Obama vs. Romney — First 2012 Presidential Debate

Framework provenance: the coding scheme set out in the doctoral thesis 'Pragmatic & Paralinguistic Features of Televised Political Discourse in a Political Debate' by Hussain Al Sharoufi was originally developed and applied manually between 1997 and 2001, before automated discourse-analysis tools or large language models were available; the present implementation digitises the same analytical scheme without introducing theoretical categories absent from the original framework. Automatic pipeline: transcription and diarisation, turn segmentation, coding of the five framework domains (cohesion & coherence, pragmatics, paralinguistics, ideology & bipartisanship, culturemic & pragmemic/ADAR), the Lexical Cohesive Trio layer (Al Sharoufi 2014, building on Hoey 1991: reference and transitional signals, patterns of lexical repetition, academic lexical phrases, and central utterances ranked by repetition density), computed repetition and gaze instruments, and a probabilistic Debate Analyser evaluation. Every record remains editable and the instruments recompute from corrections.

Status: preliminary, machine-derived descriptive case study of one recorded debate. Figures describe automated coding of this recording; they are not validated measures of debate performance. The coded corpus contains 960 observations across 145 turns, with complete gaze coverage. Paralinguistics dominates the annotation set (724 observations; 75.9% of Obama's and 73.3% of Romney's counted records), so raw totals across domains are not compared: Romney has 363 counted records and Obama 253, but no overall total is treated as a finding. Romney has more cohesion/coherence observations (52 vs. 35), repetition links (82 vs. 67, each link counted once on the repeating speaker's turn), pragmatics observations (21 vs. 7), and culturemic/pragmemic observations (20 vs. 14). Obama has a slight raw lead in ideology/bipartisanship observations (5 vs. 4). Duration-weighted gaze (the primary gaze measure) gives Romney the higher q-gaze share of his own speaking time (85.5% vs. 81.0%), while Obama has more absolute q-gaze time because he held the floor longer (2043.3 vs. 1912.1 seconds). These figures show higher coded rates for Romney on several measures in this recording; they are exploratory, unvalidated by blinded double-coding, and do not establish the quality or truth of either candidate's claims. No post-debate polling figures were supplied, so agreement with independently documented polling cannot be tested from this record. Cohesion & coherence Romney records 52 cohesion/coherence observations and 82 repetition links, compared with Obama's 35 observations and 67 links. Per 1,000 words, repetition links remain somewhat more frequent for Romney (10.3 per 1,000 words) than for Obama (9.0). Across the whole corpus, simple lexical repetition is the predominant coded type (124 instances), followed by phrasal repetition (52) and complex phrasal/propositional repetition (17). Under Hoey bonding at a 3-link cut-off there are 9 bonds and no central utterances (see the bonding matrix below). The most repeated items concern taxes, deficits, Medicare, plans and the middle class; coded intertextuality totals are not provided. Pragmatics Romney has 21 pragmatic observations against Obama's 7. The record assigns Romney one maxim breach, specifically the corpus's only coded breach of relation, while Obama has none. A larger pragmatic count is not intrinsically positive because it can include contestation or problematic conduct; nevertheless, Romney's higher count indicates more coded pragmatic activity. Both candidates repeatedly use direct rebuttal, numbered argument structure, correction, contrast, and audience address in the supplied utterances. Speaker-level coding for implicature, explicit-to-implicit movement, institutionalisation, and code-switching is absent. Paralinguistics Gaze coverage is complete at 100% of turns. Romney holds q-gaze for 85.5% of his own 2235.3 seconds of speaking time, compared with Obama's 81.0% of 2523.1 seconds. Romney therefore has the higher duration-weighted q-gaze share, with 14.5% a-gaze versus Obama's 19.0%. Obama leads in absolute q-gaze duration, 2043.3 seconds to 1912.1 seconds, because his speaking time is longer. The observable descriptions characterize both candidates as using sustained forward orientation and rhythmic hand gestures, with Romney frequently directing gaze toward the moderator or opponent and Obama combining steady forward gaze with brief lectern glances. The share of gaze records (q-gaze records ÷ all gaze records, about 50% for each candidate) is a different, secondary measure that ignores how long each gaze lasts; it is not a duration proportion and is labelled separately in the dossier. Ideology & bipartisanship This is the sparsest candidate domain: Obama has 5 observations and Romney 4. The supplied utterances