📄 We Tweet Like We Talk and Other Interesting Observations: An Analysis of English Communication Modalities

Summary

This 2014 study asks whether newer online channels such as Twitter and IRC chat resemble written prose or spontaneous speech more closely. Zayner compares four English sources: Google Books 1-grams, a Twitter stream sample, logs from two IRC channels, and the Santa Barbara Corpus of Spoken American English (SBCSAE). Because the corpora differ in size by many orders of magnitude, he keeps only each corpus's 200 most frequent words and renormalizes within that set. These 200 words cover more than half of all tokens and roughly follow a power-law (Zipf-like) curve. He then compares the corpora on frequency curves, overlap of the top-200 lists, and part-of-speech make-up. Speech turns out to be the most "central" modality, and tweets and chat sit much closer to it than to book text. Fiction behaves like the conversational modes, while nonfiction looks like the Google 1-grams. That suggests the Books corpus is dominated by nonfiction. A case study of Cory Doctorow's talks, tweets, novels and essays confirms the pattern. The work shows that crude vocabulary statistics alone can separate registers.

Key points

• Corpus sizes: Google 1-grams (2009 release) has about 271 billion tokens and about 3.7M types. Twitter has more than 1.25M unique tweets and about 12M tokens. IRC chat has about 9M tokens and about 155k types. SBCSAE is the smallest, with about 250k tokens and about 11k types.
• How quickly the top words accumulate differs by source:
  • 20% of all tokens is reached by word 5 in the 1-grams, word 8 in Talking, word 11 in IRC and word 13 in Twitter.
  • Talking reaches 50% at about word 55, while the other corpora need 100–130 words.
• The normalized top-200 frequencies fit a power law with parameters a=6 and b=1.6. Diversity is scored as RMSD from a perfectly flat (uniform) distribution.
• Doctorow's fiction matches about 77% of a multi-author fiction top-200 list. His essays and blog writing overlap poorly with everything, including book nonfiction.
• By part of speech, the 1-grams and nonfiction lean toward nouns, while fiction, chat, tweets and speech use more verbs.

Interesting ideas

• I1. Rather than comparing whole corpora of wildly unequal size, the author compares only the high-frequency core and renormalizes within it, which makes small and huge corpora directly comparable.
• I2. Since the paper finds that tweets and chat pattern like transcribed speech rather than like books, conversational data is the more natural reference point for online text.
• I3. Cheap cleaning heuristics remove noise: tokens containing a letter repeated three or more times are dropped as OCR junk, and IRC lines lose their "<name>:" addressee prefix.
• I4. A simple noun-versus-verb balance separates fiction from nonfiction without any syntax, which gives a fast register signal.
• I5. When a fitted parameter contradicted the plotted curve, the author cross-checked it with a plainer statistic (a linear slope over the tail). He also tested the population-level claims on a single author as a sanity check.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

• I1 (learn): Silver training sets and evaluation texts could be checked for register mismatch by comparing their renormalized top-N lemma profiles before trusting a score.
• I2 (tree): When parsing tweets or forum text, spoken or conversational UD treebanks may be a better training or adaptation source than newswire.
• I3 (tokens): The en tokenizer could tag such prefixes as an addressee field instead of words, and normalize letter elongations rather than discarding them, because in tweets they are often genuine.
• I4 (world): Verb-heavy narrative input such as fables could trigger stricter event-ordering and state-tracking, while noun-heavy expository text could favour entity and attribute extraction.
• I5 (learn): Evaluation could pair every aggregate metric with a simpler independent measure and a single-source control set, to catch artefacts of fitting or pooling.