Computational Model
Computational Model
Embedded Files
Proficiency-Aware Adaptation and Data Augmentation for Robust L2 ASR
General-purpose automatic speech recognition (ASR) underperforms for atypical speakers, such as second-language (L2) learners, reinforcing bias and limiting use in education and accessibility. Using the Common European Framework of Reference for Languages (CEFR)-graded Speak & Improve corpus, we show that naïve fine-tuning of Whisper reduces the average word error rate (WER) but simultaneously widens performance disparities and disproportionately harms lower-proficiency learners. To address this, we propose two strategies: (i) proficiency-aware multitask learning, jointly optimizing ASR with proficiency classification, and (ii) targeted augmentation, applying spectrogram masking to low-proficiency speech to counter imbalance. These approaches reduce WER by up to 29.4% (relative) and insertion/deletion errors by as much as 58.6% (relative). Crucially, despite the dataset’s severe imbalance reflecting real-world distributions, both strategies consistently narrow proficiency gaps, advancing equitable ASR for L2 learners.
CtrlAct: Grounding LLMs to Bridge the Gap between Embodied Instruction and Action
Large Language Models (LLMs) show strong natural language understanding but often fail in embodied AI settings that demand physical validity and causal reasoning. We evaluate open-source models across four tasks: Goal Interpretation, Subgoal Decomposition, Action Sequencing, and Transition Modeling. Our analysis shows that LLMs often generate full action sequences without considering intermediate environmental feedback, which leads to runtime failures. To address this issue, we encode physical constraints as declarative rules in system prompts and applies Supervised Fine-Tuning (SFT) to align the model with domain dynamics. These interventions improve physical validity, but their effectiveness varies by task. This study clarifies how prompt engineering and SFT affect embodied performance, revealing both the capabilities and the persistent constraints of current open-source models.
Debiasing Static Embeddings for Hate Speech Detection
This project looks at how hidden biases in language models can influence hate speech detection systems. Specifically, it explores two common debiasing techniques—hard-debiasing and soft-debiasing—to see how they affect both the representation of social stereotypes in embedding spaces and the overall fairness of detection outcomes. A key focus is on whether these methods help reduce the over-flagging of content from marginalized voices, while still accurately identifying harmful speech targeted at these groups. The work also examines how well standard bias metrics align with actual system behavior in downstream tasks like hate speech classification.
Quantum NLP model on Natural Language Inference
This project applies quantum circuits to natural language inference, showing that quantum models can match classical baselines while using far fewer parameters. We introduce a novel metric to measure learning efficiency and find that quantum models learn with far greater efficiency per parameter—sometimes by several orders of magnitude—making them well-suited for data-scarce, structure-aware scenarios. A new cluster-based architecture also improves generalization by sharing parameters across related words.
A Neural Model Approach to Individual Differences in Syntactic Adaptation
Syntactic adaptation refers to the process where sentence comprehension andproduction are aligned with syntactic structures in recent input. Building upon previous neural modeling research, we introduce a novel neural model method capable of simulating individual responses duringadaptation and capturing their variances.
This paper investigates the application of Shengzhou dialect tone sandhi rules in the context of sprechstimme in Shaoxing opera. Our study utilizes a corpus of Shaoxing sprechstimme, where we closely examine the tone realization of one lexical item in various contexts. Surprisingly, our findings reveal that the tone realization patterns observed in the corpus cannot be fully explained by the Shengzhou tone sandhi system. As a result, we suggest potential solutions to address the observed discrepancies.
Recommender systems are widely studied for creating user-friendly platforms (e.g., online shopping, videostreaming, playlist generating). Although many researchershave used machine learning to increase model performance, not much attention has been paid to howindividual users are affected by order-sensitive contexts.
Previous approaches to studying sequence effects can be expensive and difficult to understand. This paperlooks into whether case-based reasoning systems incorporatedwith dynamical system techniques can model context effects for individual users. It aims to introducean explainable model for addressing sequence-sensitive context effects with better transparency and efficiency.
Previous approaches to studying sequence effects can be expensive and difficult to understand. This paperlooks into whether case-based reasoning systems incorporatedwith dynamical system techniques can model context effects for individual users. It aims to introducean explainable model for addressing sequence-sensitive context effects with better transparency and efficiency.
This paper investigates the contextual dependency of relative adjectives (e.g., long, short) and their cognitive underpinnings, arguing that such language behavior is grounded in cognition rather than purely language-internal mechanisms. By analyzing cross-linguistic data and comparing relative adjectives to relational nouns, the paper suggests that these adjectives exhibit universal traits indicative of an underlying cognitive factor. The study adopts a situated, embodied, and dynamic (SED) approach, evaluating existing SED models on relational categories and proposing a pilot study to model communication of relational categories like "long" and "short" with continuous-time recurrent neural network (CTRNN) models.
