In-domain Paired Bilingual Dialogue Datasets

• In-domain paired bilingual dialogue datasets are specialized language resources that align turn-based dialogues across languages within specific domains.
• They employ rigorous alignment and annotation methods, including human translation and automatic projection, to ensure context and slot consistency.
• These resources underpin advances in neural dialogue systems, context-aware machine translation, and cross-lingual recommendation.

In-domain paired bilingual dialogue datasets are specialized language resources in which turn-by-turn, contextually consistent dialogue utterances from a single domain are aligned across two (or more) languages. Unlike parallel sentences in general-purpose bitexts or open-domain chat collections, these datasets capture the discourse structures, role hierarchies, slot-value annotations, and pragmatic phenomena unique to specific application settings, such as task-oriented assistants, media subtitles, conversational recommendation, or code translation. Robust in-domain paired dialogue resources underpin the development and evaluation of advanced neural dialogue systems, context-aware machine translation (MT), dialogue state tracking (DST), cross-lingual recommendation, and dialogue-grounded generative models in multi- or bilingual contexts.

1. Types and Scope of In-domain Paired Bilingual Dialogue Datasets

In-domain paired bilingual dialogue corpora span multiple subgenres and application verticals:

• Task-Oriented Dialogue (TOD): Datasets such as IndoToD (English–Indonesian) (Kautsar et al., 2023), X-RiSAWOZ (Chinese–English, French, Hindi, Korean, code-mixed) (Moradshahi et al., 2023), and those derived from MultiWOZ (Moradshahi et al., 2021) capture slot-filled conversations in domains including restaurant search, navigation, weather, scheduling, and travel.
• Media and Subtitle Dialogue: Corpora built from TV and movie scripts plus subtitles—e.g., the English–Chinese "Friends" corpus (Wang et al., 2016)—provide richly annotated, multi-speaker, situational dialogues.
• Conversational Recommendation: DuRecDial 2.0 (English–Chinese) targets human-to-human recommendation, integrating profiles, knowledge graphs, and context-rich recommendation turns (Liu et al., 2021).
• Chat-Translation Benchmarks: Corpora such as BConTrasT (En–De) and BMELD (En–Zh) (Liang et al., 2021) construct one-to-one aligned chat sessions in domains like customer service or TV-series dialogue.
• Specialized Code or Speech Translation Datasets: For cross-language code translation, datasets with multi-turn LLM-driven reasoning (Fortran↔C++, C++↔CUDA) (Chen et al., 29 Nov 2025) are available. For prosody mapping, audio-focused corpora like DRAL (EN–ES) (Ward et al., 2022) provide turn- and fragment-level aligned speech.
• Domain-constrained Fictional Worlds: The Harry Potter Dialogue (HPD) dataset (English–Chinese), covering all book dialogues and annotating character attributes, relations, and scenes (Chen et al., 2022).
• Complex Spoken Dialogue Benchmarks: C³ (EN–ZH, CN) provides evaluation instances targeting ambiguity, context dependencies, and multi-turn spoken interaction (Ma et al., 30 Jul 2025).

This broad spectrum is distinguished by strict in-domain focus (e.g., all dialogue occurs within a constrained activity or world), explicit bilingual alignment (utterance turn-pairs), and detailed metadata supporting downstream semantic and pragmatic tasks.

2. Construction Methodologies and Alignment Procedures

Methodological rigor in aligning dialogues across languages is critical to maintain discourse structure, slot consistency, and conversational flow. Key techniques include:

