Literature Review: Code-Switching Morphosyntactic Analysis

Status: Reference Last updated: 2026-04-21 09:01 EDT

This review surveys computational approaches to morphosyntactic analysis of code-switched text, with particular attention to techniques relevant to batchalign3’s L2 morphotag feature (default-on since 2026-04-15). The goal is to situate our approach within the existing literature and identify whether we are reinventing known techniques or contributing something novel.

Background: The Problem

Code-switching (CS) is the alternation between two or more languages within a single conversation, utterance, or even word. In TalkBank’s CHAT format, word-level code-switching is marked with @s:

*EVA: ich möchte film@s studies@s machen .

The challenge: morphosyntactic analysis (POS tagging, lemmatization, dependency parsing) requires language-specific models, but code-switched text mixes languages at the word level. Monolingual NLP models degrade at switch points because they encounter out-of-vocabulary words from the other language.

Prior Art

Solorio & Liu (2008): Two Monolingual Taggers with Supervised Merger

The foundational work on POS tagging for code-switched text. Solorio and Liu ran separate English and Spanish POS taggers on English-Spanish code-switched data, then trained an SVM classifier to combine their outputs. Key finding: using the output of both monolingual taggers as features gave the best results, outperforming either tagger alone.

Relevance to our work: This is the closest antecedent to our two-model approach. The difference: they trained a supervised classifier to merge tagger outputs (requiring annotated code-switching training data), while our structural merge uses a rule-based deprel constraint that requires no code-switching training data at all.

Citation: Solorio, T. & Liu, Y. (2008). Part-of-Speech Tagging for English-Spanish Code-Switched Text. Proceedings of EMNLP 2008.

Bhat et al. (2018): Neural Stacking for Code-Switching Dependency Parsing

Built the first Hindi-English code-switching Universal Dependencies treebank and proposed “neural stacking”, base monolingual parsers (Hindi and English) whose hidden representations are fed to a stacking parser trained on code-switched data. Achieved 90.5% POS accuracy and 71.0% LAS on code-switched dependency parsing.

Relevance to our work: The neural stacking approach is architecturally related to our structural merge, both use monolingual models as a foundation and combine their outputs. The critical difference: Bhat et al. require a code-switching treebank to train the stacking layer. Our approach requires only monolingual Stanza models and a hand-crafted deprel-to-POS constraint table.

Citation: Bhat, I.A., Bhat, R.A., Shrivastava, M. & Sharma, D. (2018). Universal Dependency Parsing for Hindi-English Code-switching. Proceedings of NAACL-HLT 2018.

Soto & Hirschberg (2018): Joint POS and Language ID Tagging

Demonstrated that jointly predicting language ID and POS improves both tasks for code-switched text. Adding language ID as a feature to the POS tagger yields significant accuracy gains, because knowing the language constrains the set of valid POS tags.

Relevance to our work: We effectively have perfect language ID (the @s marker in CHAT), which is a stronger signal than predicted LID. Our approach exploits this by routing @s words to the correct language-specific model. Soto & Hirschberg’s finding validates that language-aware POS tagging outperforms language-agnostic approaches.

Citation: Soto, V. & Hirschberg, J. (2018). Joint Part-of-Speech and Language ID Tagging for Code-Switched Data. Proceedings of the Third Workshop on Computational Approaches to Code-Switching.

LinCE Benchmark (Aguilar et al., 2020)

A centralized benchmark for code-switching NLP evaluation, covering four language pairs (Spanish-English, Nepali-English, Hindi-English, MSA-Egyptian Arabic) and four tasks (language identification, NER, POS tagging, sentiment analysis). Established that multilingual BERT underperforms specialized approaches on code-switching tasks.

Relevance to our work: LinCE confirms that code-switching POS tagging is a recognized benchmark task with established baselines. Our approach could be evaluated against LinCE’s Spanish-English POS dataset for quantitative comparison, though our focus is on CHAT transcript morphology rather than social media text.

Citation: Aguilar, G., Kar, S., & Solorio, T. (2020). LinCE: A Centralized Benchmark for Linguistic Code-switching Evaluation. Proceedings of LREC 2020.

CS-ELMo (Winata et al., 2020): Transfer Learning with Morphological Clues

Extended ELMo with a position-aware attention mechanism that enhances morphological clues from character n-grams. The bottom layers of the ELMo architecture learn morphological patterns that transfer across languages, establishing state of the art on code-switching NER and POS tasks.

Relevance to our work: CS-ELMo uses subword/character-level features to capture cross-language morphological patterns. Our approach operates at the word level and doesn’t exploit subword features, this is a potential improvement direction.

Citation: Winata, G.I., Cahyawijaya, S., Lin, Z., Liu, Z., & Fung, P. (2020). From English to Code-Switching: Transfer Learning with Strong Morphological Clues. Proceedings of ACL 2020.

“Parsing the Switch”: BiLingua Parser (2025)

Very recent work using GPT-4 with linguistically-informed prompting to produce UD annotations for code-switched text. The BiLingua Parser combines few-shot LLM prompting with expert review. LLM-based annotations outperform conventional parsers in syntactic accuracy, particularly at switch points where monolingual models typically fail.

Relevance to our work: Confirms that monolingual parsers fail at switch points, exactly the problem our secondary dispatch addresses. Their LLM-based approach is more powerful but requires API access to proprietary models, while our approach uses open Stanza models locally.

Citation: (2025). Parsing the Switch: LLM-Based UD Annotation for Complex Code-Switched and Low-Resource Languages. Findings of EMNLP 2025.

