Referential Coherence in a Public Speaker Across Fifteen Years

Abstract

This paper presents a preliminary, citizen-conducted analysis of referential chain structure in spontaneous speech by Donald J. Trump across four matched samples spanning 2011 to 2026. Using a visual mapping method, we tracked direct-object referents sentence by sentence, recording hop density (the rate of topic transitions), loop-back frequency (unannounced returns to earlier referents), construction abandonment, and the emergence of dominant anchor topics. Hop density was stable between 2011 and 2016, approximately doubled in the April 2026 sample, and remained elevated in the June 2026 sample. Loop-backs increased from zero (2011) to two (2016) to four (April 2026); the June 2026 sample shows three loop-backs with a notably high density of abandoned mid-sentence constructions — five in approximately 90 seconds, more than any prior sample. Anchor gravity in the June 2026 sample attaches to the same referent (nuclear weapon / Iran nuclear deal) as the April sample, via a largely independent chain of associations. The analysis is preliminary and does not constitute a clinical assessment. It establishes a consistent directional trend that warrants closer examination by researchers with access to larger corpora and formal coding instruments.

Introduction

Why discourse coherence matters — and why it can be studied

Language is the most accessible window into cognition. When a person speaks spontaneously — responding to questions, narrating events, fielding interruptions — the structure of their speech reflects the organisation of their thinking in real time. Researchers in clinical linguistics, neuropsychology, and gerontology have long used connected speech analysis as a sensitive, non-invasive indicator of cognitive change. Unlike formal memory tests, it requires no clinical setting and leaves a permanent public record.

The question addressed here is narrow and methodological: can a consistent directional change in discourse coherence be detected across a fifteen-year sample of one speaker's spontaneous speech? The speaker is Donald J. Trump. The choice reflects the availability of a long, well-documented public record of press conferences and interviews in a consistent format — solo Q&A with journalists — that spans decades and is well-suited to this kind of analysis.

This analysis does not claim to diagnose any condition. It does not assert that observed changes are pathological. It asks only: is there a measurable trend, and if so, what does it look like?

The analysis originated as an observation while watching a press conference on 6 April 2026 — a noticeable pattern of circling back to the same topics — and evolved into a structured comparison across time points. A fourth sample, drawn from a Meet the Press interview recorded in Wisconsin on 6 June 2026, was added to test whether the April patterns were consistent or anomalous.

Background

Tangential speech and referential chaining

All spontaneous adult speech is somewhat tangential. Speakers digress, use illustrative analogies, and follow associative paths away from a main topic before returning. This is normal and well-documented in spoken language corpora. The relevant clinical literature is concerned not with the presence of tangentiality, but with its density, structure, and directionality.

Researchers studying age-related and condition-related changes in discourse have identified several markers of concern when they appear in combination and increase over time. The foundational work here is the Nun Study, begun in 1986 by epidemiologist David Snowdon at the University of Minnesota. Snowdon and colleagues — including psycholinguist Susan Kemper of the University of Kansas — analysed autobiographical essays written by Catholic sisters in young adulthood and found that low "idea density" (the number of discrete propositions per ten words) and low grammatical complexity in early life were strong predictors of Alzheimer's disease and poor cognitive function decades later.^1,2 Kemper's associated work traced language decline across the lifespan using the same cohort, documenting how syntactic complexity and propositional density both diminish with age and cognitive decline.³ More recently, a substantial body of work using natural language processing applied to connected speech has confirmed that linguistic features — including syntactic complexity, lexical diversity, and speech disfluency — can detect early Alzheimer's disease with accuracy rates of 80–88% in controlled studies.^4,5

Increased referential hop density — more rapid switching between topics within a given time window
Loop-back frequency — returning to earlier referents without explicitly signalling the return, suggesting the speaker retains a background awareness of an "anchor" topic even while departing from it
Abandoned constructions — syntactic structures that begin but do not complete, where the speaker pivots mid-sentence
Anchor gravity — a single topic that exerts disproportionate pull, being revisited repeatedly across a passage regardless of what intervening topics were introduced

What makes the loop-back with anchor gravity pattern particularly notable is that it is qualitatively different from ordinary digression. Ordinary digressions are centrifugal — the speaker moves outward from a topic and keeps going. The loop-back pattern is partly centripetal: the speaker moves away but the discourse keeps collapsing back to the same point, often via different associative paths each time.