• Annotation Projection and Speaker Tagging: As in the "Friends" corpus (Wang et al., 2016), monolingual scripts are aligned to bilingual subtitles via vector-space retrieval. Speaker and dialogue-boundary tags are then projected onto the paired subtitle sentences using cosine similarity in a TF–IDF space.
• Human Translation and Post-Editing: IndoToD uses delexicalization followed by manual translation at the utterance level, preserving alignment and minimizing annotation overhead (Kautsar et al., 2023). X-RiSAWOZ combines machine translation with hybrid alignment (dictionary and neural-attention based), then applies post-editing with slot- and entity-checkers in a custom GUI (Moradshahi et al., 2023).
• Knowledge and Profile Enrichment: DuRecDial 2.0 collects and manually aligns profiles, goals, and knowledge triples in both languages, enforcing strict quality controls (≤1% error threshold for translators) (Liu et al., 2021).
• Constrained Automatic Translation: Contextual Semantic Parsing for Multilingual TOD leverages neural attention alignments to map slot values exactly and localize ontology mentions in each language, yielding turn-aligned bilingual representations (Moradshahi et al., 2021).
• Audio and Prosody-Synchronous Design: DRAL’s protocol records spontaneous dialogues, then segment-wise re-enactment in the target language, enabling near-synchronous speech fragment pairs annotated for timing and pragmatic function (Ward et al., 2022).
• Dual-LLM Reasoning for Code: In code translation, an LLM Questioner–Solver framework generates multi-turn paired code explanations and translations, validated iteratively with external compilation and unit testing (Chen et al., 29 Nov 2025).

These methodologies yield not only raw bilingual utterance pairs, but also guarantee projection of annotations such as speaker roles, slot–value states, knowledge context, and other task-specific ontological labels.

3. Dataset Structure, Annotation Schemes, and Metadata

In-domain paired bilingual dialogue datasets share a set of canonical schema features:

Dataset	Alignment Unit	Key Annotations
IndoToD (Kautsar et al., 2023)	Utterance/turn	Delexicalized slot spans, belief state, KB references
"Friends" (Wang et al., 2016)	Utterance/scene	Speaker ID, boundary tags, script ↔ subtitle mapping
DuRecDial 2.0 (Liu et al., 2021)	Sub-dialog/turn	Profile, goals, KG triples, context, response
X-RiSAWOZ (Moradshahi et al., 2023)	Turn	Belief state, API acts, slot-span alignment
HPD (Chen et al., 2022)	Turn	Scene summary, character relations, 13+ attributes
DRAL (Ward et al., 2022)	Audio fragment	Speaker, fragment timing, dialogue act coverage
C³ (Ma et al., 30 Jul 2025)	Audio/text snippet	Category, subcategory, ambiguity/contextual challenge

Annotation schemas are typically designed to facilitate role transfer (via speaker or role tags), track stateful fields (belief or slot-value pairs), or capture scene/intent information. For instance, each IndoToD utterance is annotated with a current belief state $b_t = \{(s_1, v_1), \dots\}$, while subtitle corpora encode speaker and shot/scene boundaries (Wang et al., 2016). X-RiSAWOZ aligns turn-level slot values via hybrid span-marking, with all languages using shared slot ontologies (Moradshahi et al., 2023). HPD encodes speakers, turn text, evolving character relations, and attributes at every point in the story (Chen et al., 2022).

4. Evaluation Protocols and Benchmarks

Assessment of bilingual dialogue models built on these datasets employs automatic and human-centric metrics:

• Translation and Generation Quality: BLEU, TER, and token-level F1 remain standard. For example, BLEU is defined as $\mathrm{BLEU} = \exp(\sum_{n=1}^N w_n \log p_n)$ with $p_n$ the $n$-gram precision (Kautsar et al., 2023, Liang et al., 2021, Moradshahi et al., 2023).
• Dialogue Success/Uptake Metrics: Match Rate (fraction of dialogues matching user goal), Success F1 (requested slots fulfilled), and Combined Score are used for TOD (Kautsar et al., 2023). Joint Goal Accuracy (JGA) in DST tasks measures the rate of exact slot-value set matches (Moradshahi et al., 2021, Moradshahi et al., 2023).
• Contextual and Semantic Fidelity: Consistency, coherence, and fluency (often via human judgment) evaluate chat translation (Liang et al., 2021). For DRAL, prosodic feature preservation is validated by cross-language F0 correlation and dynamic time warping (Ward et al., 2022).
• Task-specific Criteria: For code translation, compilation rate, unit test pass/fail, and functional correctness assess translation usability (Chen et al., 29 Nov 2025). C³ evaluates ambiguity resolution (phonological, semantic, context) with both automatic (LLM-based) and human scores (Ma et al., 30 Jul 2025).
• Quality Control in Construction: Most corpora conduct multi-level validation (cross-annotator checking, post-lexicalization review, or crowd-comparison), with error rates strictly bounded (e.g., <1% in DuRecDial 2.0) (Liu et al., 2021).