CHILDES MOR Program (Sagae et al., 2010; MacWhinney, 2012)

The direct predecessor to our work within TalkBank. The MOR program provides automatic morphological analysis for CHILDES transcripts using hand-crafted rule-based grammars. For bilingual corpora, separate MOR grammars for each language are applied. MOR achieves 98% accuracy on adult English corpora and 97% on child language.

Relevance to our work: Our L2 morphotag feature is the neural successor to MOR’s bilingual capability. MOR uses hand-crafted lexicons (limited vocabulary, requires manual maintenance); we use Stanza’s neural models (open vocabulary, no manual lexicon needed). The CHAT @s marker convention that we exploit was designed for MOR’s bilingual processing.

Citations:

• Sagae, K., Davis, E., Lavie, A., MacWhinney, B. & Wintner, S. (2010). Morphosyntactic annotation of CHILDES transcripts. J. Child Language.
• MacWhinney, B. (2012). Morphosyntactic Analysis of the CHILDES and TalkBank Corpora. Proceedings of LREC 2012.

Code-Switching UD Treebanks

Several code-switching treebanks exist in the Universal Dependencies framework, providing gold-standard annotations for evaluation:

• UD Hindi-English HIENCS (Bhat et al., 2018), Hindi-English code-switching tweets
• UD Turkish-German SAGT (Cetinoglu, 2022), Turkish-German conversational code-switching

These could serve as evaluation resources if we wanted to quantitatively validate our approach against gold-standard code-switching annotations.

Matrix Language Frame Model (Myers-Scotton, 1993)

The dominant linguistic theory of code-switching. Defines the “matrix language” (the language providing the grammatical frame) and the “embedded language” (the language contributing inserted elements). The Morpheme Order Principle and System Morpheme Principle govern how elements from the two languages combine.

Relevance to our work: Our architecture aligns naturally with the MLF model. The primary language in batchalign3 IS the matrix language (providing the syntactic frame via dependency parsing). The @s words ARE embedded language elements. Our merge algorithm respects the matrix language’s structural frame (deprel, head) while filling in the embedded language’s morphology (lemma, features), this is essentially a computational implementation of the MLF’s Morpheme Order Principle.

Situating Our Approach

What We Share with Prior Work

What Appears Novel

1. Deprel as cross-linguistic POS constraint. Using the primary model’s UD dependency relation to constrain the secondary model’s POS tag, without any code-switching training data. The UD deprel is cross-linguistically valid by design (a word with deprel advmod must be an adverb in any language), but this property has not been exploited for code-switching POS tagging in the literature.

2. Closed-class function word override. When the secondary model returns a closed-class POS (DET, ADP, SCONJ, CCONJ, AUX, PART, PRON), trusting it over the structural constraint. This heuristic addresses the specific failure mode where the primary model mislabels foreign function words (e.g., parsing Spanish los as English proper noun “Los”). Not described in prior work.

3. GRA deprel correction from resolved POS. Upgrading the primary model’s structural parse (e.g., FLAT to DET, OBL to DET) based on the resolved POS. Prior work on code-switching dependency parsing focuses on building dedicated CS parsers; we post-correct the monolingual parser’s output.

4. Zero code-switching training data requirement. Most approaches (Solorio 2008, Bhat 2018, CS-ELMo 2020) require annotated code-switching data for training or fine-tuning. Our approach needs only standard monolingual Stanza models plus a hand-crafted deprel-to-POS mapping table (a pure function with ~30 lines).

5. Integration with CHAT/TalkBank ecosystem. The @s marker provides reliable word-level language identification without any LID model. This is a unique advantage of the CHAT annotation framework that prior computational work on code-switching does not have access to (they must predict LID from text alone).

Known Limitations Relative to Prior Art

1. No subword features. CS-ELMo (2020) shows that character-level morphological features transfer across languages and improve CS POS tagging. Our approach operates purely at the word level.

2. No joint training. Approaches like neural stacking (Bhat 2018) learn to handle switch-point phenomena from code-switching data. Our rule-based merge cannot learn from data.

3. Primary model’s structural parse may be unreliable. The 2025 BiLingua paper confirms that monolingual parsers degrade at switch points. Our function-word override addresses the most common failure mode, but content words at switch boundaries may still get wrong structural analyses.

4. No evaluation on standard CS benchmarks. We have evaluated on TalkBank bilingual data but not on LinCE or other standard code-switching benchmarks. Quantitative comparison with prior work would require evaluation on shared datasets.

Recommendations

1. Our approach is viable and reasonably novel. We are not reinventing the wheel, the two-model combination idea is well established (Solorio 2008). But our specific implementation (deprel constraint + closed-class override + zero CS training data) is a new combination of known principles.

2. Cite Solorio 2008 and Bhat 2018 as the key antecedents. Our approach can be described as “a rule-based structural merge variant of the two-model pipeline approach (Solorio 2008), exploiting UD’s cross-linguistic deprel semantics for zero-shot code-switching morphosyntax.”

3. Consider evaluation on HIENCS (Hindi-English CS UD treebank) to quantitatively validate our structural merge against gold-standard code-switching annotations.

4. The Matrix Language Frame alignment should be mentioned in any publication, our architecture is a natural computational realization of the MLF model’s matrix/embedded language distinction.

Related

• Experimental L2 Morphotag, design document for the feature
• L2 & Language Switching, current behavior reference

This page last changed: 2026-06-19 (commit c82a6d03). The whole book last changed: 2026-09-16 (commit 34d249d8).