Investigating the content and form of referring expressions in Mandarin: introducing the Mtuna corpus

Kees van Deemter, Le Sun, Rint Sybesma, Xiao Li, Chen Bo, Muyun YangINLG 2017 · DOI · OpenReview

Abstract

Most computational models for Referring Expression Generation (REG) are grounded in Western European languages, leaving a significant gap in our understanding of East Asian languages like Mandarin Chinese. This paper introduces the Mtuna Corpus, the first "Data-Text" corpus for Referring Expressions (REs) in Mandarin. By analyzing how native speakers describe objects in controlled settings, the researchers challenge existing linguistic theories regarding "canonical" patterns and highlight the heavy reliance on context in Mandarin communication.

Introduction

Referring Expression Generation (REG) is a core task in Natural Language Generation (NLG), focusing on how systems can identify specific objects for a listener. However, current models are predominantly trained on English-centric datasets like the TUNA corpus. Mandarin Chinese presents a unique linguistic landscape: it lacks articles to mark definiteness and does not consistently mark grammatical number. To address the lack of standardized data and test theoretical assumptions about Mandarin reference, the authors developed the Mtuna Corpus.

Methodology and Experimental Design

The researchers employed a web-based language elicitation experiment following the TUNA paradigm. Participants were presented with grids containing a target object—distinguished by a red border—and several distractor objects. They were tasked with typing a description that would allow a listener to uniquely identify the target.

The experimental interface for the furniture domain, showing a target green desk highlighted in a red box.

The study involved 37 speakers and covered 40 pictorial scenes across two domains: furniture and people. A critical component of the design was the variation of sentence position. Half of the participants produced REs in a pre-verbal position, while the other half used a post-verbal position. This allowed the team to investigate the long-held linguistic belief that Mandarin syntax strictly dictates definiteness (i.e., pre-verbal positions are reserved for definite references).

The experimental interface for the people domain, where participants identify specific individuals among distractors.

Key Insights: The Dominance of Bare Nouns

Traditional linguistic literature often highlights three primary patterns for reference in Mandarin: Demonstrative + Classifier + Noun (DCN), Demonstrative + Noun (DN), and the bare Noun Group (N). Surprisingly, the Mtuna data showed that the N pattern (bare nouns) dominated almost exclusively for singular references.

In the experimental trials, the N pattern was used 315 times in the pre-verbal position and 326 times in the post-verbal position. Remarkably, the "canonical" DN and DCN patterns were used zero times for singular targets. These findings suggest that Mandarin is a "cool" language—one that relies on communicative context and default assumptions (e.g., assuming a bare noun is singular unless context suggests otherwise) rather than explicit grammatical markers.

Results and Discussion

The analysis of the Mtuna Corpus yielded several key results:

  • Number Marking: Speakers were far more likely to explicitly mark number (e.g., using liang for "two") when referring to plural sets than when referring to singular objects.
  • Definiteness and Syntax: The relationship between sentence position and definiteness proved more nuanced than a simple rule. Indefinite markers appeared in pre-verbal positions more frequently than expected, suggesting that the "pre-verbal = definite" mapping is a preference rather than a strict requirement.
  • Maximality and Givenness: Explicit markers for "all" (dou) were rare (14 occurrences), and no explicit markers for "givenness" or familiarity were found, further reinforcing the language's reliance on underspecification.

Conclusion and Future Work

The Mtuna Corpus provides a much-needed benchmark for Mandarin NLG tasks and offers empirical evidence that challenges traditional linguistic descriptions of referring expressions. While the study's reliance on written input and a specific participant demographic are noted limitations, it opens the door for future research into how listeners interpret these underspecified expressions. The researchers plan to further investigate the role of "syntactic sugar"—function words that enhance naturalness without adding semantic content—and the cognitive processes behind "default" interpretations in Mandarin.

The Mtuna Corpus is freely available for research purposes, encouraging further exploration of cross-linguistic diversity in computational linguistics.

Summary updated

BibTeX
@inproceedings{deemter2017investigating,
  author = {Kees van Deemter and Le Sun and Rint Sybesma and Xiao Li and Chen Bo and Muyun Yang},
  title = {{Investigating the content and form of referring expressions in Mandarin: introducing the Mtuna corpus}},
  booktitle = {INLG},
  year = {2017},
  doi = {10.18653/v1/w17-3532},
}
Figure