Source-linked AI summary

PingPong: A Natural Benchmark for Multi-Turn Code-Switching Dialogues

Mohammad Rifqi Farhansyah, Hanif Muhammad Zhafran, Farid Adilazuarda, Shamsuddeen Hassan Muhammad, Maryam Ibrahim Mukhtar, Nedjma Ousidhoum, Genta Indra Winata, Ayu Purwarianti, Alham Fikri Aji

arXiv:2601.17277v1cs.CL

TL;DR

Existing code-switching dialogue benchmarks are underexplored, outdated, and increasingly misaligned with everyday multilingual communication. PINGPONG introduces a human-authored, multi-party benchmark spanning five language combinations and three downstream tasks, while evaluations reveal persistent limitations in current language models.

  • Problem

    Existing code-switching dialogue benchmarks are underexplored, increasingly outdated, and less reflective of everyday multilingual communication.

  • Method

    PINGPONG uses manually crowdsourced conversations from fluent multilingual speakers across five language combinations, supporting Question Answering, Summarization, and Topic Classification.

  • Results

    Evaluations of open-source and proprietary language models reveal persistent limitations in modeling code-switched, multi-party conversations, with regionally designed models generally outperforming general-purpose multilingual models.

  • Takeaways & Limitations

    PINGPONG identifies current NLP limitations and provides a foundation for developing more robust, inclusive systems for the world’s multilingual majority.

  • Takeaways & Limitations

    The benchmark covers four under-studied languages across five combinations and three geographic regions, so its scope is not exhaustive and remains extendable.

Abstract

from arXiv · show

Code-switching is a widespread practice among the world's multilingual majority, yet few benchmarks accurately reflect its complexity in everyday communication. We present PingPong, a benchmark for natural multi-party code-switching dialogues covering five language-combination variations, some of which are trilingual. Our dataset consists of human-authored conversations among 2 to 4 participants covering authentic, multi-threaded structures where replies frequently reference much earlier points in the dialogue. We demonstrate that our data is significantly more natural and structurally diverse than machine-generated alternatives, offering greater variation in message length, speaker dominance, and reply distance. Based on these dialogues, we define three downstream tasks: Question Answering, Dialogue Summarization, and Topic Classification. Evaluations of several state-of-the-art language models on PingPong reveal that performance remains limited on code-switched inputs, underscoring the urgent need for more robust NLP systems capable of addressing the intricacies of real-world multilingual discourse.

1 Introduction

PINGPONG introduces a benchmark for natural, multi-party code-switching dialogues that addresses limitations in existing resources through broad language coverage, diverse conversational structure, and evaluation across three downstream tasks.

  • Motivation: PINGPONG targets underexplored code-switched dialogue, while existing benchmarks are increasingly outdated and less reflective of everyday code-switching.The benchmark is motivated by insufficient evaluation of multilingual dialogue systems and limitations in prior datasets.
  • Benchmark Design: PINGPONG contains authentic interactions among 2 to 4 speakers, including dialogues from 17 to 189 turns with long-distance replies to earlier messages.These properties are intended to capture the fluid dynamics of multilingual communication in conversational settings.
  • Benchmark Design: The benchmark spans five language combinations across high-resource, low-resource, and diverse-script settings.This broad coverage supports evaluation across varied multilingual scenarios.
  • Evaluation Tasks: PINGPONG evaluates Question Answering, Dialogue Summarization, and Topic Classification under code-switching conditions.Together, these tasks assess language understanding and generation across multiple aspects of multilingual dialogue.
  • Evaluation Tasks: Experiments with open-source and proprietary LLMs reveal persistent limitations in modeling code-switched, multi-party conversations.The evaluation framework is designed to expose challenges for current systems.

2 Related Work

Existing code-switching datasets cover many NLP tasks but remain limited in release sustainability, linguistic diversity, language coverage, and conversational realism. Their evaluation tasks also often fail to reflect contemporary NLP challenges.

  • Dataset limitations: Existing code-switching datasets face limitations in public availability, linguistic diversity, and language or regional coverage.Several datasets are not publicly released, most focus on bilingual settings, and many languages remain under-explored.
  • Dataset limitations: Most existing benchmarks focus exclusively on bilingual code-switching, with only one dataset explicitly addressing trilingual scenarios.
  • Conversational realism: Many datasets do not capture spontaneous, interactive code-switching in real-world conversational settings.
  • Evaluation scope: Existing evaluation tasks are frequently limited to traditional benchmarks that no longer reflect contemporary NLP challenges.

