SAID: Style Transfer for Dialogue Rewriting

• SAID is a style transfer task that rephrases multi-party dialogues into structured third-person narratives by resolving speaker references and standardizing informal language.
• It employs coupled transformations including coreference resolution, syntactic rewriting, formalization, and emotion attribution to maintain content and tone.
• This approach improves extractive summarization performance by converting conversational inputs into self-contained narrative statements suitable for news-style systems.

Searching arXiv for the specified SAID paper and closely related dialogue perspective-shift work. SAID, in the sense introduced in "He Said, She Said: Style Transfer for Shifting the Perspective of Dialogues" (Bertsch et al., 2022), is a style transfer task that reframes a dialogue from informal first person to a formal third-person rephrasing of the text. It converts multi-party conversational utterances into line-by-line narrative statements that preserve content and tone while resolving speaker references, standardizing informal language, and making affect explicit. The task is defined at the utterance level and also in a conversation-level variant, and was proposed as a preprocessing transformation that makes dialogue more compatible with extractive news-style summarization systems (Bertsch et al., 2022).

1. Definition and formalization

SAID denotes shifting dialogues into “he said, she said” third-person accounts. The input is a multi-party dialogue $D$ consisting of $T$ utterances, $D = \{u_1, \ldots, u_T\}$, where each utterance $u_t$ is produced by a speaker $s(u_t)$. The target is a third-person narrative $T = \{y_1, \ldots, y_T\}$ aligned at the utterance level (Bertsch et al., 2022).

The line-by-line perspective shift is modeled as a function $f$ that takes the full conversation and a target index $t$ and produces a single third-person rephrasing:

$y_t = f(D, t).$

A conversation-level variant directly produces the entire perspective-shifted conversation:

$Y = g(D) = \{y_1, \ldots, y_T\}.$

The formulation is explicitly not summarization. Annotation guidelines require each $T$0 to include the speaker’s name, remove first-person pronouns, convert questions to descriptive statements, and interpret emoticons or slang into formal descriptors when possible. The goal is a standardized, faithful rephrasing of each utterance with tone preserved rather than condensation of content (Bertsch et al., 2022).

This definition places SAID at the intersection of style transfer, dialogue understanding, and controlled rewriting. A plausible implication is that the task is structurally harder than conventional formality transfer because it couples entity resolution, syntactic transformation, and pragmatic interpretation within each rewritten turn.

2. Core linguistic operations

The operation $T$1 requires four coupled transformations: coreference resolution, syntactic rewriting, formalization, and emotion or stance attribution (Bertsch et al., 2022).

Coreference resolution converts first- and second-person pronouns and other referring expressions into explicit entities, preferably speaker names. First-person singular typically maps to $T$2, but second-person and first-person plural often depend on dialogue context or cataphora. Ambiguous “we” is resolved to “<speaker> and others” when necessary.

Syntactic rewriting transforms first- and second-person syntax into third person. The paper gives the example “Sam: I am busy” becoming “Sam is busy.” This rewriting is not merely lexical substitution; it must preserve predicate structure while changing person, agreement, and narrative framing.

Formalization standardizes grammar and punctuation and removes or reinterprets slang, emojis, and unconventional punctuation. The paper emphasizes that this increases edit distance substantially and differs from conventional formality corpora, which often involve relatively small edits.

Emotion or stance attribution preserves informal tonal markers by encoding them as explicit descriptors, adverbs, or short sentences. “Sam: lol that’s wild!!!” may become “Sam is amused.” or “Sam says, jokingly.” This operation is central because pure normalization would otherwise erase stance information (Bertsch et al., 2022).

These components are tightly interdependent. Formalization without coreference resolution can leave unresolved “I” or “you,” while pronoun replacement without affect modeling can flatten the utterance’s interpersonal force. This suggests that SAID is best understood as discourse-sensitive rewriting rather than simple stylistic normalization.

3. Dataset, annotation, and descriptive statistics

The perspective-shift dataset is a newly annotated subset of SAMSum, described in the paper as a synthetic text-message conversation corpus. From the original SAMSum splits, 314 train, 368 validation, and 151 test conversations are used; 151 SAMSum test conversations are reserved for testing, and the remaining 682 are used for training and validation. Conversations range from 3–30 lines and 2–20 speakers (Bertsch et al., 2022).

