Frequent Frames, Flexible Frames and the Noun-Verb Asymmetry

Frequent Frames, Flexible Frames and the Noun-Verb Asymmetry Daniel Freudenthal, Julian Pine School of Psychology, University of Liverpool Gary Jones School of Psychology, Nottingham Trent University Fernand Gobet School of Psychology, University of Liverpool Abstract terms of the grammatical category of the items that occur in the central position. The notion of a frequent frame is therefore thought to provide a powerful cue that children might employ in the acquisition of syntactic categories. More recent work has confirmed the utility of frequent frames for French (Chemla et al. 2009), but results have been less promising for languages with relatively free word order such as Dutch (Erkelens, 2009) and German (Stumper et al. 2011). A major difference between the approaches of Redington et al. and Mintz is that the approach described by Redington is inherently graded and frequency sensitive in nature. Thus, in this approach, co-occurrence statistics are collected across all uses of a particular word. Depending on the exact implementation, the approach can also show varying degrees of frequency sensitivity with context vectors containing (rank orders of) word counts. Similarity is then expressed as a correlation-like measure across context vectors, which can be interpreted as a probability of two items being of the same class. This graded context-sensitivity is absent from Mintz’s approach. Thus, while Mintz’s analysis is restricted to the 45 most frequent frames, it clusters together all items that co-occur in one of these frames. The approach therefore ignores many contexts in which a word may occur, and instead clusters items on the basis of (potentially one) occurrence in specific high frequency contexts. Typically, mechanisms for extracting grammatical categories are evaluated in terms of accuracy (the extent to which items that are clustered together belong to the same syntactic category) and completeness (the extent to all items within one syntactic category are clustered together). St. Clair et al. (2010), as well as Monaghan (2004), compared frames and independent contexts as used by Redington et al. in the context of connectionist simulations, and found that frames were accurate but resulted in low completeness while independent contexts performed similarly in terms of accuracy but outperformed frames in terms of completeness. However, while high accuracy is clearly a desirable property of a mechanism that derives syntactic categories (children for instance make very few word class errors), it is less clear if high completeness is desirable, particularly if one is interested in modeling children’s early linguistic abilities. Thus, while children ultimately develop In this paper we compare several mechanisms for using distributional statistics to derive word class information. We contrast three different ways of computing statistics for independent left and right neighbours with the notion of a frequent frame. We also investigate the role of utterance boundaries as context items and weighting of frequency information in terms of the successful simulation of the noun-verb asymmetry. It is argued that independent contexts can classify items with a higher degree of accuracy than frequent frames, a finding that is more pronounced for larger input sets. Frequent frames classify a larger number of items, but do so with lower accuracy. Utterance boundaries are useful for the development of a noun category, particularly at intermediate levels of frequency sensitivity. Keywords: Word class derivation, independent contexts, frequent frames. Introduction Several authors have shown that distributional statistics can provide powerful cues for acquiring syntactic categories; words that belong to the same syntactic category tend to be preceded and followed by the same words. Thus, nouns tend to be preceded by determiners and adjectives and followed by verbs. Redington, Chater and Finch (1998), building on work by Chater and Finch (1992), investigate several variants of the same basic principle: for any of a set of target words, a context vector was derived that contained (rank orders of) counts of the 150 most frequent words in the corpus, in positions preceding and following the target words. Redington et al. computed correlations between the context vectors of the target words, which were then used as input to a Hierarchical Cluster Analysis, and concluded that the resulting classes mapped closely onto broad syntactic classes. Redington et al. explore a number of variants of the basic mechanism, but get their best results by using a context of one preceding and one following word, and using a rank order correlation as their distance measure. An alternative mechanism for acquiring syntactic categories has been proposed by Mintz (2003). Mintz introduces the notion of a frequent ‘frame’: two lexical items with one word intervening (e.g. He X to). Mintz argues that the (45) most frequent frames in the (English) corpora he analyses show high internal consistency in