logseq/pages/@TSE-2025-01-0025_Proof_hi.md

tags:: [[#zotero]]
title:: @TSE-2025-01-0025_Proof_hi
item-type:: [[document]]
original-title:: TSE-2025-01-0025_Proof_hi
language:: en
links:: [Local library](zotero://select/library/items/TNT49ESZ), [Web library](https://www.zotero.org/users/1039502/items/TNT49ESZ)

- ### Attachments
	- [Responses_to_Reviewer](zotero://select/library/items/RCSRLPWZ) {{zotero-imported-file RCSRLPWZ, "Responses_to_Reviewer.pdf"}}
	- [PDF](zotero://select/library/items/PFCQPJV8) {{zotero-imported-file PFCQPJV8, "TSE-2025-01-0025_Proof_hi.pdf"}}
- ### Notes
	- # Annotazioni 
	   (11/4/2025, 16:52:40)
	  
	  - “Software researchers and industry experts are increasingly focused on automating the task of requirement analysis.” (“TSE-2025-01-0025_Proof_hi”, p. 1) #5fb236
	  * *
	  
	   
	  
	  - “While numerous methodologies and tools exist to automate the translation of user requirements into UML diagrams, practical usability remains challenging due to their inherent complexity” (“TSE-2025-01-0025_Proof_hi”, p. 1) #5fb236
	  * *
	  
	   
	  
	  - “This research addresses the challenge of converting textual software requirements into UML class diagrams.” (“TSE-2025-01-0025_Proof_hi”, p. 1) #a28ae5
	  * *
	  
	   
	  
	  - “An intermediate layer is introduced, representing diagram elements as statements in the PlantUML language” (“TSE-2025-01-0025_Proof_hi”, p. 1) #a28ae5
	  * *
	  
	   
	  
	  - “The study employs the natural language processing Transformer (NLP Transformer) as a Neural Machine Translation (NMT) tool, exploring two approaches: sequence-to-sequence (Seq2Seq) and sequence-to-Abstract-Syntax-Tree (Seq2Ast).” (“TSE-2025-01-0025_Proof_hi”, p. 1) #a28ae5
	  * *
	  
	   
	  
	  - “The Seq2Seq approach demonstrates gains in precision, recall, and over-specification at 74.16%, 81.95%, and 28.91%, respectively” (“TSE-2025-01-0025_Proof_hi”, p. 1) #a28ae5
	  * *
	  
	   
	  
	  - “Seq2Ast approach achieves higher values, with precision, recall, and over-specification at 86.19%, 90.81%, and 33.38%.” (“TSE-2025-01-0025_Proof_hi”, p. 1) #a28ae5
	  * *
	  
	   
	  
	  - “dramatic” (“TSE-2025-01-0025_Proof_hi”, p. 1) #ff6666
	  * *
	  
	   
	  
	  - “Companies and organizations are investing heavily in AI research and development, with applications being developed in fields such as robotics, manufacturing, agriculture, and logistics.” (“TSE-2025-01-0025_Proof_hi”, p. 1) #5fb236
	  * *
	  
	   
	  
	  - “Machine learning (ML) and deep learning (DL) methods can be employed during the requirements-gathering phase [7].” (“TSE-2025-01-0025_Proof_hi”, p. 1) #5fb236
	  * *
	  
	   
	  
	  - “In the subsequent phase of the SDLC, the design stage, ML techniques prove beneficial by automating diagram creation, thus reducing time and effort consumption” (“TSE-2025-01-0025_Proof_hi”, p. 1) #5fb236
	  * *
	  
	   
	  
	  - “This requirement analysis process is resource-intensive, demanding time and expertise to establish the foundational rules for software development.” (“TSE-2025-01-0025_Proof_hi”, p. 1) #5fb236
	  * *
	  
	   
	  
	  - “The research’s general problem is to study the possibility of converting textual data written in natural language to the ontology domain.” (“TSE-2025-01-0025_Proof_hi”, p. 2) #5fb236
	  * *
	  
