Indeed, it is possible that very frequent words such as articles or prepositions generate holistic processing without a repetition, but because of the weak number of those stimuli and their great heterogeneity, ERPs technique doesn't allow us to test such a hypothesis.
Finally, it is possible that whole-form and combinatorial processing are recruited to different extents in the perception of opaque compared to transparent, and frequent compared to infrequent words, as would be predicted by dual-route models (Baayen et al., 1997; Isel et al., 2003; Koester et al., 2004, 2007, 2009; Sandra, 1990; Zwitserlood, 1994).
For the reduced-hybrid H-R2 where frequent words are represented as word units while infrequent words are represented as PM units, a hybrid n-best list illustrated in Fig. 8 c is used in the second-pass re-scoring.
The top 30K frequent words are selected as the vocabulary.
The top 64,000 most frequent words are selected as the vocabulary and other words are mapped to a special token.
The Findex system provides an overview of the results by presenting a list of the most frequent words and phrases as result categories next to the actual results.
To create the word cloud the top 100 most frequent words were extracted (as marked with the vertical line in the frequency plot).
In the first variation (H-R1), the vocabulary size is reduced to be equal to the size of the word-based RNNLM vocabulary (N) by including only frequent words and PMs as shown in Fig. 6 a.
Since the input text of H-R2 is a mix of words and PMs, the top-N" most frequent words are determined first and kept as word units in the vocabulary.
Accordingly, they will be sensitive and prone to pick up the regularities they hear more often, such as frequent words, sounds, inflections and grammar constructions (Saffran et alia 1996; De Villiers 1985).