contain bipartisan and partisan framing, including appeals to cross-party ideas, cooperation, and contrasts between public and private provision. The near-even and very small counts do not support a strong comparative judgment about ideological framing, media commodification, or bipartisan performance. Culturemic & pragmemic (ADAR) Romney has 20 culturemic/pragmemic observations and Obama 14. Both candidates orient discourse toward culturally salient constituencies and values through references to the middle class, seniors, families, work, future generations, fairness, choice, responsibility, and personal narratives. Obama's explicitly coded audience-orientation example addresses viewers of a specified age about a policy consequence. Romney's higher count indicates more frequent coded cultural-pragmemic activation in this recording, but only three aggregate lexical-field entries are supplied and they are not allocated by speaker. Lexical Cohesive Trio (Al Sharoufi 2014, building on Hoey 1991) Element 1 — reference and transitional signals; element 2 — patterns of lexical repetition; element 3 — academic lexical phrases. Counts below use accepted records only. Trio profile by speaker: - Jim Lehrer: reference & transitional signals 23, repetition links 72, lexical phrases 0 - Barack Obama: reference & transitional signals 21, repetition links 67, lexical phrases 2 - Mitt Romney: reference & transitional signals 31, repetition links 82, lexical phrases 1 Repetition patterns (Hoey taxonomy): - Simple lexical: 124 - Phrasal repetition: 52 - Complex phrasal propositional: 17 - Co-reference: 10 - Complex lexical: 9 - Propositional: 4 - Simple paraphrase: 4 - Superordinate: 1 Lexical-phrase functions: - Reporting results: 2 - Discussing findings: 1 Bonding matrix (Hoey cut-off: 3 links form a bond): - Matrix cells: 10440; empty 98.29%, one-link 1.42%, bonding 0.09%. - Bonds: 9. Central utterances: 0. Marginal utterances (co-ordinates 0,0): 129. Central utterances by bonding co-ordinates (bonds with earlier, later utterances): - No utterance reached the 3-bond cut-off in this corpus. Turn-onset and turn-offset gaze (share of the speaker's observed turns opened or closed in eye-to-eye gaze): - Jim Lehrer: opens eye-to-eye 86.0% (49/57), closes eye-to-eye 93.0% (53/57) - Mitt Romney: opens eye-to-eye 98.0% (49/50), closes eye-to-eye 96.0% (48/50) - Barack Obama: opens eye-to-eye 84.2% (32/38), closes eye-to-eye 89.5% (34/38) Relational lexical web (cohesion) Each participant's repeated lexical items are read as a network: an item is a node, and two items are joined where they meet in the same turn. The web makes visible the fabric a speaker weaves with repetition — where a speaker returns to the same small company of words, the web tightens into a dense knot; where repetitions never meet, it stays loose. Density is realised joins over possible joins; mean joins per item is the average company each word keeps; the largest cluster is the share of items bound into one connected web rather than scattered into isolated threads. Only counted repetition links feed the web, so every node and join can be traced back to the transcript. - Jim Lehrer: 38 repeated items, 78 joins, density 11.1%, mean joins per item 4.1, largest connected cluster 74% of items. Most repeated: Medicare (7), economy (5), governor (5), federal government (4), support (4). - Barack Obama: 52 repeated items, 172 joins, density 13.0%, mean joins per item 6.6, largest connected cluster 83% of items. Most repeated: education (4), Governor Romney (4), role (4), economy / economic (3), make sure (2). - Mitt Romney: 51 repeated items, 123 joins, density 9.6%, mean joins per item 4.8, largest connected cluster 33% of items. Most repeated: Medicare (13), regulation (6), Simpson-Bowles (4), economy (3), role of government (3). Rhetorical moves within the pragmatic domain: none counted in this corpus. Reliability, coverage and exclusions Records held: 960. Counted in every figure: 960 (100%). Accepted means the researcher confirmed the machine-proposed record as coded; qualified means confirmed with an amendment. Machine-proposed: 960. Awaiting review: 0. Rejected on review: 0. Acceptance rate of the automatic pass: 100% of reviewed records were accepted or qualified. A 100% acceptance rate is a workflow status, not evidence of validity. Coverage: 145 of 145 turns carry a counted observation (100%); 145 carry a gaze observation (100%). Exclusions: 0 counted records are not attached to a turn, 0 are not attributed to a participant, and 724 carry no verbatim evidence quote (these are gaze records, anchored to a timecode rather than a quotation). Inter-rater reliability: no blinded double-coding has yet been recorded for this study, so no Cohen's kappa can be reported, and no machine-versus-human precision or recall has been estimated. Until that validation exists, every comparative reading in this report is exploratory.

