Dialogue Act Script (DAS) for Multilingual Dialogue
- Dialogue Act Script (DAS) is a structured, intent-based intermediate representation that abstracts dialogue into functional scripts to preserve communicative intent across languages.
- It encodes turn-level speaker intents using function-call-like expressions with dialogue act labels and minimal semantic parameters to enable cultural adaptation.
- Empirical evaluations show DAS mitigates anglocentric bias, enhances meaning preservation and cultural relevance, and offers a robust alternative to direct translation.
Dialogue Act Script (DAS) is a structured, intent-based intermediate representation for multilingual dialogue generation and localization. Rather than translating dialogue utterances directly, DAS first abstracts each turn into a functional representation of what the speaker is doing, then localizes that abstract script, and only afterward realizes a new dialogue in the target language. In this formulation, DAS “represents communicative intent using a predefined set of dialogue acts and parameters,” with the goal of mitigating anglocentric bias and translationese while preserving conversational purpose across languages (Vasselli et al., 26 Sep 2025).
1. Conceptual basis and intellectual context
DAS belongs to the broader research tradition that treats dialogue acts as a general representation of speakers’ intents rather than as surface strings. In dialogue-act research, this representation has been used to describe open-domain conversational functions, task-oriented intents, feedback, and social obligations, often through standardized or reduced taxonomies designed to align heterogeneous corpora (Mezza et al., 2018). Related task-oriented work has likewise proposed universal act inventories so that annotated datasets with incompatible schemas can be mapped into a shared action layer for training and transfer (Paul et al., 2019).
A second neighboring line of work concerns the unit to which dialogue acts attach. Vietnamese dialogue-act segmentation research, following ISO 24617-2, defines a functional segment as the “minimal stretch of communicative behavior that has a communicative function,” and shows that one turn may contain multiple act-bearing spans (Ngo et al., 2017). This matters for DAS because functional abstraction presupposes a choice of annotation unit: the act-bearing turn or span must first be isolated before it can be encoded as a script element.
A third neighboring tradition emphasizes relational structure. “Dependency Dialogue Acts” places dialog acts on response edges rather than on utterance units, supports multiple response relations, and targets multi-party, multi-threaded conversation (Cai et al., 2023). DAS does not adopt that graph formalism. Its core abstraction is instead a turn-level, function-call-like script that captures speaker, communicative function, and minimal semantic parameters for later localization and realization (Vasselli et al., 26 Sep 2025). This suggests a different emphasis: not response-graph analysis or corpus normalization for its own sake, but controlled regeneration of dialogue from abstract intent.
2. Representational structure
In the proposed framework, a DAS representation consists of turn-level structured expressions encoding the speaker, one or more dialogue act labels, and the parameters needed to preserve meaning. The paper states that DAS is “agnostic to the specific taxonomy used,” but instantiates a custom schema of 15 dialogue acts for multilingual dialogue generation (Vasselli et al., 26 Sep 2025).
| Dialogue act | Function | Example |
|---|---|---|
| Inquire | Seeks information or clarification | “What time does the meeting start?” |
| Clarify | Resolves ambiguity or misunderstanding | “I meant next Tuesday.” |
| Inform | Provides factual information | “This policy was updated last week.” |
| Express | Communicates emotion, attitude, or opinion | “That’s an excellent idea!” |
| Agree | Affirms a previous statement | “Yeah, that makes sense to me.” |
| Disagree | Explicitly disagrees | “That doesn’t seem right to me.” |
| Commit | Promises or agrees to future action | “Yes, I’ll take care of that.” |
| Acknowledge | Neutral receipt or backchannel | “I see.” |
| Seek Action | Requests or commands an action | “Could you please send me the file?” |
| Suggest | Proposes an action or idea | “Why don’t you try restarting your computer?” |
| Offer | Voluntarily provides help or a resource | “Would you like some water?” |
| Reject | Declines a proposal, offer, or request | “I’m sorry, but I’ll have to pass.” |
| Encourage | Praise, motivation, or reassurance | “Don’t worry, you’ll figure it out!” |
| Manage Topic | Opens, changes, or closes topics | “Let’s move on to the next point.” |
| Social Interaction | Greeting or small talk | “Hello. How are you?” |
The notation is function-call-like. The paper gives examples such as inform(subject=restaurant, attribute=famous, object=Cuervo_Gold_margaritas), inquire(topic=menu, subject=house_specials, availability=yes_no), and customer.seek_action(action=bring, object=Cuervo_Gold_margarita) (Vasselli et al., 26 Sep 2025). The parameters are intentionally open-ended: they are “whatever is necessary to capture the meaning of the utterance” and should be “the minimum amount of information necessary to convey all of the information of the sentence.” Observed fields include topic, subject, object, attribute, location, availability, action, and options (Vasselli et al., 26 Sep 2025).
The representation also permits multiple functions for a single turn. The paper gives the example "No, I don't think so" with functions ["disagree()", "express(doubt)"], indicating that DAS is compositional at the turn level rather than constrained to a single flat label (Vasselli et al., 26 Sep 2025). Beyond turn-level acts, the framework can include broader context such as speaker identity, scenario information, and character biographies. In the implemented pipeline, generated context specifies at minimum each speaker’s name, gender, age, and presumed relationship to the other speaker (Vasselli et al., 26 Sep 2025).
