Pre-trained code models are essential for various code intelligence tasks. Yet, their effectiveness is heavily influenced by the quality of the pre-training dataset, particularly human-written reference comments, which usually serve as a bridge between the programming language and natural language. One significant challenge is that such comments could become inconsistent with the corresponding code as the software evolves, leading to suboptimal model performance. Large language models (LLMs) have demonstrated superior capabilities in generating high-quality code comments. This work investigates whether substituting original human-written comments with LLM-generated ones can improve pre-training datasets for more effective pretrained code models. As existing reference-based metrics cannot evaluate the quality of human-written reference comments themselves, to enable direct comparison between LLM-generated and human reference comments, we introduce two auxiliary tasks as novel reference-free metrics, including code-comment inconsistency detection and semantic code search. Experimental results show that LLM-generated comments exhibit superior semantic consistency with the code compared to human-written reference comments. Our manual evaluation also corroborates this conclusion, which indicates the potential of utilizing LLMs to enhance the quality of the pre-training dataset. Based on this finding, we rebuilt the CodeSearchNet dataset with LLM-generated comments and re-pre-trained the CodeT5 model. Evaluations on multiple code intelligence tasks demonstrate that models pretrained by LLM-enhanced data outperform their counterparts (pre-trained by original human reference comments data) on code summarization, code generation, and code translation tasks. This research validates the feasibility of rebuilding the pre-training dataset by LLMs to advance code intelligence tasks. It advocates rethinking the reliance on human reference comments for coderelated tasks.

Transformers are now widely utilized in code intelligence tasks. To better fit highly structured source code, various structure information is passed into Transformer, such as positional encoding and abstract syntax tree (AST) based structures. However, it is still not clear how these structural features affect code intelligence tasks, such as code summarization. Addressing this problem is of vital importance for designing Transformer-based code models. Existing works are keen to introduce various structural information into Transformers while lacking persuasive analysis to reveal their contributions and interaction effects. In this paper, we conduct an empirical study of frequently-used code structure features for code representation, including two types of position encoding features and AST-based structure features. We propose a couple of probing tasks to detect how these structure features perform in Transformer and conduct comprehensive ablation studies to investigate how these structural features affect code semantic summarization tasks. To further validate the effectiveness of code structure features in code summarization tasks, we assess Transformer models equipped with these code structure features on a structural dependent summarization dataset. Our experimental results reveal several findings that may inspire future study: (1) there is a conflict between the influence of the absolute positional embeddings and relative positional embeddings in Transformer; (2) AST-based code structure features and relative position encoding features show a strong correlation and much contribution overlap for code semantic summarization tasks indeed exists between them; (3) Transformer models still have space for further improvement in explicitly understanding code structure information.

Summary We present EcForest, an extractive summarization model through Enhanced Sentence Embedding and Cascade Forest. Sentence representation is of great significance for many summarization methods. Bag‐of‐words mostly fails to grasp the semantics, and typical embedding models cannot capture more complex semantic features, such as polysemy and the meaning of a phrase, which is usually ignored by simply averaging the word embeddings included in a sentence. To this end, we propose Enhanced Sentence Embedding (ESE) model to solve such drawbacks via mapping several valid features to dense vectors. Essentially, the enhanced sentence embedding is a novel model for improving the distributed representation of sentence. Our sentence embedding model is universally applicable and it can be adapted to other NLP tasks. Moreover, deep forest is used as a sentence extraction algorithm for its robustness to the hyper‐parameters and its efficient training algorithm compared to deep neural network. The evaluation of variant models proposed in this work proves the validation of the enhanced sentence embedding. The comparison results between EcForest and several baselines on two different datasets demonstrate that the proposed summarization model performs better than or with high competitiveness to the state‐of‐the‐art.

In the era of information overload, text summarization has become a focus of attention in a number of diverse fields such as, question answering systems, intelligence analysis, news recommendation systems, search results in web search engines, and so on. A good document representation is the key point in any successful summarizer. Learning this representation becomes a very active research in natural language processing field (NLP). Traditional approaches mostly fail to deliver a good representation. Word embedding has proved an excellent performance in learning the representation. In this paper, a modified BM25 with Word Embeddings are used to build the sentence vectors from word vectors. The entire document is represented as a set of sentence vectors. Then, the similarity between every pair of sentence vectors is computed. After that, TextRank, a graph-based model, is used to rank the sentences. The summary is generated by picking the top-ranked sentences according to the compression rate. Two well-known datasets, DUC2002 and DUC2004, are used to evaluate the models. The experimental results show that the proposed models perform comprehensively better compared to the state-of-the-art methods.