	   
	  
	  - “Within software engineering, UML diagrams act as essential ontologies, in this research specifically addressing the extraction of UML class diagram from software requirements” (“TSE-2025-01-0025_Proof_hi”, p. 2) #a28ae5
	  * *
	  
	   
	  
	  - “The study also discusses the possibility of applying the methods developed here to other types of UML diagrams.” (“TSE-2025-01-0025_Proof_hi”, p. 2) #a28ae5
	  * *
	  
	   
	  
	  - “diagrams rely on a more constrained set of visual conventions for conveying information” (“TSE-2025-01-0025_Proof_hi”, p. 2) #5fb236
	  * *
	  
	   
	  
	  - “this process of information distillation carries the inherent risk of data loss, potentially leading to unforeseen consequences downstream.” (“TSE-2025-01-0025_Proof_hi”, p. 2) #a28ae5
	  * *
	  
	   
	  
	  - “Can textual data be transformed into diagrams, and if so, what factors determine the feasibility?” (“TSE-2025-01-0025_Proof_hi”, p. 2) #2ea8e5
	  * *
	  
	   
	  
	  - “Can textual data be automatically converted into UML diagrams through existing models or established mapping techniques?” (“TSE-2025-01-0025_Proof_hi”, p. 2) #ffd400
	  *This research query is very similar to the previous one. How are they different? This needs to be clarified. *
	  
	   
	  
	  - “Linear text, composed of sentences and words, depicts situations in a sequential manner.” (“TSE-2025-01-0025_Proof_hi”, p. 2) #2ea8e5
	  * *
	  
	   
	  
	  - “can we represent a diagram as a collection of distinct parts, independent of resorting to natural language descriptions?” (“TSE-2025-01-0025_Proof_hi”, p. 2) #2ea8e5
	  * *
	  
	   
	  
	  - “Consequently, formulating a set of guidelines is pivotal in this approach, along with possessing preliminary training or knowledge.” (“TSE-2025-01-0025_Proof_hi”, p. 2) #5fb236
	  * *
	  
	   
	  
	  - “discussing generalizing the approach to the rest of the UML diagrams.” (“TSE-2025-01-0025_Proof_hi”, p. 2) #ffd400
	  * *
	  
	   
	  
	  - “The last critical question requires an intermediate layer between natural language and ontology. This layer must fit between the textual and ontology representation domains.” (“TSE-2025-01-0025_Proof_hi”, p. 2) #ffd400
	  *not clear *
	  
	   
	  
	  - “This framework aims to leverage textual software requirements as input to generate both the corresponding software code and project structures” (“TSE-2025-01-0025_Proof_hi”, p. 2) #5fb236
	  * *
	  
	   
	  
	  - “the textual data into UML diagrams, establishing a software project structure, and generating code based on the structural or behavioral diagrams and the software design patterns” (“TSE-2025-01-0025_Proof_hi”, p. 2) #5fb236
	  * *
	  
	   
	  
	  - “• Formulating the task of generating class diagrams from software requirements as a machine translation task.” (“TSE-2025-01-0025_Proof_hi”, p. 2) #2ea8e5
	  * *
	  
	   
	  
	  - “troducing the GenClass model, a class diagram generation model proficient in producing class diagrams from natural language.” (“TSE-2025-01-0025_Proof_hi”, p. 2) #2ea8e5
	  * *
	  
	   
	  
	  - “Exploring potential strategies to achieve this translation in the optimal manner.” (“TSE-2025-01-0025_Proof_hi”, p. 2) #2ea8e5
	  * *
	  
	   
	  
	  - “Our suggested method is evaluated, and the results are reported in section IV.” (“TSE-2025-01-0025_Proof_hi”, p. 2) #ffd400
	  * *
	  
	   
	  
