The recently introduced "words as classifiers" model of grounded semantics (Kennington & Schlangen 2015) views words as classifiers on (perceptual) contexts, and composes the meaning of a phrase through 'soft' intersection of the denotations of its component words. The model is trained from instances of referential language use, and was first evaluated with references in a game-playing scenario with a small number of different types of objects. Here, we apply this model to a large set of real-world photographs (SAIAPR TC-12, (Escalante et al. 2010)) that contain objects with a much larger variety of types, and we show that it achieves good performance in a reference resolution task on this data set. We also extend the model to deal with quantification and negation, and evaluate these extensions, with good results. To investigate what the classifiers learn, we introduce 'intensional' and 'denotational' word vectors, and show that they capture meaning similarity in a way that is different from and complementary to word2vec word embeddings.

该研究通过预先训练的卷积神经网络提取图像特征，并将其与图片中的位置信息相结合，应用“以词为分类器”的模型研究语言和感知之间的联系，成果是一种能够在寻求表达的参考物边界盒子上表现出色的策略，并具有更简单和灵活的概念。

利用“词作分类器”模型解决照片中物体的引用问题