Methods

Passage selection and coding

Passage selection

Four passages were selected from YouTube-sourced transcripts (auto-generated captions, extracted using yt-dlp). Each passage was drawn from a solo press conference or one-on-one interview in which the speaker was responding directly to a journalist's question, minimising the performance-genre effects associated with rallies or scripted speeches. Each passage runs approximately 90 seconds of spontaneous speech.

2011: Television interview (Piers Morgan Tonight, CNN). Question: "What would you do to fix America?" Passage: approximately 4:22–8:00.
2016: Press conference, Doral, Florida, July 27. Responding to questions about Russian hacking and the DNC email leak. Passage: approximately 5:00–7:00.
April 2026: White House press conference, April 6, 2026 (described by the White House as "President Trump Holds a Press Conference"). Responding to questions about the Obama-era Iran nuclear deal. Passage: approximately 56:58–58:16.
June 2026 New: Meet the Press interview with Kristen Welker, recorded at Custer Farms, Chippewa Falls, Wisconsin, June 6, 2026. Responding to the question "Why didn't you negotiate a better deal [after ripping up the JCPOA]?" Passage: approximately 10:37–12:17.

Coding scheme

Each passage was coded for four variables. Coding was performed by reviewing the transcript clause by clause and identifying each distinct direct-object referent — person, place, object, institution, or abstract concept — as it was introduced or switched to.

Referential hop density: total number of distinct referents introduced or transitioned to across the passage.
Loop-back frequency: number of occasions on which a previously-introduced and -departed-from referent is returned to without explicit flagging.
Maximum anchor returns: highest number of times any single referent is revisited within the passage.
Construction abandonment: mid-sentence structures that begin but do not complete grammatically before the speaker pivots.

Note on the June 2026 transcript

The Meet the Press interview was conducted in a barn during stormy weather, interrupted multiple times by rain hitting the metal roof. An NBC-produced professional transcript was available but was not used for coding, as professional transcription typically cleans up false starts, abandoned constructions, and incomplete utterances — precisely the features being measured. The YouTube auto-generated caption file was used instead, consistent with the method applied to all other samples. The raw captions show a notably high number of mid-sentence pivots in this passage, some of which reflect interruption by the interviewer and some of which appear to be spontaneous.

Limitations

Four passages is still not a corpus. Findings are indicative, not conclusive.
Passage topics differ across time points; topic complexity may independently affect discourse organisation.
Coding was performed by a single analyst with AI assistance, without inter-rater reliability testing.
Auto-generated captions introduce transcription errors that may affect clause boundary identification.
No clinical baseline or formal diagnostic instrument has been applied. This is discourse analysis, not neuropsychological assessment.
The speaker is a practiced performer. Some apparent tangentiality may reflect deliberate rhetorical strategy rather than spontaneous cognitive organisation.
The June 2026 interview was conducted in adverse weather conditions with multiple external interruptions, which may independently affect speech organisation.

Results

Referential chain maps

The four diagrams below map each passage as a directed sequence of topic nodes. Each node represents a distinct referent. Solid arrows indicate forward transitions. Red dashed arrows indicate loop-backs. Teal nodes mark loop-back destinations — topics the speaker has returned to. The sequence number at upper-left of each node records its position in the chain; ↩n indicates which earlier node is being revisited. ⚠ marks an abandoned mid-sentence construction.

New referent

Loop-back destination

Forward hop

Loop-back

⚠ Abandoned construction

Figure 1 — 2011 passage (~4:22–8:00)

Figure 1. 2011 interview. Twelve referential hops in approximately 90 seconds. All forward motion — no topic is revisited once departed. The chain is wide-ranging but directional.

Figure 2 — 2016 passage (~5:00–7:00)

Figure 2. 2016 press conference. Twelve referential hops — same density as 2011. Two loop-backs appear for the first time. Russia is revisited twice via different associative paths. The chain still completes logically.