Status: preliminary, machine-derived descriptive case study of one recorded debate. In one automatically coded recording of the first 2012 Obama–Romney debate, the system produced different counts and normalised rates across its predefined discourse and gaze categories. Those counts are descriptive outputs of this implementation and require independent coding validation and audit before they can be interpreted as stable differences in the candidates' discourse. The defensible output is the five-domain profile set out in the statistical dossier below — verbal domains per 1,000 words, gaze per minute of speaking time — together with exploratory tests that are Holm-corrected across the declared family and an order-aware block bootstrap that respects the sequence of the debate. No overall density total is reported: the domains are heterogeneous and the gaze layer makes up about three-quarters of each candidate's records. This is not a statement of who won the debate, a ranking of candidates, a validated quality measure, or a prediction of an election.

Single-debate design: the data are one fixed recording with 38 Obama and 50 Romney turns. Adjacent turns answer the same question, opponent and moderator, so they are serially dependent; every p-value is exploratory and conditional on this recording. Multiplicity: p-values are reported nominally and Holm-adjusted across the declared family of candidate tests. Nominal results such as p = .028 or p = .042 do not survive adjustment and are hypothesis-generating only. Validation: no blinded double-coding, per-category agreement, or machine-versus-human precision/recall is yet reported; automated coding of implicature, maxims, ideology, culturemic patterns and gaze is unvalidated. Role contrasts that include the moderator describe institutional roles and are not used in the candidate comparison. Generalisation needs multiple debates, speakers, moderators, formats and topics. Transcription, turn segmentation, and framework coding in this report were produced automatically and are marked as machine-derived, each carrying a confidence score. Gaze and speech-state readings are visual estimates sampled from the recording, not manual frame-by-frame coding; treat low-confidence gaze records as indicative only. No independently documented post-debate polling figures are supplied, so the research question's polling-agreement component cannot be evaluated without introducing prohibited outside evidence. Speaker-resolved repetition types are not supplied; under Hoey bonding no utterance reached the central cut-off, so there is no central-utterance matrix to report. Intertextuality, institutionalisation, implicature, explicit-to-implicit movement, and code-switching lack speaker-level totals. The three lexical-field entries are aggregate and too sparse for a robust ADAR lexical-field comparison. Ideology/bipartisanship has only nine candidate observations, limiting comparative inference in that domain. Speech-state figures are counts rather than duration totals, so fluent, hesitant, and phrase-pausing gaze cannot be aggregated into reliable duration-weighted speaker proportions. This is a description of counted framework evidence within the coded corpus, not a debate verdict or an election prediction. Raw observation totals reflect annotation opportunities, speaking patterns, and coding density as well as discursive behavior; higher counts are not automatically better. Policy merit, factual correctness beyond the coded maxim record, party preference, polling, and real-world outcomes are outside the permitted evidence base. The moderator is analyzed as part of the institutional corpus but excluded from the candidate comparison.