📄 Syntactic Patterns in a Sample of Technical English

Summary

Streeter asks how to tell an important frequency difference between texts from one that chance alone could produce. The question matters for statistical studies of style and authorship. He argues that variation inside one text must be understood before comparisons between texts can be trusted. The material, hand-coded into syntactic categories, comes from two linguistics books: all 8 chapters of Introduction to Transformational Grammars by Emmon Bach, and three chapters of the work by Kenneth Pike titled Language in Relation to a Unified Theory of the Structure of Human Behavior. Clauses were sorted into five types (be, active, passive, there, it). Sentences were coded as sequences of main, subordinate and embedded clauses. Chi-square tests then checked whether the chapters of each book behave like samples from one population.

The main conclusion is that homogeneity is not an absolute property of a text. It also depends on sample size and significance level, because any real difference becomes "significant" once the sample is large enough. Streeter therefore recommends relative consistency measures, Mosteller's index of contextuality and a "rejection size", for ranking how stable each feature is.

Key points

• Data: Bach (1964, whole book except the chapter exercises) and Pike (pp. 25–82, without bibliographies). Both come from one discipline, so differences can be credited to the writers rather than to the subject.
• Coding was manual, with no conversion program. Main clauses are marked "M", embedded clauses go in parentheses, and clause types are numbered (e.g. M4(3) is a main transitive clause containing an embedded be clause). Statistics were run on the Michigan Terminal System.
• Mosteller's index of contextuality is chi-square divided by sample size, times 1000. In the Brown corpus the least context-sensitive words were to, and and the, with indexes of 6.2, 6.9 and 9.6.
• Rejection size is the sample size needed to reject the null hypothesis at .001. Unlike the contextuality index, it takes degrees of freedom into account, and a higher value means more uniformity.
• Table 6 compares Bach and Pike. The two writers agree closely on the main/subordinate ratio for passive clauses (0.65) but differ strongly for there clauses (29.15). Length features such as paragraph length in sentences (98.4) vary the most.

Interesting ideas

• I1. Scaling chi-square to a fixed notional sample of 1000 units turns a yes/no significance verdict into a ranking of features by how much context moves them.
• I2. The "rejection size" index asks how much data is needed before a difference shows up, which makes consistency concrete and accounts for degrees of freedom.
• I3. Before claiming two texts differ, measure how much chapters of one text differ from each other. That internal variation is the baseline.
• I4. A coarse clause typology (be, active, passive, existential there, extraposed it) with M-marking and parenthesised embedding gives a compact, readable signature of sentence structure.
• I5. Picking both samples from one discipline keeps subject matter from being mistaken for authorial style.

For Mova

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

• I1 (learn): Rank deprel and POS distributions by their contextuality across treebank sections. This shows which parser statistics are stable and which drift by genre.
• I2 (learn): When comparing two parser versions or silver-data witnesses, report how many sentences it takes for their difference to become detectable. A raw p-value alone says little.
• I3 (learn): Use segment splits of the same corpus as a negative control, so cross-genre or cross-author score gaps are only trusted if they exceed this within-corpus spread.
• I4 (tree): Derive this signature from UD trees (cop, nsubj:pass, expl, ccomp/acl nesting) as an explainable summary of a sentence. It could also serve as a quick sanity check on parser output distributions.
• I5 (answer): When building test sets for readers, hold the topic fixed while varying the author or style. Accuracy changes can then be attributed to syntax rather than to content.