Jethro's Braindump

Beam Search Promotes Uniform Information Density

Using beam search for text generation shockingly produces text better than using exact search, despite the high search error rate. Why is this so?

Beam search has an inductive bias that promotes uniform information density (UID). This bias is paramount in producing text that humans prefer. There is also a strong relationship between UID and metrics like BLEU.

The uniform information density hypothesis has roots in cognitive science, and it states:

Within the bounds defined by grammar, speakers prefer utterances that distribute information uniformly across the signal (information density). Where speakers have a choice between several variants to encode their message, they prefer the variant with more uniform information density (ceteris paribus)

That is, humans prefer sentences that distribution their information more uniformly.

Beam search can then be formulated as regularized decoding. Their experiments show that encouraging UID allieviates the text degradation that occurs when using high beam widths.