Following Qing and Franke (2014), the current study simulates experiment data of a subject labelling task. It then (1) repeats Qing and Franke’s (2014) Bayesian model (henceforth, QF model) prediction, (2) presents a modified QF model with continuous prior (henceforth, continuous QF), (3) presents a modified continuous QF model with competing reference points (henceforth, multi-reference QF), and (4) compared the prediction of the three models with the simulated experiment data.
Experiment
Experiment
Does Precision Affect Categorization? Magnitude Categorization and Measurement Scales
This study explores the impact of precision on magnitude categorization and measurement scales in linguistic contexts. Using a novel experimental design, participants were introduced to a morph sequence with novel labels and trained on degree measurements under high-precision and low-precision conditions. The study aimed to investigate whether the precision of numerical information affects participants' ability to learn, generalize, and communicate about relative magnitudes. Results from various testing blocks, including XAB, differentiation, and equally tasks, were analyzed to assess the influence of precision on participants' categorical judgments and their application of linguistic labels. The findings offer insights into how precision in communication shapes perception and categorization, with implications for understanding numerical cognition and linguistic relativity.
From Distributional Structure to Meaning: Learning, Syntax, and the Emergence of Categorical and Graded Interpretations
Using novel adjectives mapped onto an arbitrary morph space with both discrete and continuous variance, we manipulate learning task and test generalization across syntactic frames in English and Mandarin. Learning task systematically shifts alignment between adjectives and perceptual dimensions. This alignment effect is amplified in Mandarin, where syntax overtly distinguishes categorical from graded predication. Critically, alignments do not reconfigure across post-learning tasks: category assignments remain stable, while graded information surfaces selectively in intensifier choice.
Formal Logic Model
Formal Logic Model
(Presented at ESSLLI 2024)
Traditional degree semantics approaches have aimed to pin down the inherent and static class of adjectives. This paper presents a novel dynamic perspective, where the classification of an adjective is dy- namic and syntactically dependent. Using measurement theory and fuzzy set analysis, the proposed framework defines dynamic patterns of adjec- tive classes with a set of axioms and integrates these patterns with syn- tactic structures to explain the flexibility and constraints observed in ad- jectival expressions. Employing Mandarin data, the paper illustrates how different syntactic constructions select specific adjective classes, thereby affecting their interpretation and distribution. This approach not only accommodates cross-linguistic variations but also provides a more com- prehensive understanding of the semantics of adjectives.
Traditional degree semantics approaches have aimed to pin down the inherent and static class of adjectives. This paper presents a novel dynamic perspective, where the classification of an adjective is dy- namic and syntactically dependent. Using measurement theory and fuzzy set analysis, the proposed framework defines dynamic patterns of adjec- tive classes with a set of axioms and integrates these patterns with syn- tactic structures to explain the flexibility and constraints observed in ad- jectival expressions. Employing Mandarin data, the paper illustrates how different syntactic constructions select specific adjective classes, thereby affecting their interpretation and distribution. This approach not only accommodates cross-linguistic variations but also provides a more com- prehensive understanding of the semantics of adjectives.
Much research has shown that measure phrases elicit an asymmetry in acceptability between positive vs. negative members of gradable adjective antonym pairs (3m tall vs. *3m short) (Kennedy 1999, Schwarzschild 2005, and others). As in other languages, the measure phrase in Mandarin yi mi “1 meter” is accepted as direct modifier of gao “tall” and rejected by the negative antonym ai “short”. However, when the adjective combines with name "so", this measure phrase is accepted by both positive and negative adjectives. In addition, name adjective expressions are ambiguous between two readings. Namely, the surprise reading and equative reading. This paper accounts for the new data with two proposals: (1) name is a presupposition filter; (2) There is a silent morpheme Vec in name adjective expressions, which is in complementary distribution with overt measure phrases.
This paper presents new data from Mandarin and German showing that some negative adjectives can license measure phrases (0.4 mm bao'0.4 mm thin') regardless of surface positions. This observation challenges existing work on measure phrase constructions such as Schwarzschild (2005) and Breakstone (2012), which argue that measure phrases cannot directly modify negative adjectives. This paper offers a novel analysis of measure phrase constructions. I propose (i) measure phrases are inherently downward monotonic, and this polarity is sensitive to contextual and lexical information;(ii) measure phrases denote the maximum value of the set of degrees. This analysis explains the markedness of negative adjective in measure phrase constructions crosslinguistically, along with the variation of grammatical judgement in some languages and contexts, and among different speakers.
Other theoretical work
Other theoretical work
L Sun, in press. The role of polarity in measure phrase constructions. Proc of NACCL-32.
A O’Neil, L Sun, Z Liu, manuscript. Implementing Dependency Grammar for ditransitive constructions: Data from English, Mandarin and Zulu
L Sun, manuscript. Neurological and Cognitive Reality of tall and short
L Sun, manuscript. “He is tall”: A typological survey on predicative adjectives
L Sun, manuscript. Tone System and Tone Sandhi in Nanjing Mandarin
Page updated
Google Sites
Report abuse