Figure 3 — April 2026 passage (~56:58–58:16)

Figure 3. April 2026 press conference. Twenty-three referential hops — approximately double the 2011 and 2016 density. Four loop-backs. "Nuclear weapon" revisited four times. Two constructions abandoned mid-sentence. Chain does not resolve to the original question.

Figure 4 — June 2026 passage (~10:37–12:17) New

Figure 4. June 2026 Meet the Press interview. Eighteen referential hops — elevated but below the April 2026 peak. Three loop-backs. Five abandoned constructions — more than the other three samples combined. The nuclear weapon / Iran deal anchor recurs independently from the April 2026 sample, via a different chain of associations. The passage ends with a specific factual claim (21–22% of missiles remaining) rather than dissolving mid-chain, suggesting the interviewer's direct question imposed some closure.

Summary comparison Updated

Measure	2011	2016	Apr 2026	Jun 2026
Referential hop density	12	12	23	18
Loop-back frequency	0	2	4	3
Max returns to single topic	0×	2×	4×	2×
Abandoned constructions	0	0	2	5
Chain resolves logically	Yes	Yes	Partial	Partial
Dominant anchor topic	None	Russia (weak)	Nuclear weapon (strong)	Nuclear / deal (moderate)
Analogies completed	Yes	Yes	No (landlord)	N/A — none attempted

Discussion

What the four-point trend suggests

The most important finding from the 2011 sample remains that this speaker has always been tangential. The baseline chain is wide-ranging but directional — centrifugal, not centripetal. This matters because it means the changes observed in the 2026 samples cannot be attributed to a speaker who was always disorganised.

The 2016 passage introduces loop-backs for the first time, but at moderate density and without construction abandonment. The chain still completes logically.

The April 2026 passage represents a clear departure: doubled hop density, four loop-backs, two abandoned constructions, and a single referent ("nuclear weapon") exerting strong gravitational pull across the entire passage.

The June 2026 passage is instructive in a different way. Hop density is lower than April (18 vs 23), loop-backs are slightly fewer (3 vs 4), and the anchor gravity is less extreme. One reading is that this represents natural session-to-session variation. But a second feature stands out sharply: five abandoned constructions in approximately 90 seconds — more than the other three samples combined. The raw transcript shows repeated false starts and mid-sentence pivots ("Had I not — Had I — I terminated it"; "It was — You know, it expired long ago"; "I'm moving very fast. I'm — I'm into 3 months"). These are the kind of incomplete utterances that professional transcription typically cleans away, which is precisely why the auto-generated caption source is methodologically important here.

The June 2026 sample suggests a pattern that is not simply "more or less tangential" but that varies in which features dominate. In April, the dominant feature was anchor gravity — compulsive return to a single referent. In June, the dominant feature is construction abandonment — sentences that begin and do not complete. Both are markers identified in the discourse coherence literature as associated with cognitive load or decline. Their appearance in different proportions across two sessions two months apart is consistent with a fluctuating rather than fixed pattern, which is itself clinically significant.

A further observation: the June 2026 sample attaches to the same anchor topic (nuclear weapon / Iran deal) as the April sample, but arrives there via a completely independent chain of associations. In April, the path was: Obama chose Iran over Israel → Jewish voters → cash on 757 → Arab world → Iraq → nuclear deal. In June, the path was: Iran pride → 47 years → nuclear weapon close twice → JCPOA terminated → Obama. The destination is the same; the route is entirely different. This is consistent with an anchor topic that has strong independent salience rather than one being rehearsed from memory.

Confounds and alternative explanations

Topic complexity: Both 2026 passages involve the same complex topic (Iran nuclear programme), which may independently drive elevated loop-backs to that referent. The 2011 and 2016 passages involved different topics. This remains the strongest methodological confound.
Weather and interruption: The June 2026 interview was conducted in a barn during a storm, with rain hitting the metal roof causing multiple interruptions and audibility problems. These conditions may independently drive construction abandonment as the speaker loses thread due to external distraction.
Emotional salience: The speaker may have stronger emotional investment in Iran-related topics that produces rhetorical repetition rather than involuntary anchoring.
Deliberate style: Apparent tangentiality may reflect rhetorical strategy. However, the five abandoned constructions in the June sample are difficult to explain as deliberate — a strategic speaker does not typically begin sentences and fail to complete them.
Interviewer effects: Welker's persistent follow-up questions in June may have disrupted the speaker's processing in ways that drove abandonment, rather than the abandonment being spontaneous.