3 PINGPONG Dataset

PINGPONG is built from manually collected, native-speaker conversations across five language combinations, including trilingual settings. Its dialogues support question answering, summarization, and topic classification while preserving natural multi-party and code-switching structure.

  • Data collection: PINGPONG covers five language combinations, including Indonesian–English, three trilingual combinations, Hausa–English, and Algerian Arabic–Standard Arabic–French.Native speakers collect dialogues and curate downstream tasks for each language combination.
  • Conversational Dialogue Collection: Participants hold 15-minute conversations on assigned topics and may code-switch at word, sentence, paragraph, or utterance levels.Discord replies allow participants to respond directly to specific utterances while preserving dialogue structure.
  • Downstream tasks: PINGPONG supports question answering, dialogue summarization, and multi-label topic classification across predefined topic categories.Topics include science and technology, entertainment, society and culture, education, and daily life.
  • Dataset statistics: Dialogue statistics measure linguistic complexity using Code-Mixing Index and Switch Point Fraction.CMI measures mixing intensity, while SPF measures the frequency of transitions between languages.
  • Dataset analysis: Human dialogues show participation imbalance, multi-threaded replies, and highly variable utterance lengths, unlike more monotonous machine-generated dialogues.Machine-generated conversations tend toward rigid turn-taking, uniform speaker participation, consistent utterance lengths, and immediate linear replies.

4 Experimental Setup

The experiments evaluate diverse model families across PINGPONG’s three tasks using task-specific prompts, shot settings, reasoning configurations, output constraints, and metrics. Models span English-centric, multilingual, regional, instruction-tuned, and reasoning-oriented systems.

  • Models: The study evaluates models across question answering, dialogue summarization, and topic classification.Models are categorized by linguistic coverage and include base, reasoning-oriented, and instruction-tuned variants.
  • Models: The evaluated models include English-centric, multilingual, and region-specific systems covering the study’s target languages.
  • Reasoning behavior: The experiments compare reasoning-enabled behavior with settings where models generate explicit reasoning traces through prompting.Qwen3 provides native thinking traces, while other models can be prompted to produce reasoning steps.
  • Prompting: Prompting varies the number of in-context examples, with QA using zero- and one-shot settings and summarization using zero-, one-, and three-shot settings.Examples are drawn from dialogues outside the evaluation set.
  • Output format: Models are instructed to return JSON outputs, and selected configurations include reasoning traces for analyzing model logic.
  • Evaluation metrics: QA and topic classification use accuracy, while dialogue summarization uses ROUGE, METEOR, CHRF++, and BERTScore, with ROUGE-L reported in the main paper.

5 Results and Analysis

Performance is generally poor across PINGPONG’s downstream tasks, although regionally designed models generally outperform general-purpose multilingual models. Reasoning usually helps, while few-shot prompting provides inconsistent benefits outside summarization, and language-level comparisons remain difficult to interpret.

  • Overall Results: Most evaluated models perform poorly across PINGPONG’s downstream tasks, indicating substantial challenges for current systems.
  • Overall Results: Regionally designed models generally outperform general-purpose multilingual models.The authors attribute the observed benefit to specialization for particular linguistic and cultural characteristics.
  • Limitations: Non-parallel dialogue content prevents reliable conclusions about whether any language is inherently more difficult than another.Observed performance differences remain tied to the specific conversation content of each language.
  • Reasoning Behavior: Reasoning-enabled configurations generally improve performance across the benchmark’s tasks.The authors associate this pattern with additional computation and internal verification from reasoning traces.
  • Prompting: Few-shot examples do not consistently improve QA or topic classification, but they significantly improve dialogue summarization.
  • Reasoning Behavior: Reasoning improves answerable questions, whereas unanswerable cases benefit only from explicit Qwen3 reasoning traces and may decline with implicit reasoning.Few-shot effects are model-dependent, ranging from marginal gains to no improvement or slight regressions.

6 Conclusion

PINGPONG captures authentic human multilingual discourse across five language combinations and shows that current state-of-the-art models still struggle with natural, multi-party code-switching.

