Two Pointers, One Recipient: Linguistic Orientation Where Reply Edges and @-Addresses Diverge in YouTube Comment Trees Max Sedlmair · O
Authors/Creators
Description
In brief. A YouTube reply carries two indications of whom it answers. The platform records a parent comment, and the text often begins with an @-name. Usually they agree. Sometimes they do not — and when they diverge, the reply’s wording tracks the person named in the text, not the person the platform recorded. The divergence is not random: it occurs in 41.5% of replies at the deepest available level and in none at all one level up.
Research on threaded comment data treats the recorded reply edge as the conversational relation. Herring and Chae (2021) distinguish reply relation, addressee relation and content relation conceptually, and code them by hand on 600 comments; the alignment literature on Twitter and Reddit takes the observed reply edge as the relevant interlocutor. What has not been tested is which of two competing, observable recipient signals the surface language follows when they disagree.
This study uses a corpus of 52,865 YouTube comments from 16 videos, retrieved with genuine parent edges to depth 3 under two ranking conditions. Replies whose leading @-name resolves unambiguously to exactly one account in the same thread yield 9,542 analysable cases, of which 7,629 agree with the recorded parent and 1,913 diverge.
The divergence is structurally located. At depth 2 it occurs in 0 of 4,838 cases. At depth 3 — the deepest level YouTube stores — it occurs in 1,886 of 4,548 (41.5%). This is the signature of a display constraint, not of speaker behaviour: the platform stores four levels, and a reply that belongs deeper is attached to the last node with room. In 848 length-matched divergent cases, after deduplication across the two ranking strata, content-word overlap with the @-target exceeds overlap with the recorded parent (paired difference +0.0284, permutation p < 0.001, 63.0% of individual cases). A positive control on 7,289 agreement cases establishes that the measure detects conversational reference rather than shared topic: the parent scores 0.1145 against 0.0550 for a random third comment from the same thread. Function-word profiles show no difference (p = 0.30), locating the effect in contentrather than in style.
The recorded parent is not noise — it also beats the placebo — but where the two pointers disagree, the text follows the name.
Notes (English)
The author holds a Staatsexamen in English philology from the University of Regensburg. He has no institutional affiliation and no formal training in corpus methods, quantitative analysis or computer science. The work was carried out with a large language model (Claude, Anthropic).
The model proposed the paired-competition design; wrote all collection, extraction and analysis code; computed every figure reported here; identified the depth distribution in §4.1; built the positive control and the placebo; and drafted this text, including this section.
The author specified the corpus design and froze it before coding; carried out the collection; set the case-count gate that this analysis had to pass before being written up; rejected two earlier projects on the same corpus when the data did not support them; and made every decision about what this paper would claim.
Errors on record. A first version of this analysis used a different candidate-selection rule for each subgroup and produced an apparent effect in opposite directions. The model built it; the author’s request for the controls exposed it. That version also had an empty positive control — a coding error in which agreement cases were silently discarded — which meant the measure was unvalidated while its output was being interpreted. Both are reported in §3.2 and §4.2 rather than only here. Neither would have been visible in a finished analysis.
How to read this paper. The author can assess the descriptive claims and cannot independently assess the statistical ones. Every number was produced by code he did not write. Stating this is an invitation rather than a disclaimer: this paper is submitted for refutation. The claims most in need of adversarial checking are the zero divergence rate at depth 2 (§4.1), the validity of the similarity measure established by the positive control (§4.2), and the ranking-stratum imbalance (§6). The deposited scripts and result tables exist for that purpose.
Data availability
Analysis scripts, per-case result tables, per-video summaries and the corpus manifest are deposited with this preprint under CC0. Scripts use the Python standard library only. Result tables contain the similarity scores and structural variables but no comment text; comment text is not redistributed, in line with platform terms. Video identifiers and retrieval parameters permit reconstruction of the corpus.