	  - “This process aims to convert human natural language information into a format that machines can process, store, and analyze, making it more accessible for computational tasks.” (“TSE-2025-01-0025_Proof_hi”, p. 3) #5fb236
	  * *
	  
	   
	  
	  - “Automatic knowledge representation often involves the creation of knowledge graphs, ontologies, databases, or other structured data models” (“TSE-2025-01-0025_Proof_hi”, p. 3) #5fb236
	  * *
	  
	   
	  
	  - “Upadhyay and Fujii proposed a software system for knowledge support that extracts sentences and keywords from an extensive document database” (“TSE-2025-01-0025_Proof_hi”, p. 3) #5fb236
	  * *
	  
	   
	  
	  - “Artequakt project [22] connects a knowledge extraction tool with an ontology to provide continuous knowledge assistance, guiding the information extraction process in a structured manner.” (“TSE-2025-01-0025_Proof_hi”, p. 3) #5fb236
	  * *
	  
	   
	  
	  - “The semiautomated approach combines manual input and automated tools to transform textual requirements into UML diagrams.” (“TSE-2025-01-0025_Proof_hi”, p. 3) #5fb236
	  * *
	  
	   
	  
	  - “Overall, MOVA heavily relies on human intervention due to its inability to identify OO concepts autonomously” (“TSE-2025-01-0025_Proof_hi”, p. 3) #5fb236
	  * *
	  
	   
	  
	  - “The literature review highlights gaps in heuristic and pattern-based software modeling methods, which lack adaptability and semantic depth. Recent advancements suggest that treating the problem as a machine translation task can offer a promising solution. In our work, we try to employ the Transformer to solve the problem of converting textual requirements into precise class diagrams. Our approach attempts to address the limitations of heuristic methods, which are interpretable and fast but struggle in unstructured environments. While statistical models are powerful, they are often viewed as black boxes, lacking the transparency of heuristic methods.” (“TSE-2025-01-0025_Proof_hi”, p. 4) #a28ae5
	  * *
	  
	   
	  
	  - “the input is text written in English, while the output is a set of PlantUML statements.” (“TSE-2025-01-0025_Proof_hi”, p. 4) #5fb236
	  * *
	  
	   
	  
	  - “PlantUML syntax that defines an element, a relationship, or any component of a class diagram.” (“TSE-2025-01-0025_Proof_hi”, p. 4) #5fb236
	  * *
	  
	   
	  
	  - “NLP, tokenization involves breaking a sentence into smaller units to simplify assigning the most appropriate meaning to each part.” (“TSE-2025-01-0025_Proof_hi”, p. 4) #5fb236
	  * *
	  
	   
	  
	  - “with type” (“TSE-2025-01-0025_Proof_hi”, p. 5) #ffd400
	  *where is the type? Maybe you wanted to say Class with Attribute? *
	  
	   
	  
	  - “Conventionally, we will rely on data written in natural language, which will be input to the encoder, called source data, and data written in PlantUML, which is the output of the decoder, called target data” (“TSE-2025-01-0025_Proof_hi”, p. 5) #a28ae5
	  * *
	  
	   
	  
	  - “Whitespace tokenization was used for the source data” (“TSE-2025-01-0025_Proof_hi”, p. 5) #5fb236
	  * *
	  
	   
	  
	  - “Stop words removal to remove the most frequent words from the corpus records. Coreference resolution is a technique used to determine when two or more expressions in a text refer to the same entity. These expressions can be pronouns, noun phrases, or other references. It aims to link these expressions together so that we know which words or phrases in a text are talking about the same thing.” (“TSE-2025-01-0025_Proof_hi”, p. 5) #5fb236
	  * *
	  
	   
	  
	  - “Byte-Pair Encoding (BPE)” (“TSE-2025-01-0025_Proof_hi”, p. 5) #5fb236
	  * *
	  
	   
	  
	  - “Neural machine translation (NMT)” (“TSE-2025-01-0025_Proof_hi”, p. 5) #a28ae5
	  * *
	  