Turns are preserved under annotation. Mean words per turn increase from 8.4 in the original dialogues to 11.0 in the shifted version. The average word-wise edit distance between an original and shifted utterance is 8.5 words. Emojis are removed or reinterpreted at an average rate of 0.1 per utterance, while pronoun resolution edits average 0.9 per utterance. The part-of-speech distribution also shifts: interjections decrease by 65.8%; adjectives and adverbs slightly decrease overall by 5.1%, but in utterances with emojis, adjectives and adverbs increase by 12.8% due to explicit emotion descriptors (Bertsch et al., 2022).

Because of SAMSum’s licensing, the annotated perspective-shift dataset cannot be released, although code was publicly released. This limits exact dataset replication while leaving the modeling and training pipeline reproducible at the code level.

The descriptive statistics are important because they show that SAID is not a light-touch rewrite task. The increase in average turn length and the large edit distance indicate substantial structural intervention, especially in utterances with informal affect markers.

4. Modeling approaches and comparative performance

The paper evaluates several baseline approaches to SAID, centered on BART-large fine-tuning and heuristic or formality-transfer baselines (Bertsch et al., 2022).

Four BART-large input formulations are studied: no context, left context only, left and right context, and conversation-level generation. The best-performing configuration is the left-and-right-context line-by-line model. It achieves ROUGE-1/2/L of 63.57/40.74/62.04 and a BARTScore of -2.36. Human judges preferred it over the no-context model 86% of the time, with Cohen’s $T$3 (Bertsch et al., 2022).

A rule-based heuristic prepends “<speaker> says,” replaces “I” with the speaker name, and appends a period to each utterance. It is efficient but ignores complex coreference and emotion attribution. It obtains 61.77/35.93/55.34 on ROUGE-1/2/L with a BARTScore of -2.80, which appears superficially competitive on overlap metrics but is described as weak overall because of hallucinations and unhandled pronouns (Bertsch et al., 2022).

The paper also investigates whether generic formality transfer helps. Sequential fine-tuning on GYAFC and then on perspective shift (“FORMALITY + PS”) yields 62.00/39.14/60.38 and -2.37, slightly worse than direct perspective-shift training. “FORMALITY ONLY” performs poorly in zero-shot transfer with 51.25/22.12/49.96 and -2.57. The stated explanation is that formalization biases toward small edits, whereas perspective shift requires larger pronoun and emotion reinterpretations (Bertsch et al., 2022).

The conversation-level model produces 63.20/35.04/51.80 with a BARTScore of -2.67. Although its ROUGE-1 is competitive, it is described as more abstractive, with weaker alignment and variable line count. This exposes a central modeling tension: unrestricted generation can improve narrative fluency while degrading utterance-level correspondence.

Approach	ROUGE-1/2/L	BARTScore
No context	62.57 / 40.45 / 61.41	-2.38
Left context only	60.80 / 37.50 / 59.27	-2.39
Left and right context	63.57 / 40.74 / 62.04	-2.36
Conversation-level	63.20 / 35.04 / 51.80	-2.67

A plausible implication is that the full dialogue context is useful not because the target is globally abstractive, but because local utterance rewriting depends on future and past mentions for resolving “you,” “we,” and late-introduced named entities.

5. Downstream summarization effects

A principal contribution of the work is the claim that applying SAID before summarization substantially improves the performance of extractive news-style summarizers on dialogue data (Bertsch et al., 2022).

In oracle extractive summarization, the paper defines

$T$4

Under this protocol, a longest-3 baseline reaches ROUGE-1/2/L of 32.46/10.27/29.92. Oracle extraction over original SAMSum reaches 45.89/16.35/34.80, whereas oracle extraction over perspective-shifted SAMSum reaches 50.63/21.40/39.11, approaching abstractive BART-large at 52.86/28.58/43.73 (Bertsch et al., 2022).