Benchmarking reveals that data quality and alignment method significantly impact system performance. For instance, slot-value localized translation in X-RiSAWOZ boosts BLEU by 10–15 points in slot-sensitive contexts (Moradshahi et al., 2023), and integrated speaker boundary annotation in "Friends" scripts improves BLEU by ~0.5 (Wang et al., 2016).

5. Impact, Limitations, and Application Scenarios

The availability of in-domain paired bilingual dialogue corpora has enabled the development of:

• Context-Aware Translation Models: Exploiting dialogue history, speaker role, and multi-turn context for improved translation consistency (Wang et al., 2016, Maruf et al., 2018, Liang et al., 2021).
• End-to-End Dialogue Agent Training: Zero/few-shot transfer across languages using high-quality parallel data for multilingual assistant deployment (Moradshahi et al., 2023, Kautsar et al., 2023).
• Dialogue-Grounded Evaluation: Fine-grained metrics capturing not only sentence-level correctness but multi-turn coherence, intent revelation, and pragmatic alignment (Liu et al., 2021, Chen et al., 2022).
• Speech/Prosody Cross-Language Studies: Paired spoken corpora like DRAL enable modeling of prosodic mapping in speech-to-speech MT (Ward et al., 2022).
• Code Translation Reasoning: Dialogue-based code translation datasets with integrated toolchain feedback have advanced code reasoning in low-resource domains (Chen et al., 29 Nov 2025).

Limitations include domain bias (e.g., TV scripts, Harry Potter fiction, UTEP student conversations in DRAL), labor-intensive annotation and post-editing, and challenges in scaling to low-resource languages or spoken genres. Ambiguity handling in real dialogues, prosodic and paralinguistic fidelity, and cross-domain entity generalization present ongoing challenges (Ward et al., 2022, Ma et al., 30 Jul 2025).

6. Data Access, Licensing, and Best Practices

Most leading datasets are released under open-access licenses:

Dataset / Resource	Access URL	License
IndoToD	https://github.com/dehanalkautsar/IndoToD	MIT
X-RiSAWOZ	[see paper for link]	Open-source
"Friends" (EN–ZH)	http://computing.dcu.ie/~lwang/resource.html	Not specified
HPD (Harry Potter)	[see paper for link]	Not specified
DuRecDial 2.0	[see paper for link]	Not specified
BConTrasT / BMELD	https://github.com/Unbabel/BConTrasT, https://github.com/XL2248/CPCC	Public domain
DRAL	https://www.cs.utep.edu/nigel/dral/	Not specified
C³	https://huggingface.co/datasets/ChengqianMa/C3	CC-BY-4.0

Best practices for downstream use include leveraging structured metadata (speaker, slot, scene), respecting annotation scope (single-reference caveats for BLEU), and tailoring model architectures to utilize in-domain signals (e.g., speaker-conditioned LMs, factor-based NMT, context-truncated parsers).

7. Summary and Significance

In-domain paired bilingual dialogue datasets form the empirical backbone for cross-lingual dialogue system research, robust neural MT under conversational conditions, and advanced knowledge-grounded dialogue reasoning. Their precise alignment, rich annotation, and context-preserving structure facilitate progress beyond sentence-level translation, supporting the design of next-generation multilingual conversational AI, recommendation interfaces, and dialogue evaluation paradigms. Ongoing improvements in alignment, slot/entity consistency, multi-modal support, and scalable annotation protocols will further enhance their applicability across languages, domains, and modalities (Wang et al., 2016, Moradshahi et al., 2023, Chen et al., 2022, Moradshahi et al., 2021).