Abstract
Contextual influences on language often exhibit substantial cross-lingual regularities; for example, we are more verbose in situations that require finer distinctions. However, these regularities are sometimes obscured by semantic and syntactic differences. Using a newly-collected dataset of color reference games in Mandarin Chinese (which we release to the public), we confirm that a variety of constructions display the same sensitivity to contextual difficulty in Chinese and English. We then show that a neural speaker agent trained on bilingual data with a simple multitask learning approach displays more human-like patterns of context dependence and is more pragmatically informative than its monolingual Chinese counterpart. Moreover, this is not at the expense of language-specific semantic understanding: the resulting speaker model learns the different basic color term systems of English and Chinese (with noteworthy cross-lingual influences), and it can identify synonyms between the two languages using vector analogy operations on its output layer, despite having no exposure to parallel data.
Introduction
In human communication, language is used pragmatically to distinguish a target from its surroundings. A speaker's choice of words depends heavily on the context: while "the green shirt" might suffice in a room of red clothes, a more specific description like "the bright, yellowish-green one" is necessary when surrounded by various shades of teal. While these pragmatic patterns—such as increased verbosity in fine-grained contexts—appear universal, they are mediated by language-specific syntax and semantics. This research explores whether a neural speaker model can improve its pragmatic performance in one language (Mandarin Chinese) by leveraging data from another (English).
Methods and Datasets
The study utilizes a Reference Game framework. In this setup, a "speaker" model is presented with three colors and must generate a natural language description for a specific "target" color, enabling a "listener" to identify it correctly.
The researchers used an existing large-scale English color dataset and collected a new dataset of 5,774 Mandarin Chinese messages from human players. Contexts were categorized by difficulty: "Far" (distinct colors), "Split" (two similar colors), and "Close" (all three colors are very similar).
The model architecture consists of an LSTM encoder for the three colors and an LSTM decoder for token generation. To create a bilingual model, the vocabulary was expanded to include both languages, and a "language flag" was added to the input. This multitask approach forces the model to share hidden representations for colors while specializing its output for the target language.
Universal Pragmatics and Model Performance
Analysis of the human data reveals striking cross-lingual regularities. In both English and Mandarin, speakers adapt to contextual difficulty in similar ways: as colors become harder to distinguish, they use more words, more negations (e.g., "not the bright one"), and more superlatives (e.g., "the darkest one").
The bilingual training approach yielded significant improvements in Pragmatic Informativeness—the ability of generated text to successfully guide a model-derived listener. For Mandarin, informativeness rose from 67.75% in the monolingual model to 72.89% in the bilingual version. Interestingly, the bilingual model better captured human-like patterns of context dependence, avoiding the over-verbosity often seen in monolingual models during "easy" contexts.
Semantic Integrity and Synonym Discovery
Despite sharing internal representations, the model maintained language-specific semantic boundaries. For instance, it respected the Mandarin usage of "red" (红) for shades that English speakers might split into "red" and "pink." Furthermore, the model demonstrated an emergent ability to identify synonyms across languages. By performing vector analogy operations on the model's output layer, the researchers could accurately "translate" words (e.g., mapping "蓝色" to "blue") without any exposure to parallel translation data.
Conclusion
This work demonstrates that bilingual training is a powerful tool for improving pragmatic performance, particularly in lower-resource settings like the Mandarin color dataset. By learning shared communication strategies, the model becomes a more effective and human-like communicator. Future research will explore whether these benefits persist across languages with more divergent color systems or syntactic structures.
Summary updated