  • PINGPONG captures authentic complexity in human multilingual discourse across five language combinations.
  • Human-authored conversations among 2 to 4 participants include multi-threaded dynamics and varied speaker dominance often missing from synthetic corpora.
  • The dataset analysis finds greater naturalness and structural diversity than machine-generated alternatives, particularly in message length and reply distances.
  • PINGPONG evaluates Question Answering, Dialogue Summarization, and Topic Classification to address gaps in prior benchmarks.
  • State-of-the-art models still struggle with the intricacies of natural, multi-party code-switching.

Limitations

The benchmark covers four complementary under-studied languages across five combinations and three geographic regions, but does not capture every code-switching variation.

  • The study covers 4 complementary under-studied languages across 5 language combinations spanning 3 geographic regions.
  • The benchmark’s scope is not exhaustive across all code-switching variations in its five language combinations.
  • The authors describe the dataset as a scalable foundation that can be extended to additional languages and regions.
  • The study uses a relaxed CMI formulation because reliable tools for advanced CMI metrics are unavailable for the under-studied languages.

Ethical Considerations

The benchmark’s data-collection materials organize recruitment, language assessment, dialogue construction, and downstream annotation procedures for multilingual participants and tasks.

  • Ethical Considerations: Annotators receive compensation above the local minimum wage and detailed instructions, with demographic collection conducted using informed consent.
  • Ethical Considerations: Annotators are instructed not to use offensive language, although some instances may be missed and later reported cases may prompt benchmark updates.
  • Data Collection: Recruitment forms collect demographic information, self-reported language use, and language-assessment responses for annotator selection.
  • Data Collection: Language assessment asks respondents to compose a short paragraph using a specified language combination to simulate natural code-switching behavior.
  • Annotation: Guidelines cover dialogue construction, Question Answering, Dialogue Summarization, and Naturalness scoring to support consistent dataset creation.

F Quantitative Results

The quantitative-results section organizes findings across experimental configurations, including Question Answering comparisons of reasoning settings and prompting strategies.

  • The section provides a comprehensive overview of quantitative findings across all experimental configurations.
  • Question Answering: Question Answering results compare reasoning-enabled with non-reasoning configurations.
  • Question Answering: Question Answering results also compare zero-shot with one-shot prompting strategies.
  • Question Answering: Table 8 reports per-language-combination accuracy statistics for reasoning versus non-reasoning approaches.
  • Question Answering: Table 9 reports per-language-combination accuracy statistics for zero-shot versus one-shot approaches.

F.2 Text Summarization

The text summarization experiments examine reasoning mechanisms and prompt shot size, with detailed results reported across five language combinations. They also compare human-written and machine-generated conversations.

  • The text summarization evaluation examines the role of reasoning mechanisms and prompt shot size.
  • Reasoning vs. Non-Reasoning: Reasoning and non-reasoning approaches are compared across five language combinations using detailed metric tables.The comparisons cover ID-EN, JV-ID-EN, SU-ID-EN, HA-EN, and ar_DZ-FR.
  • 0-shot vs. Few-shot: Zero-shot and few-shot approaches are compared across the same five language combinations for text summarization.The reported metrics include ROUGE, METEOR, CHRF++, and BERTScore measures.
  • Human-written and machine-generated conversations are compared using overall quantitative statistics.

G Additional Results

The additional results describe unanswerable-question categories, evaluation metrics across language combinations and prompting conditions, and annotation and prompt guidelines. They also include comparisons between human-written and machine-generated conversational text.

  • Question Answering annotation targets five unanswerable categories: Negation, Antonym, Entity-Swap, Mutual-Exclusion, and Impossible-Condition.Three language combinations produced up to five unanswerable questions per instance: ID-EN, JV-ID-EN, and SU-ID-EN.
  • Text summarization tables report ROUGE-1 through ROUGE-4, METEOR, CHRF++, and BERTScore precision, recall, and F1 across five language combinations.
  • Additional summarization tables compare zero-shot with few-shot approaches across the same five language combinations.
  • The results include full quantitative statistics comparing human-written and machine-generated conversational patterns.
  • The appendix provides annotator guidelines for Question Answering, Dialogue Summarization, and human-versus-machine naturalness judgments.
  • Prompt templates specify task instructions for Question Answering, Dialogue Summarization, Topic Classification, and machine-generated conversational text.The Question Answering and Dialogue Summarization templates include optional few-shot components.
Loading 2601.17277v1…