	   
	  
	  - “It has outperformed traditional phrase-based systems and effectively overcome issues like the requirement for manually crafted features” (“TSE-2025-01-0025_Proof_hi”, p. 5) #5fb236
	  * *
	  
	   
	  
	  - “Encoder and a Decoder: the former processes input text sequences, while the latter produces translated output sequences.” (“TSE-2025-01-0025_Proof_hi”, p. 5) #5fb236
	  * *
	  
	   
	  
	  - “attention mechanism that aligns target tokens with their corresponding source tokens” (“TSE-2025-01-0025_Proof_hi”, p. 5) #5fb236
	  * *
	  
	   
	  
	  - “we employ two prevalent methods, sequenceto-sequence (Seq2Seq) and Sequence-to-Abstract-Syntax-Tree (Seq2Ast), to assess their capacity for translation and determine their comparative effectiveness.” (“TSE-2025-01-0025_Proof_hi”, p. 5) #a28ae5
	  * *
	  
	   
	  
	  - “These sublayers include multi-head self-attention, a fully connected feed-forward network, and decoder self-attention (as multihead attention in the case of decoders)” (“TSE-2025-01-0025_Proof_hi”, p. 6) #5fb236
	  * *
	  
	   
	  
	  - “Encoder” (“TSE-2025-01-0025_Proof_hi”, p. 6) #2ea8e5
	  * *
	  
	   
	  
	  - “The encoder’s role is to progress through the input time steps and transform the entire sequence into a constant-length vector known as a context vector.” (“TSE-2025-01-0025_Proof_hi”, p. 6) #5fb236
	  * *
	  
	   
	  
	  - “The first sub-layer employs a multi-head self-attention mechanism, while the second sub-layer uses a straightforward, position-wise, fully connected feed-forward network” (“TSE-2025-01-0025_Proof_hi”, p. 6) #5fb236
	  * *
	  
	   
	  
	  - “In other words, the output of each sub-layer is passed through a normalization layer.” (“TSE-2025-01-0025_Proof_hi”, p. 6) #5fb236
	  * *
	  
	   
	  
	  - “Decoder” (“TSE-2025-01-0025_Proof_hi”, p. 6) #2ea8e5
	  * *
	  
	   
	  
	  - “This additional sub-layer conducts multihead attention over the encoder stack’s output” (“TSE-2025-01-0025_Proof_hi”, p. 6) #5fb236
	  * *
	  
	   
	  
	  - “Self-Attention mechanisms” (“TSE-2025-01-0025_Proof_hi”, p. 6) #2ea8e5
	  * *
	  
	   
	  
	  - “Within the selfattention layer, an input represented as a vector, denoted as x, transforms another vector called z” (“TSE-2025-01-0025_Proof_hi”, p. 6) #a28ae5
	  * *
	  
	   
	  
	  - “q (queries), k (keys), and v (values).” (“TSE-2025-01-0025_Proof_hi”, p. 6) #a28ae5
	  * *
	  
	   
	  
	  - “These encodings are computed using sine and cosine functions” (“TSE-2025-01-0025_Proof_hi”, p. 6) #5fb236
	  * *
	  
	   
	  
	  - “The selection of the sinusoidal approach was motivated by the potential to enable the model to make predictions for sequence lengths that extend beyond those encountered during its training.” (“TSE-2025-01-0025_Proof_hi”, p. 6) #5fb236
	  * *
	  
	   
	  
	  - “The process followed in this research involves creating an intermediate representation using the PlantUML language, structured as a tree.” (“TSE-2025-01-0025_Proof_hi”, p. 6) #5fb236
	  * *
	  
	   
	  
	  - “each child node corresponds to a distinct part of the class diagram,” (“TSE-2025-01-0025_Proof_hi”, p. 7) #a28ae5
	  * *
	  
	   
	  
