Fix-Filter-Fix: Intuitively Connect Any Models for Effective Bug Fixing

Hong, Haiwen; Zhang, Jingfeng; Zhang, Yin; Wan, Yao; Sui, Yulei

doi:10.18653/v1/2021.emnlp-main.282

Cited by 5 publications

(2 citation statements)

References 18 publications

(17 reference statements)

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Designing novel training objectives, investigating if such tradeoffs occur, and if they can be mitigated, are all ample avenues for future work. Some particularly interesting training objectives to investigate include ones that model program paths, such as Flow2Vec [68], that leverage contrastive learning such as ContraFlow [69], or approaches that combine multiple models such as Fix-Filter-Fix [70]. However some of these approaches would need INSPECT to be extended.…”

Section: Performance Of Decodersmentioning

confidence: 99%

What do pre-trained code models know about code?

Karmakar

Robbes

2021

2021 36th IEEE/ACM International Conference on Automated Software Engineering (ASE)

View full text Add to dashboard Cite

Pre-trained models of source code have recently been successfully applied to a wide variety of Software Engineering tasks; they have also seen some practical adoption in practice, e.g. for code completion. Yet, we still know very little about what these pre-trained models learn about source code. In this article, we use probing-simple diagnostic tasks that do not further train the models-to discover to what extent pre-trained models learn about specific aspects of source code. We use an extensible framework to define 15 probing tasks that exercise surface, syntactic, structural and semantic characteristics of source code. We probe 8 pre-trained source code models, as well as a natural language model (BERT) as our baseline. We find that models that incorporate some structural information (such as GraphCodeBERT) have a better representation of source code characteristics. Surprisingly, we find that for some probing tasks, BERT is competitive with the source code models, indicating that there are ample opportunities to improve source-code specific pre-training on the respective code characteristics. We encourage other researchers to evaluate their models with our probing task suite, so that they may peer into the hidden layers of the models and identify what intrinsic code characteristics are encoded.

show abstract

Section: Performance Of Decodersmentioning

confidence: 99%

What do pre-trained code models know about code?

Karmakar

Robbes

2021

2021 36th IEEE/ACM International Conference on Automated Software Engineering (ASE)

View full text Add to dashboard Cite

show abstract

“…F 3 [21] proposes an intuitive yet effective general framework to concatenate different learners with the filter mechanism to filter out unchanged fixes. Flow2Vec [22] presents a new code embedding approach that preserves interprocedural, context-sensitive and alias-aware value-flows in the low-dimensional vector space to better support subsequent learning tasks. These methods exhibit better feature learning in shared and task-specific representations and achieve better performances than typical MTL methods.…”

Section: Multi-task Learning (Mtl)mentioning

confidence: 99%

Multi-Task Deep Learning with Task Attention for Post-Click Conversion Rate Prediction

Luo¹,

Zhou²,

Ding³

et al. 2023

Intelligent Automation &Amp; Soft Computing

View full text Add to dashboard Cite

Online advertising has gained much attention on various platforms as a hugely lucrative market. In promoting content and advertisements in real life, the acquisition of user target actions is usually a multi-step process, such as impres-sion→click→conversion, which means the process from the delivery of the recommended item to the user's click to the final conversion. Due to data sparsity or sample selection bias, it is difficult for the trained model to achieve the business goal of the target campaign. Multi-task learning, a classical solution to this problem, aims to generalize better on the original task given several related tasks by exploiting the knowledge between tasks to share the same feature and label space. Adaptively learned task relations bring better performance to make full use of the correlation between tasks. We train a general model capable of capturing the relationships between various tasks on all existing active tasks from a meta-learning perspective. In addition, this paper proposes a Multi-task Attention Network (MAN) to identify commonalities and differences between tasks in the feature space. The model performance is improved by explicitly learning the stacking of task relationships in the label space. To illustrate the effectiveness of our method, experiments are conducted on Alibaba Click and Conversion Prediction (Ali-CCP) dataset. Experimental results show that the method outperforms the state-of-the-art multi-task learning methods.

show abstract