For zero-shot extractive summarization using PreSumm trained on CNN/DailyMail, the original SAMSum dialogues yield 35.00/12.09/30.76, while SAID-transformed dialogues yield 37.12/13.14/31.49. For supervised extractive summarization, training on original SAMSum gives 32.19/9.86/28.52, whereas training on perspective-shifted SAMSum gives 39.58/15.03/33.94 (Bertsch et al., 2022).

The paper also reports better faithfulness and fluency for extractive summaries derived from SAID outputs. On 100 sampled summaries, extractive over PS shows 5% hallucinations and 3% contradictions, compared with 22% hallucinations and 18% contradictions for abstractive BART. GPT-2 perplexity is 31.07 for extractive summaries from shifted dialogues versus 48.77 for extractions from original dialogues (Bertsch et al., 2022).

These findings support the paper’s central argument that perspective-shifted utterances are more suitable extraction units. They are self-contained, third-person, and entity-resolved, which better matches the assumptions of extractive news summarizers trained outside the dialogue domain.

6. Examples, failure modes, and practical significance

The paper’s canonical example begins with:

Laura: I need a new printer :/ Jamie: you're sure you need a new one? Jamie: I mean you can buy a second hand one

and yields:

Laura is frustrated that she needs a new printer. Jamie asks if Laura is sure she needs a new one. Jamie clarifies that Laura could buy a secondhand printer. (Bertsch et al., 2022)

Here, “:/” is rendered as an affective statement, pronouns are resolved to speaker names, and questions become declarative statements. A second example shows the risks of insufficient context: “Igor tells John that he has so much to do at work and he is so demotivated” may hallucinate the addressee if the context model incorrectly infers “John” (Bertsch et al., 2022).

The paper identifies several failure modes. Coreference errors and named entity hallucination are especially common in no-context settings or in early cataphoric turns. Emotion misattribution remains difficult because interpretation of informal markers is subjective. The transformation can also lose discourse nuance by linearizing dialogue into third-person statements and discarding interactive structure such as repairs, overlaps, and speaker relations. Generalization is limited by training on short text-message turns; long media interview segments, multi-word speaker names, and domain-specific formatting degrade performance (Bertsch et al., 2022).

Practical implications are broader than summarization alone. The paper states that SAID can transform chat transcripts, support tickets, interviews, and meeting minutes into coherent third-person accounts, facilitating downstream summarization, auditing, and archival. It may also help information extraction by clarifying who did what, when, and with what affect. The authors additionally argue that SAID is easier to annotate than full summarization, and that a perspective-shift model trained on 545 dialogues improves models more than a 20× larger summarization annotation effort would be required to match (Bertsch et al., 2022).

A plausible implication is that SAID functions as a style-and-structure adapter between conversational corpora and downstream systems trained on expository prose. Its value lies less in replacing summarization than in converting dialogue into a representation that established summarization and IE pipelines can already process effectively.

7. Position within dialogue modeling research

Within the paper’s framing, SAID defines a new style transfer task rather than a new summarization objective. That distinction matters because the transformation preserves utterance-level alignment and tone, whereas dialogue summarization typically performs compression and abstraction (Bertsch et al., 2022).

The work also highlights a common misconception: that dialogue preprocessing into formal prose is equivalent to generic formality transfer. The paper’s experiments show otherwise. Models trained on GYAFC underperform because perspective shift demands larger structural and semantic edits, particularly explicit coreference resolution and reinterpretation of informal affect (Bertsch et al., 2022).

Another misconception is that extractive summarization is inherently ill-suited to dialogue. The results suggest a narrower claim: extraction from raw dialogue is poor because turns are interdependent, speaker-centric, and often linguistically incomplete, but extraction from perspective-shifted dialogue becomes markedly more viable. This suggests that some of the apparent mismatch between dialogue and extractive summarization is representational rather than intrinsic.

In this sense, SAID occupies a distinctive place in dialogue processing. It is a rewriting layer that exposes entities, stance, and proposition structure in a format more compatible with narrative-style NLP systems. Its limitations—subjective affect interpretation, possible flattening of interactional nuance, and domain dependence—also define the boundaries of that representation.