Who Echoes Whom: Propagating Asymmetric Accommodation to Infer Hierarchy Between People Who Never Talk

People in an organization unconsciously copy the small words of those they depend on. Subordinates echo a boss's "the," "we," and "would" more than the boss echoes theirs. This copying, called linguistic accommodation, is one of the best text-based signals of power. It has a hard limit, though: it exists only between people who actually exchange messages. In the Enron email corpus, most pairs in the corporate hierarchy never write to each other, so any method that reads only the text between two people tops out at 59.61% accuracy over all dominance pairs. Plain degree centrality, which just counts how many people someone talks to, reaches 83.88% (A Comprehensive Gold Standard for the Enron Organizational Hierarchy). That paper closed by calling for hybrid models of text and network structure, and it did not build one.

The proposal is to turn each email relationship into a directed piece of evidence about rank, then let that evidence spread through the communication graph. Each edge carries a length-corrected measure of how much more one person echoes the other than the reverse. A model then propagates these local judgments across paths, so that if A defers to B and B defers to C, the model can rank A below C without A and C ever corresponding. Put simply, the claim is that asymmetric accommodation is the edge signal that hybrid text-graph models have been missing. Once that signal is propagated over the graph, it should beat centrality on exactly the non-communicating pairs where text-only methods are mathematically excluded.

Core Claim

Existing hybrid text-graph models treat the text on an edge as topical content. Examples include GLEM, which alternates training a language model and a graph network that teach each other through pseudo-labels (Learning on Large-scale Text-attributed Graphs via Variational Inference), and DyGRASP, which fuses LLM summaries of timestamped edge text with temporal graph networks and is evaluated on the same Enron graph (Global-Recent Semantic Reasoning on Dynamic Text-Attributed Graphs with Large Language Models). Accommodation work, by contrast, shows that power lives in content-poor function words and in the direction of copying (Echoes of power). Topical embeddings largely discard function words, and averaging over edges discards direction.

We claim that a directed, length-normalized accommodation asymmetry on each edge, propagated by a ranking model over the graph, will do three things. It will recover hierarchy on non-communicating pairs above the centrality baseline. It will beat topical text-graph embeddings given the same graph. And its advantage will vanish if the edge signal is replaced by message volume or made symmetric. That last prediction is what makes the claim falsifiable: if volume-weighted edges do as well, the linguistics adds nothing beyond structure.

Research Program

Measuring the edge signal. For every reply pair in the Enron threads, where person $a$ writes and person $b$ replies, we compute coordination on function-word classes in the style of Echoes of power. That paper's measure is the chance that $b$ uses a marker class given that $a$ just used it, minus $b$'s baseline rate for that class. We then correct this for message length. The methodology review in Coordinating Chaos shows that unnormalized measures partly track verbosity, so we subtract the coordination expected under a null in which $b$'s actual reply is swapped for another of $b$'s replies of similar length. Pairs have few exchanges, so we shrink per-pair estimates toward speaker-level means. This is in the spirit of the normalization behind the Zelig Quotient (Finding Zelig in Text). The edge weight is

$$w_{ab} = \bar C(b \to a) - \bar C(a \to b),$$

where $\bar C(b\to a)$ is the shrunk, length-corrected coordination of $b$ toward $a$ averaged over marker classes. A positive $w_{ab}$ means $b$ defers to $a$ more than the reverse.

We deliberately orient the signal around the target of coordination. The reason is that Echoes of power found strong support for people coordinating toward high-status targets. It also found that Wikipedia administrators coordinate more, not less, as speakers. A score that rewards a person for "not echoing" would therefore be miscalibrated, while a score that rewards "being echoed" is supported. The same paper's finding that Supreme Court advocates echo unfavorable Justices most is a warning that dependence, not formal rank, drives the signal. We treat this as a known confound to analyze, not hide.

Propagating it. Each person $i$ gets a scalar status $s_i$, read out from a node embedding $h_i$. That embedding combines two sources. The first is a sentence encoder over the person's sent mail. The second is directed message passing in which incoming and outgoing edges are aggregated by separate encoders, following the dual-direction design of Directed Homophily-Aware Graph Neural Network. Each edge's features are $w_{ab}$, message volume, and an "upward email" score from an n-gram classifier. Training uses gold dominance pairs $(i \succ j)$ among training employees:

$$\mathcal{L} = -\sum_{(i \succ j)} \log \sigma(s_i - s_j) \;-\; \lambda \sum_{(a,b)\in E} |w_{ab}| \log \sigma\big(\mathrm{sign}(w_{ab})(s_a - s_b)\big).$$

Here $\sigma$ is the logistic function and $E$ is the set of communicating pairs. The second term asks the learned ranks to agree with the direction of accommodation, weighted by how strong the asymmetry is, and $\lambda$ sets how much that unlabeled evidence counts. Setting the first term to zero gives a label-free variant. A scalar rank is the smallest geometry that can answer "who is above whom." We do not adopt hyperbolic embeddings in the style of Language Models as Hierarchy Encoders, which encodes taxonomy depth in a Poincaré ball. The evaluation here only asks for the direction of comparable pairs, and that paper itself notes the risk that specializing the geometry degrades general representations.

