similarity, respectively. Experimental results show that our model learns meaningful representations which are more accurate than models based on individual modalities or different modality integration mechanisms. The presented model is furthermore able to predict textual attributes for new concepts given their visual attribute predictions only, which we demonstrate by comparing model output with human generated attributes. Finally, we show the model’s effectiveness in an image-based task on visual category learning, in which images are used as a stand-in for real-world objects.