	  - “Each class node further branches into nodes that define its attributes, methods and relationships.” (“TSE-2025-01-0025_Proof_hi”, p. 7) #5fb236
	  * *
	  
	   
	  
	  - “Relationships are depicted by relationship nodes that include child nodes representing the related classes, establishing connections and interactions between different classes within the diagram.” (“TSE-2025-01-0025_Proof_hi”, p. 7) #a28ae5
	  * *
	  
	   
	  
	  - “This ensures that the most interconnected classes are established early, providing a strong foundation for the diagram” (“TSE-2025-01-0025_Proof_hi”, p. 7) #5fb236
	  * *
	  
	   
	  
	  - “Out-of-vocabulary (OOV) pertains to words identified during testing or validation that do not exist in the training data.” (“TSE-2025-01-0025_Proof_hi”, p. 7) #a28ae5
	  * *
	  
	   
	  
	  - “Out-of-vocabulary tokens are substituted with a distinct symbol denoted as 〈UNK〉” (“TSE-2025-01-0025_Proof_hi”, p. 7) #5fb236
	  * *
	  
	   
	  
	  - “non-terminal nodes are characterized by their element type attribute” (“TSE-2025-01-0025_Proof_hi”, p. 7) #5fb236
	  * *
	  
	   
	  
	  - “GenClass” (“TSE-2025-01-0025_Proof_hi”, p. 7) #ffd400
	  *Is it mentioned before? *
	  
	   
	  
	  - “,” (“TSE-2025-01-0025_Proof_hi”, p. 7) #ff6666
	  * *
	  
	   
	  
	  - “sequance” (“TSE-2025-01-0025_Proof_hi”, p. 7) #ff6666
	  *sequence *
	  
	   
	  
	  - “E. Enhancing through Over-specification” (“TSE-2025-01-0025_Proof_hi”, p. 7) #5fb236
	  * *
	  
	   
	  
	  - “In most cases, there are attributes, methods, and sometimes classes within the class diagrams that do not appear or are not mentioned explicitly within the textual requirements of the software.” (“TSE-2025-01-0025_Proof_hi”, p. 7) #5fb236
	  * *
	  
	   
	  
	  - “In this section, some special processing operations in the ontology domain were considered, which would improve the resulting final diagrams by finding the missing attributes, methods, or relationships” (“TSE-2025-01-0025_Proof_hi”, p. 9) #5fb236
	  * *
	  
	   
	  
	  - “FPgrowth algorithm can be applied by identifying frequent individual items in the dataset, i.e., items that appear above a specified minimum support threshold.” (“TSE-2025-01-0025_Proof_hi”, p. 9) #5fb236
	  * *
	  
	   
	  
	  - “FP-Growth is particularly effective for analyzing and identifying associations or patterns among items.” (“TSE-2025-01-0025_Proof_hi”, p. 9) #5fb236
	  * *
	  
	   
	  
	  - “This section covers an analysis of the proposed framework’s performance and results across various datasets.” (“TSE-2025-01-0025_Proof_hi”, p. 9) #5fb236
	  * *
	  
	   
	  
	  - “top-tier” (“TSE-2025-01-0025_Proof_hi”, p. 10) #ffd400
	  *Not sure, they are toy models in the end. *
	  
	   
	  
	  - “These diagrams vary in size, ranging from a few to numerous classes, and are broken down into discrete classes and their relationships” (“TSE-2025-01-0025_Proof_hi”, p. 10) #5fb236
	  * *
	  
	   
	  
	  - “subset contains a collection of 649 UML class components.” (“TSE-2025-01-0025_Proof_hi”, p. 10) #5fb236
	  * *
	  
	   
	  
	  - “The PlantUCD (PlantUML Class Diagram) dataset was developed by aggregating diverse software problem statements and descriptions from various sources, including Stack Overflow, TutorialsPoint, and relevant literature” (“TSE-2025-01-0025_Proof_hi”, p. 10) #5fb236
	  * *
	  
