The increasing use of foundation models highlights the urgent need to address and eliminate implicit biases present in them that arise during pretraining. In this paper, we introduce PEFTDebias, a novel approach that employs parameter-efficient fine-tuning (PEFT) to mitigate the biases within foundation models. PEFTDebias consists of two main phases: an upstream phase for acquiring debiasing parameters along a specific bias axis, and a downstream phase where these parameters are incorporated into the model and frozen during the fine-tuning process. By evaluating on four datasets across two bias axes namely gender and race, we find that downstream biases can be effectively reduced with PEFTs. In addition, we show that these parameters possess axis-specific debiasing characteristics, enabling their effective transferability in mitigating biases in various downstream tasks.

Since the rise of neural natural-language-to-code models (NL→Code) that can generate long expressions and statements rather than a single next-token, one of the major problems has been reliably evaluating their generated output. In this paper, we propose CodeBERTScore: an evaluation metric for code generation, which builds on BERTScore (Zhang et al., 2020). Instead of encoding only the generated tokens as in BERTScore, CodeBERTScore also encodes the natural language input preceding the generated code, thus modeling the consistency between the generated code and its given natural language context as well. We perform an extensive evaluation of CodeBERTScore across four programming languages. We find that CodeBERTScore achieves a higher correlation with human preference and with functional correctness than all existing metrics. That is, generated code that receives a higher score by CodeBERTScore is more likely to be preferred by humans, as well as to function correctly when executed. We release five language-specific pretrained models to use with our publicly available code. Our language-specific models have been downloaded more than **1,000,000** times from the Huggingface Hub. Our code and data are available at https://github.com/neulab/code-bert-score

In this paper, we present our submission to the DialDoc shared task based on the MultiDoc2Dial dataset. MultiDoc2Dial is a conversational question answering dataset that grounds dialogues in multiple documents. The task involves grounding a user’s query in a document followed by generating an appropriate response. We propose several improvements over the baseline’s retriever-reader architecture to aid in modeling goal-oriented dialogues grounded in multiple documents. Our proposed approach employs sparse representations for passage retrieval, a passage re-ranker, the fusion-in-decoder architecture for generation, and a curriculum learning training paradigm. Our approach shows a 12 point improvement in BLEU score compared to the baseline RAG model.

Videos of group interactions contain a wealth of information beyond the information directly communicated in a transcript of the discussion. Tracking who has participated throughout an extended interaction and what each of their trajectories has been in relation to one another is the foundation for joint activity understanding, though it comes with some unique challenges in videos of tightly coupled group work. Motivated by insights into the properties of such scenarios, including group composition and the properties of task-oriented, goal directed tasks, we present a successful proof-of-concept. In particular, we present a transfer experiment to a dyadic robot construction task, an ablation study, and a qualitative analysis.

Code-switched (CS) data is ubiquitous in today’s globalized world, but the dearth of annotated datasets in code-switching poses a significant challenge for learning diverse tasks across different language pairs. Parameter-efficient prompt-tuning approaches conditioned on frozen language models have shown promise for transfer learning in limited-resource setups. In this paper, we propose a novel instance-based prompt composition technique, PRO-CS, for CS tasks that combine language and task knowledge. We compare our approach with prompt-tuning and fine-tuning for code-switched tasks on 10 datasets across 4 language pairs. Our model outperforms the prompt-tuning approach by significant margins across all datasets and outperforms or remains at par with fine-tuning by using just 0.18% of total parameters. We also achieve competitive results when compared with the fine-tuned model in the low-resource cross-lingual and cross-task setting, indicating the effectiveness of our approach to incorporate new code-switched tasks.

People speaking different kinds of languages search for information in a cross-lingual manner. They tend to ask questions in their language and expect the answer to be in the same language, despite the evidence lying in another language. In this paper, we present our approach for this task of cross-lingual open-domain question-answering. Our proposed method employs a passage reranker, the fusion-in-decoder technique for generation, and a wiki data entity-based post-processing system to tackle the inability to generate entities across all languages. Our end-2-end pipeline shows an improvement of 3 and 4.6 points on F1 and EM metrics respectively, when compared with the baseline CORA model on the XOR-TyDi dataset. We also evaluate the effectiveness of our proposed techniques in the zero-shot setting using the MKQA dataset and show an improvement of 5 points in F1 for high-resource and 3 points improvement for low-resource zero-shot languages. Our team, CMUmQA’s submission in the MIA-Shared task ranked 1st in the constrained setup for the dev and 2nd in the test setting.