3. Encoding, localization, and realization workflow
The operational pipeline has three main stages: encoding, localization, and decoding (Vasselli et al., 26 Sep 2025).
In the encoding stage, each source utterance is converted into a DAS representation by classifying its dialogue act and extracting only the essential components needed to preserve its function, such as speaker, action, relevant conditions, and timeframe. The prompt used for encoding explicitly instructs the model to capture “what the speaker is doing ... rather than how they say it,” and to record only the minimum information necessary to reconstruct meaning (Vasselli et al., 26 Sep 2025). Encoding is performed with GPT-4o or GPT-4o-mini at temperature $0$ to maximize consistency (Vasselli et al., 26 Sep 2025).
A separate context-generation step produces scenario summaries and character biographies. These are then localized before script realization. The localization prompt instructs the system to adapt names, locations, social dynamics, common objects, politeness levels, family dynamics, and other cultural norms while preserving the original structure and meaning of the script (Vasselli et al., 26 Sep 2025). DAS localization itself is constrained: it should “not remove, condense, or add new topics,” and should keep all turns intact while adjusting only culturally relevant references (Vasselli et al., 26 Sep 2025).
In the decoding stage, the localized DAS is turned into target-language dialogue. The decoder is instructed to maintain the same turn structure, keep one line per turn, and aim for cultural authenticity even when names, places, or foods need to change (Vasselli et al., 26 Sep 2025). Temperature is set to $0.2$ for localization and decoding, permitting controlled variation (Vasselli et al., 26 Sep 2025). The framework supports both turn-by-turn decoding for interactive systems and whole-dialog decoding for script localization or synthetic dataset creation (Vasselli et al., 26 Sep 2025).
A canonical example is the restaurant dialogue in which Cuervo_Gold_margaritas is localized to Negroni for Italian, and the option pair blended, on_the_rocks becomes with_ice, without_ice. The resulting Italian realization preserves the communicative structure while altering culturally loaded content (Vasselli et al., 26 Sep 2025). This is the central procedural distinction between DAS and direct translation: surface strings are not translated first; abstract intent is localized and then re-realized.
4. Empirical evaluation
The principal evaluation uses DailyDialog as the English source pool, with 80 selected dialogues of 8 to 16 turns and average length 10.92 turns, balanced across the 10 DailyDialog topics with 8 dialogues per topic (Vasselli et al., 26 Sep 2025). XDailyDialog provides professionally translated dialogues in Italian, German, and Chinese, and COD is used for slot-level localization comparison in Arabic, Indonesian, Russian, and Swahili (Vasselli et al., 26 Sep 2025).
DAS annotation reliability is high. On the paper’s DAS function annotation task, inter-annotator agreement between two humans is $0.844$, GPT-4o-mini vs. humans is $0.765$ and $0.746$, GPT-4o vs. humans is $0.822$ and $0.769$, and GPT-4o vs. GPT-4o-mini is $0.805$ (Vasselli et al., 26 Sep 2025). On the DailyDialog 4-way taxonomy, reported F1 scores are 0.92 for Inform, 0.94 for Question, 0.63 for Directive, and 0.64 for Commissive (Vasselli et al., 26 Sep 2025). These numbers support the claim that LLMs can apply the DAS-like function schema with substantial consistency.
Meaning preservation is evaluated by decoding DAS back into English. Human evaluation reports Fluency $0.727$ for DAS versus $0.455$ for the original English, Logical Flow $0.2$0 versus $0.2$1, Situational $0.2$2 versus $0.2$3, and Meaning Preservation $0.2$4 (Vasselli et al., 26 Sep 2025). Automatic metrics show a paraphrase baseline BERTScore of $0.2$5, GPT-4o-mini DAS decode $0.2$6, and GPT-4o DAS decode $0.2$7; BLEU and ChrF++ are lower for DAS than for paraphrasing, which the paper interprets as greater lexical and structural variation rather than semantic failure (Vasselli et al., 26 Sep 2025).
For slot-level localization against human-localized COD outlines, DAS achieves average Precision $0.2$8, Recall $0.2$9, and F1 $0.844$0, with Arabic F1 $0.844$1, Indonesian $0.844$2, Russian $0.844$3, and Swahili $0.844$4 (Vasselli et al., 26 Sep 2025). The reported pattern is high precision and lower recall: DAS usually changes the right items, but sometimes misses changes that human localizers would make (Vasselli et al., 26 Sep 2025).
The strongest empirical claim concerns human preference. Against professional human translations, DAS is preferred across Fluency, Coherence, Cultural Relevance, and Situational Appropriateness in Italian, German, and Chinese, with $0.844$5 throughout (Vasselli et al., 26 Sep 2025). In Italian, for example, DAS wins 84.6% to 3.1% on Fluency and 85.4% to 3.1% on Cultural Relevance; in German, 81.5% to 1.5% on Fluency; in Chinese, 75.3% to 10.1% on Fluency and 68.5% to 9.0% on Cultural Relevance (Vasselli et al., 26 Sep 2025). DAS also outperforms direct GPT-4o translation and a single-prompt “translate and localize” baseline, again with $0.844$6 (Vasselli et al., 26 Sep 2025). The empirical conclusion is therefore narrow but strong: for the evaluated languages and datasets, abstract intent localization outperforms utterance-level translation as a route to naturalistic multilingual dialogue.