These confounds are real and should be addressed in any more rigorous analysis. The topic confound in particular — both 2026 samples involve the same subject matter — means that the anchor gravity finding cannot yet be cleanly separated from topic-driven repetition. Passages from 2026 on different topics would help isolate this.

Conclusion

A consistent signal across four samples

Across four matched passages spanning fifteen years, a consistent directional trend is visible in three of four measures: hop density increased, loop-back frequency increased, and construction abandonment — absent from both earlier samples — appeared and then intensified. The fourth measure, anchor gravity, is strong in the April 2026 sample and moderate in June.

The speaker's baseline style was always tangential. The 2026 samples are not simply "more of the same." They show qualitatively new features — centripetal rather than centrifugal organisation, and sentences that begin and do not complete — that were absent from the baseline.

The addition of the June 2026 sample strengthens the analysis in one important respect: it shows that the April 2026 findings were not an isolated anomaly. The same cluster of features — elevated hop density, loop-backs, construction abandonment, nuclear weapon anchor — recurs in a second 2026 sample drawn from a different context. The specific profile differs between the two sessions, but the general pattern persists.

This analysis is preliminary. Four passages does not make a corpus. A rigorous analysis would require a larger set of matched passages, formal inter-rater reliability testing, and ideally comparison with a control group of age-matched speakers in similar contexts. Researchers in discourse analysis and clinical linguistics have the tools to do this work. The transcripts are publicly available. The question is legitimate and the method is reproducible.

This text makes no clinical claims. It notes only that a pattern appears to be present, that it is consistent across the sample, and that it points in one direction.

References

Cited works

Snowdon, D.A., Kemper, S.J., Mortimer, J.A., Greiner, L.H., Wekstein, D.R., & Markesbery, W.R. (1996). Linguistic ability in early life and cognitive function and Alzheimer's disease in late life: Findings from the Nun Study. JAMA, 275(7), 528–532.
Snowdon, D.A., Greiner, L.H., Kemper, S.J., Nanayakkara, N., & Mortimer, J.A. (1999). Linguistic ability in early life and longevity: Findings from the Nun Study. In J.-M. Robine et al. (Eds.), The Paradoxes of Longevity. Springer.
Kemper, S., Greiner, L.H., Marquis, J.G., Prenovost, K., & Mitzner, T.L. (2001). Language decline across the life span: Findings from the Nun Study. Psychology and Aging, 16(2), 227–239.
Chou, Y.H., et al. (2024). Screening for early Alzheimer's disease: enhancing diagnosis with linguistic features and biomarkers. Frontiers in Aging Neuroscience, 16, 1451326.
Balabin, L., et al. (2025). Natural language processing-based classification of early Alzheimer's disease from connected speech. Alzheimer's & Dementia. https://doi.org/10.1002/alz.14530

Sources

Transcript sources

2011: "Donald Trump Interview — Piers Morgan Tonight." YouTube. Transcript extracted via yt-dlp from auto-generated captions.
2016: "Donald Trump Press Conference July 27 2016 — Doral, Florida." YouTube. Transcript extracted via yt-dlp from auto-generated captions.
April 2026: "President Trump Holds a Press Conference, Apr. 6, 2026." White House. YouTube. Transcript extracted via yt-dlp from auto-generated captions.
June 2026: "Trump: Meet the Press Full Interview." NBC News / Meet the Press with Kristen Welker, June 8, 2026. YouTube (youtu.be/4EusZcKt5fs). Transcript extracted via yt-dlp from auto-generated captions.

Auto-generated captions contain transcription errors. Passages were reviewed against audio before coding. The June 2026 interview was also available in a professional NBC transcript, which was not used for coding to maintain methodological consistency across samples.