Computational models of word meaning in use

点击翻译按钮获取中文摘要

The human capacity for understanding meaning is impressive: in order to understand the image a speaker is trying to evoke with a sentence like The bat has red eyes, we need to make a host of decisions. These range from coarse-grained ones (are we talking about the club or the flying mammal?), to fine-grained ones (does the speaker mean that the eyes are bloodshot or that the irises are red?). The apparent ease with which humans process meaning stands in stark contrast with the current capacities of computational systems. Nevertheless, allowing computers to arrive at a deeper understanding of meaning is important if we want to improve appliances like Siri and Alexa, and more generally, the Artificial Intelligence (AI) systems that increasingly form part of our everyday life.******In this project, I will continue my line of work integrating insights from different scientific disciplines to build computational software systems or 'models' that carry out language-related tasks in human-like ways. The specific project goals are centred around the question how a computational model can correctly identify the intended interpretation of a word in context, both at the coarser-grained (whether bat refers to the club or the mammal) and finer-grained (what is meant with red in red eyes) level. I will explore novel computational formulations of how words contribute to the interpretation as well as little-explored sources of information for these computational models, such as translation data. For the development of these formulations, I draw inspiration from the study of human language processing in linguistics and cognitive science -- after all, no one can interpret language as well as humans can!******An important aspect of this project is a focus on ways of assessing the validity of computational models of interpretation. Currently, the evaluation of such models relies on datasets of, for instance, pairs of words in different contexts with similarity ratings. These datasets are valuable starting points, but also display limitations. First, the items may not reflect the range of interpretation tasks an AI system faces 'in the wild', which I intend to overcome by looking more closely into what items we test the models on. Second, developing such datasets for multiple languages ensures the generalizability of the computational interpretation models to languages beyond English. Third, similarity ratings (and comparable tasks) are conscious measures, which may introduce various kinds of unwanted biases. Testing the models on data from psychological experiments on language processing that tap into unconscious processes is another way of evaluating the models I intend to develop in this project.******Together, the novel approaches that interpret language in a human-like way, as well as stronger ways of evaluating those approaches, allow for more open-ended computational tools that interpret language in a way that matches human expectations in conversation.**