TL;DR
This paper investigates how the content and communicative function of speech influence code-switching (CSW) in Spanish-English bilinguals. By analyzing named entities and dialogue acts, they found that communicative function plays a more significant role than content in shaping CSW patterns, offering new insights for building more naturalistic speech technologies.
What Problem Does This Paper Address?
Code-switching (CSW), the alternation between languages within a conversation, is a common phenomenon among bilinguals. While previous research has explored syntactic, morphological, and even paralinguistic factors influencing CSW, there's a significant gap in understanding how discourse-level aspects—specifically, what is being said (content) and why (function)—impact this linguistic behavior. This paper aims to bridge that gap by systematically analyzing the relationship between discourse content and function and CSW patterns in spontaneous Spanish-English speech. Understanding these dynamics is crucial for developing better tools for speech recognition, synthesis, and translation that can handle the complexities of multilingual communication.
Methods
The researchers employed a multi-pronged approach:
- Data Annotation: They meticulously annotated a Spanish-English spontaneous speech corpus (Bangor Miami corpus) with:
- Named Entities (NEs): Categorized into types like people (N), organizations (O), places (P), brands (B), temporal expressions (T), religious figures (R), and others (U).
- Dialogue Acts (DAs): Using the Switchboard Dialogue Acts (SwDA) tag set to capture the communicative function of utterances.
- CSW Metrics: Calculating CSW ratio, M-index, and I-index to quantify code-switching frequency and intensity.
- Statistical Analysis: They conducted chi-squared tests and odds ratio calculations to identify statistically significant associations between NE types, DA types, and CSW patterns. They also analyzed dependency distances between NEs and code-switched segments.
- Predictive Modeling: To assess the practical utility of their findings, they built two transition-based predictive models for CSW:
- Logistic Regression: For utterance-level predictions, focusing on local cues.
- Hidden Markov Model (HMM): For dialogue-level predictions, capturing temporal dependencies across turns. They performed ablation studies, systematically adding NE features and DA features to baseline models to measure their impact on prediction accuracy.
Datasets
The primary dataset used in this study is the Bangor Miami (BM) corpus. This corpus consists of 35 hours of spontaneous, informal Spanish-English conversations, comprising 46.9k transcribed utterances from 84 unique speakers. The original corpus was already annotated with token-level language identification (LID) tags. The authors then added their own annotations for Named Entities and Dialogue Acts.
Key Insights and Ideas
- Discourse Function > Content: The most striking finding is that the communicative function of an utterance (as captured by Dialogue Acts) has a greater influence on code-switching patterns than the content of the utterance (as captured by Named Entities). This suggests that why someone is speaking is a stronger driver of language choice than what they are talking about.
- NEs as "Anchors," Not Triggers: Contrary to what one might expect, Named Entities (like names of people or places) in this corpus tend to discourage code-switching rather than trigger it. This might be because NEs are often specific and might have established language preferences within the context.
- Task-Oriented vs. Exploratory Speech: Code-switched speech in the BM corpus leans towards more task-oriented, structured exchanges, focusing on clarity and goal completion. Monolingual speech, on the other hand, is more varied, incorporating feedback and exploratory acts.
- CSW Quantity and Function: Certain dialogue acts, like acknowledgments and backchannels, are associated with greater quantities of CSW, even though these DAs were more likely to be expressed in monolingual utterances. This suggests a compensatory mechanism where more complex CSW might be used to convey simpler conversational functions.
- Syntactic Structure and Function: The type of code-switching (insertional vs. alternational) is influenced by dialogue acts. For instance, self-talk is more likely to be insertional CSW, while quotations are always alternational, reflecting how different discourse acts naturally map onto different syntactic structures.
Key Results
- NE Usage: Only 9.8% of utterances contained NEs, with code-switched contexts showing slightly fewer NEs than monolingual ones.
- NE Type Influence: Temporal expressions (T-type NEs) were significantly more likely to appear in code-switched utterances, while person (N) and brand (B) NEs were less likely.
- DA Influence: Statements of opinion (sv-fx) and self-directed comments (t1) were about 1.3 times more likely to be expressed in code-switched utterances. Conversely, acknowledgments (bk), questions (qw), and non-verbal acts (x, %) were less likely to involve CSW.
- Predictive Model Performance: Incorporating DA features improved CSW prediction accuracy for both logistic regression and HMM models. NE features, however, slightly worsened performance when combined with DA features in the HMM model.
- Logistic Regression: Baseline accuracy 0.613, Baseline + DA accuracy 0.638.
- HMM: Baseline accuracy 0.990, Baseline + DA accuracy 0.991.
- These improvements, while small in absolute terms, were statistically significant.
Limitations and Future Work
- Single Language Pair and Corpus: The study is limited to Spanish-English and the Bangor Miami corpus, which may be skewed towards English. The authors acknowledge the need to extend their methods to other language pairs and cultural contexts to test robustness and generalizability.
- Annotator Proficiency: While efforts were made to ensure label quality, the use of a single annotator for NEs and DAs (though with spot checks) is noted as a potential limitation.
- Translation for Dependency Parsing: The analysis of dependency distance relied on automatic translation to English, which could introduce minor inaccuracies.
- Interaction of Content and Function: The paper briefly touches on the interaction between discourse content and function but suggests more detailed investigation is needed, potentially with multimodal architectures.
- Future Work: The authors plan to explore the interaction between content and function more deeply, experiment with multimodal architectures, and replicate their findings on other CSW datasets and language pairs. They also aim to inform the generation of more authentic, discourse-motivated bilingual speech.
Summary updated