Splits that avoid leakage. The Enron gold standard is a transitive closure over an org chart, so a random split of pairs leaks rank information. We split by employee instead. Training pairs have both endpoints in the training set of people. We evaluate on pairs with both endpoints held out and on mixed pairs, separately for pairs that communicated directly and pairs that did not.

Baselines

The naive reference point is degree centrality, which ranks each person by how many others they exchange email with. On the full gold-standard graph it reports 83.88% accuracy, and 87.58% on the communicating subgraph (A Comprehensive Gold Standard for the Enron Organizational Hierarchy). It needs no training, so we recompute it on our exact held-out pairs. Beating it on the held-out non-communicating pairs is the bar the proposal must clear, because that is where it claims to add something that structure alone cannot.

The strongest text-only competitor is the n-gram upward-email classifier from the same study, which reports 82.37% on communicating pairs. By construction it cannot exceed the 59.61% ceiling over all pairs. We retrain it on our training employees and report it on communicating held-out pairs only, where it is defined.

The strongest existing hybrid is a topical text-attributed graph model. We run two of them. The first is a GLEM-style co-trained language model and graph network (Learning on Large-scale Text-attributed Graphs via Variational Inference). The second is DyGRASP embeddings from the Enron graph in DTGB. Each gets the same pairwise ranking head and the same number of hyperparameter trials as our model. No reported result was found on Lacuna for either system on Enron hierarchy prediction, so both numbers will be produced here. They matter because they test whether general topical embeddings already absorb status, which would make the accommodation signal unnecessary.

The Bradley–Terry accommodation ranker fits $s$ with no neural network, using only the second term of the loss, optionally with a degree prior. It is the most dangerous baseline because it already propagates transitively. Finally, the ablations isolate the key ingredient. One replaces $w_{ab}$ with message volume. One makes edges symmetric by using $|w_{ab}|$ with no direction. One drops the length correction. The proposal counts as successful only if the full model beats both degree centrality and the topical hybrids on held-out non-communicating pairs, with a confidence interval obtained by bootstrapping over employees, and if the volume and symmetric ablations lose most of that gain.

Why not just…?

Why not just use centrality or PageRank? Centrality is strong where hubs and rank coincide. A Comprehensive Gold Standard for the Enron Organizational Hierarchy reports 93.75% on executive-versus-non-executive pairs, but the non-core network has 0.008% density, where hub counts are thin. In a different setting, the IETF, chairs are marked by facilitating language rather than hubness (Tracing Linguistic Markers of Influence in a Large Online Organisation). We therefore report accuracy stratified by whether a pair lies within the executive core, among non-core employees, or across the two. If our model only matches centrality inside the non-core stratum, the claim fails.

Why not just aggregate the accommodation edges with a simple ranking model and skip the graph network? This is a real threat. A Bradley–Terry fit over $w_{ab}$ already infers rank between people who never wrote to each other. If it matches the neural model, the honest conclusion is that the contribution is the edge signal and its propagation, not the architecture. We would report it that way. The experiment settles this directly by running both on identical edges and splits. The neural variant earns its place only if the text encoder and directional aggregation add measurable accuracy on the non-communicating pairs.

Why not just feed a frozen language model's embeddings of each person's mail into a standard graph network, or prompt an LLM with the history? Topical encoders are trained on content. The status cue is in function-word direction, which pooled embeddings do not represent as an asymmetry between two people. Flattened LLM prompts also lose multi-hop context, and DTGB shows that dynamic text-graph models lean on co-occurrence shortcuts (DTGB). The GLEM-style and DyGRASP baselines test the first form of this alternative head-on. The symmetric-edge ablation tests whether direction, rather than richer text, is what helps.

Proposed Experiments

The primary experiment is the Enron gold standard (1,518 employees, 13,724 dominance pairs) under the employee-disjoint split. We report pairwise accuracy separately for communicating and non-communicating held-out pairs and by organizational stratum. The decisive checks are three. The first is whether the full model beats recomputed degree centrality on non-communicating pairs. The second is whether replacing accommodation asymmetry with message volume removes the gain. The third is whether the label-free variant, trained with the first loss term dropped, already orders people better than degree. If it does, accommodation alone carries rank information through the graph.

A diagnostic test examines dependence versus rank. We look for pairs where the model disagrees with the org chart and check whether the subordinate-looking person depended on the other, for example on deal threads. This mirrors the Supreme Court advocate finding in Echoes of power.

A transfer check applies the label-free variant to Wikipedia talk-page networks and asks whether propagated scores separate administrators from other editors. The reference point there is that a coordination-feature classifier trained on Wikipedia reached 57.1% when transferred to Supreme Court status (Echoes of power). That figure comes from a different task, so it frames the expected difficulty rather than serving as a target.