TL;DR
This paper introduces a novel computational model, the Bilingual Dual-path model, to simulate Spanish-English code-switching. The model successfully generates code-switches without explicit training data, suggesting that internal linguistic factors, not just exposure, drive this phenomenon. This work offers a new tool for understanding the cognitive mechanisms behind bilingual language use.
What Problem Does This Paper Address?
Code-switching, the practice of alternating between languages within or between sentences, is a common behavior among multilingual speakers. While extensively studied by linguists and sociolinguists, the underlying cognitive mechanisms remain poorly understood. Previous computational approaches have focused on detecting or recognizing code-switches, but few have attempted to simulate the spontaneous generation of code-switches. This paper aims to bridge this gap by using a computational cognitive model to investigate whether code-switching can arise from the internal structure and distributional properties of languages, rather than solely from exposure to code-switched input.
Methods
The core of this research is the Bilingual Dual-path model, a Recurrent Neural Network (RNN) based on the Simple Recurrent Network (SRN) architecture. This model simulates bilingual sentence production by predicting words sequentially. It features two pathways:
- Meaning System: Maps words to concepts, thematic roles, event semantics, and the intended language.
- Sequencing System: Learns abstract syntactic patterns.
The model is trained on simulated Spanish and English sentences, but crucially, not on any code-switched examples. This allows researchers to test whether code-switching emerges naturally from the model's internal representations of the two languages. The model's "message" input includes semantic roles, concepts, and target language information, enabling it to generate sentences. For the experiments, the model was trained on a large set of Spanish-English sentence-message pairs. During testing, language control was manipulated to allow for the possibility of switching.
Datasets
The paper does not explicitly name or reference external datasets. Instead, it describes the generation of artificial "miniature" languages for training and testing.
- Training Data: 3040 randomly generated sentence-message pairs in English and Spanish. Crucially, this set did not contain any code-switched sentences.
- Test Data: 760 unseen sentences generated in the same manner as the training set.
Key Insights and Ideas
- Internal drivers of code-switching: The model's ability to spontaneously generate code-switches without explicit training data strongly supports the hypothesis that internal linguistic factors (like language distributional properties) play a significant role, not just external exposure to code-switched speech.
- Auxiliary phrase asymmetry: The model successfully simulated an observed asymmetry in Spanish-English progressive structures, where switches between the auxiliary verb ("estar") and participle are common, but switches between the perfect auxiliary ("haber") and participle are rare. This suggests that the grammaticalization status of auxiliaries (e.g., "estar" having more semantic weight than the highly grammaticalized "haber") can explain this asymmetry.
- Model as a research tool: The Bilingual Dual-path model provides a powerful, manipulable platform for testing hypotheses about code-switching, allowing researchers to isolate variables and explore phenomena that are difficult to study with human participants alone.
Key Results
- Spontaneous Code-Switching: The model produced code-switches in 18.09% of correctly produced sentences, even without code-switched training input.
- Alternational vs. Insertional Switching: Alternational switches (multi-word sequences) were more frequent (13.57%) than insertional switches (single words, 4.52%).
- Syntactic Category Preferences: Alternations at the Noun Phrase (NP) level were more common than at the Verb Phrase (VP) or Prepositional Phrase (PP) levels, aligning with previous linguistic observations.
- Auxiliary Phrase Asymmetry: The "haber-model" showed a strong preference for progressive participle switches (4.16%) over perfect participle switches (2.18%), replicating the observed asymmetry. Replacing "haber" with "tener" (a more semantically rich auxiliary) in the "tener-model" eliminated this preference, supporting the grammaticalization account.
Limitations and Future Work
- Artificial Languages: The use of miniature languages, while useful for control, may not fully capture the complexity of real-world languages.
- Limited Scope: The current experiments focus on Spanish-English and specific phenomena.
- Future Work: The authors suggest that the model can be easily adapted to other language pairs and used to test further hypotheses, such as the role of cognates in triggering code-switching. They also highlight the potential for exploring other code-switching phenomena and the interaction between internal factors and external influences.
Summary updated