5. Relation to adjacent dialogue-act methodologies
DAS is not a substitute for the broader dialogue-act stack; it occupies a different layer. Related work addresses the neighboring problems of act-bearing unit discovery, contextual recognition, and structured discourse representation.
Functional-segment identification research for Vietnamese models dialogue-act boundary detection as sequence labeling over turns and reports chunk-level F1 of 90.42 on message data and 73.26 on phone data with Bi-LSTM-CRF, using functional segments as the pre-labeling unit for later dialogue-act assignment (Ngo et al., 2017). This line of work is upstream of DAS whenever source turns are not already aligned with communicative functions.
Contextual dialogue-act recognition research emphasizes that utterance meaning depends on local context and speaker structure. Local contextual attention with hierarchical structure reaches 80.34% accuracy on Switchboard and 85.81% on DailyDialog by biasing inter-utterance self-attention toward nearby turns (Dai et al., 2020). Speaker-turn modeling augments utterance embeddings with conversation-invariant turn embeddings and reports 83.2 on SwDA and 91.4 on MRDA (He et al., 2021). These results indicate that act recognition benefits from structured conversational context even before any localization or generation step is applied.
A different contrast appears in graph-based annotation. Dependency Dialogue Acts treats dialog acts and rhetorical relations as labels on response edges, allows multiple response relations and multiple dialog acts per dialog unit, and targets multi-party, multi-threaded conversations (Cai et al., 2023). DAS, by contrast, encodes turn-level functions plus parameters as a generation-oriented script (Vasselli et al., 26 Sep 2025). This suggests complementary usage: DDA is suited to discourse-structure annotation, whereas DAS is suited to intent-preserving multilingual regeneration.
Standardization work provides yet another contrast. ISO-aligned and task-universal schemes map heterogeneous corpora into shared dialogue-act inventories for cross-corpus training (Mezza et al., 2018, Paul et al., 2019). DAS is “agnostic to the specific taxonomy used,” but the implemented multilingual pipeline relies on a custom 15-act schema rather than a universal ISO subset (Vasselli et al., 26 Sep 2025). A plausible implication is that DAS can sit atop either standardized or task-specific act inventories, provided those inventories support minimal, reconstructive semantic parameterization.
6. Applications, limitations, and research directions
The paper explicitly positions DAS as useful for synthetic dataset creation, culturally adaptive conversational AI, language-learning dialogue generation, low-resource multilingual dialogue workflows, and controllable dialogue generation (Vasselli et al., 26 Sep 2025). Because the representation separates communicative intent from surface form, it can preserve conversational function while changing names, locations, brands, foods, and politeness conventions at localization time (Vasselli et al., 26 Sep 2025). This suggests a broader role for DAS as an intermediate layer between source dialogue semantics and target-language realization.
Its limitations are equally explicit. The implemented system requires multiple LLM calls, so cost and complexity are nontrivial (Vasselli et al., 26 Sep 2025). The main evaluation is on open-domain chit-chat rather than customer support, healthcare, legal, or tightly constrained task domains (Vasselli et al., 26 Sep 2025). The pipeline depends on GPT-4o-family models, and the paper notes that localization may inherit their cultural biases (Vasselli et al., 26 Sep 2025). Low-resource and morphologically complex languages remain uncertain, and generated context showed a strong male-female pairing bias, with 88% of conversations exhibiting that pattern despite mitigation attempts (Vasselli et al., 26 Sep 2025).
Scaling DAS also raises the annotation-efficiency and robustness problems already documented in dialogue-act research. In educational dialogue-act classification, diversity-aware active learning reached roughly $0.844$7 with 600 acquired samples, comparable to baseline performance with 900 samples, implying about 30% annotation cost savings (Tan et al., 2023). Separate work on low-resource, imbalanced dialogue-act datasets reports that AUC-maximization methods outperformed cross-entropy at training sizes 25, 50, 100, 200, and 400, and maintained Cohen’s $0.844$8 when the target class proportion reached 60% or 80% in robustness tests (Lin et al., 2023). These findings do not concern DAS directly, but they indicate that extending DAS to broader annotation programs will likely require careful sample selection and imbalance-aware training.
The future directions named in the DAS paper are pragmatic: replacing some steps with smaller models, retrieval-based systems, rule-based localization, or human-in-the-loop components; improving evaluation of slot-level cultural appropriateness; extending to low-resource languages and specialized domains; and mitigating bias more systematically (Vasselli et al., 26 Sep 2025). Taken together, the current evidence supports a specific view of DAS: not a universal theory of dialogue structure, but a functional intermediate representation that makes multilingual dialogue localization a problem of intent preservation and culturally appropriate realization rather than sentence-by-sentence translation.
