📄 Utilization of Dependency Type per Sentence to Identify Differences among Genres of English Texts

Summary

The paper asks whether the way different genres of written English use syntax shows up in simple counts over dependency parses. It relies on type per sentence (TPS), a measure the author proposed in 2016: the number of tokens of a given dependency label in a corpus divided by the number of sentences in that corpus. The author parsed raw written texts from MASC 500K with the Stanford Lexicalized Parser v.3.7.1, which outputs Universal Dependencies labels, after some manual cleanup. A Ruby script then counted each label and computed TPS for eleven subcorpora. TPS profiles differed noticeably by genre. Essays and journals use more adverbial modification, technical reports use more adjectival modification, and fiction uses fewer passive subjects. Plotted across subcorpora, some labels seem to rise and fall together and others move in opposite directions. The author explains these patterns by what writers are trying to do, such as adding detail to nouns or elaborating on clauses. The work matters because it shows that cheap, parser-derived label frequencies can act as an interpretable stylistic and genre signal.

Key points

• Data: the written part of MASC 500K (about 500K words, drawn from OANC). Eleven subcorpora were kept for analysis: Travel, Tech, Nonf, News, Letters, Journal, Govt, Fiction, Ficlets, Essays and Blog. Spoken data, jokes and movie scripts were left out as too conversational, emails and spam because of repeated reply text, and tweets because of HTML tags.
• Pipeline: the Stanford Lexicalized Parser v.3.7.1 produced UD labels, and a custom Ruby script computed the counts and TPS. Table 1 reports 19 main labels, including nmod:of, which has the highest TPS of all the nmod subtypes.
• Sentence length confounds TPS. Ficlet averages about 9.8 words per sentence and Fiction about 12.32, while every other subcorpus is above 18. As a result, Ficlet shows low TPS for many labels.
• Notable values:
  • advmod is highest in Essay (1.139) and Journal (1.209).
  • amod peaks in Tech (2.64).
  • nsubjpass is lowest in Ficlet (0.045) and Fiction (0.051).
• Correlations were judged visually from plots, without statistical tests:
  • Positive: acl with acl:relcl, amod with compound, and amod with det.
  • Negative: dobj with nsubjpass.
  • None: dobj with iobj, and aux with dep.
  • ccomp appears paired with advcl and with xcomp.
• An earlier study (Oya 2016) compared 881 learner sentences from ICLE v2 with 881 sentences from academic abstracts. Academic text had higher TPS for all of its top 20 labels, with case, amod and compound differing the most. nsubj was nearly identical in both.

Interesting ideas

• I1. Normalizing label counts per sentence gives each corpus a compact, readable syntactic fingerprint that can be compared across corpora of any size.
• I2. The Ficlet case shows that per-sentence rates are driven partly by sentence length, so any comparison needs length controlled or reported alongside it.
• I3. Labels that share a function (amod, compound, det all modify nouns) tend to co-vary, so it is useful to group them into functional families.
• I4. Cleaning or dropping subcorpora with reply quoting, HTML or dialogue before counting keeps the statistics from reflecting noise instead of genre.
• I5. The paper proposes TPS profiles as a cue for genre and even author identity.

For Mova

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

• I1 (learn): compute TPS profiles of silver parses on a new domain and compare them to the source treebank, so that a strange spike or dip (e.g., in dep) flags parser drift before the data is used for training.
• I2 (learn): when evaluating parsers on different genres, also report rates per word or bucketed by length, as a negative control against differences that are only about length.
• I3 (frames): measure "noun modification density" or "clausal elaboration" (ccomp, advcl, xcomp) per text and use it to decide how much attributive and embedded-proposition structure the frame builder should expect.
• I4 (tokens): add a pre-tokenizer filter that strips markup and quoted reply chains, and mark conversational spans so later per-genre statistics stay clean.
• I5 (back): check generated or round-tripped text against the TPS profile of the target genre, for example a fable, so the output's syntax is explainably close to the intended style.