Cross-situational learning in a Zipfian environment

Both adults and children have shown impressive cross-situational word learning in which they leverage the statistics of word usage across many different scenes in order to isolate specific word meanings (e.g., Yu & Smith, 2007). However, relatively little is known about how this learning scales to real language. Some theoretical analyses suggest that when words follow a Zipfian distribution, as they do in natural language, it should be more difficult to learn a lexicon because of the many low-frequency words that are only observed a few times (Blythe, Smith, & Smith, 2010; Vogt, 2012). Although this effect can be mitigated somewhat by assuming mutual exclusivity (Reisenauer, Smith, & Blythe, 2013), no mathematical analyses suggest that learning in a Zipfian environment should be easier. In this work, we show the opposite of the predicted effect using cross-situational learning experiments with adults: when the distribution of words and meanings is Zipfian, learning is not impaired and is usually improved. Over a series of experiments, we provide evidence that this is because Zipfian distributions help people to disambiguate the meanings of the other words in the situation.

成年人与儿童均展现出令人瞩目的跨情境词汇学习能力，他们能够利用众多不同场景中词汇使用统计规律，以辨识特定的词汇意义（如 Yu & Smith, 2007 所述）。然而，关于此类学习如何扩展至真实语言，所知甚少。一些理论分析表明，当词汇遵循 Zipf 分布，正如自然语言中的情况，由于存在众多仅被观察几次的低频词汇，学习词汇表应更为困难（Blythe, Smith, & Smith, 2010; Vogt, 2012）。尽管通过假设互斥性（Reisenauer, Smith, & Blythe, 2013）可以在一定程度上缓解此效应，但尚无数学分析表明在 Zipf 环境中的学习应当更为容易。在本研究中，我们通过针对成年人的跨情境学习实验，揭示了与预测相反的现象：当词汇与意义的分布呈 Zipf 分布时，学习并未受损，通常反而得到改善。在一系列实验中，我们提供了证据表明，这是由于 Zipf 分布有助于人们区分情境中其他词汇的意义。