	   
	  
	  - “Over two months, 618 responses were collected and subsequently refined to extract 147 UML class diagrams.” (“TSE-2025-01-0025_Proof_hi”, p. 10) #5fb236
	  * *
	  
	   
	  
	  - “data augmentation techniques were employed to expand the sample size of requirements and corresponding PlantUML statements.” (“TSE-2025-01-0025_Proof_hi”, p. 10) #5fb236
	  * *
	  
	   
	  
	  - “client or buyer , while shop owner might be altered to shopkeeper or merchant. However, substituting synonyms alone was not sufficient; thus, we suggest rephrasing the software requirements to further increase the data volume, ensuring that the underlying system remains unchanged” (“TSE-2025-01-0025_Proof_hi”, p. 10) #ffd400
	  *But what about mutations about the structure of the model. I'm not convinced about this operated data augmentation process. *
	  
	   
	  
	  - “This augmentation encompassed entity names, attributes, and relationships, resulting in a 50% increase in the overall data volume.” (“TSE-2025-01-0025_Proof_hi”, p. 10) #ffd400
	  *... and how those changes were reflected to the corresponding textual requirements? How about mutations that do not involve only renaming, but also structure changes! *
	  
	   
	  
	  - “It measures the similarity between the generated and reference sequences.” (“TSE-2025-01-0025_Proof_hi”, p. 10) #5fb236
	  * *
	  
	   
	  
	  - “The training spans approximately 60 epochs” (“TSE-2025-01-0025_Proof_hi”, p. 11) #5fb236
	  * *
	  
	   
	  
	  - “QUANTITATIVE, QUALITATIVE ANALYSIS AND DISCUSSIONS” (“TSE-2025-01-0025_Proof_hi”, p. 12) #ffd400
	  *The evaluation does not have explciti research questions. It is important to have them upfront! *
	  
	   
	  
	  - “The results gained in Section IV-D show the superiority of the Seq2Ast approach over the Seq2Seq approach in general.” (“TSE-2025-01-0025_Proof_hi”, p. 12) #a28ae5
	  * *
	  
	   
	  
	  - “The superiority can be explained by the fact that Seq2Seq models generate output sequences solely based on input sequences without considering the syntax or structure of the PlantUML language.” (“TSE-2025-01-0025_Proof_hi”, p. 12) #a28ae5
	  * *
	  
	   
	  
	  - “Seq2Ast models are designed to transform sequences into abstract syntax trees, effectively capturing the input text’s structural and semantic elements.” (“TSE-2025-01-0025_Proof_hi”, p. 12) #a28ae5
	  * *
	  
	   
	  
	  - “Tables VI, VII, and VIII present a comparative analysis between our proposed model and alternative models gathered from diverse studies.” (“TSE-2025-01-0025_Proof_hi”, p. 12) #5fb236
	  * *
	  
	   
	  
	  - “threats to validity” (“TSE-2025-01-0025_Proof_hi”, p. 13) #ffd400
	  *I would discuss this in a dedicated threats to validity section! *
	  
	   
	  
	  - “We know that UML diagrams can be categorized into structural and behavioral types. If we can express any diagram category using the PlantUML language, it opens the possibility of extending the methodology employed in this research to other diagram types.” (“TSE-2025-01-0025_Proof_hi”, p. 14) #ffd400
	  *The paper does not give any evidence to answer this question. *
	  
	   
	  
	  - “[49] Songyang-dev, “GitHub - songyang-dev/uml-classes-and-specs: Repository that contains the data used for ‘Extraction of UML Class Diagrams from Natural Language Specification’ (Yang et al. 2022),” GitHub, Jul. 21, 2022. Available: https://github.com/songyang-dev/uml-classes-and-specs. [Accessed: Oct. 22, 2024]” (“TSE-2025-01-0025_Proof_hi”, p. 15) #5fb236
	  * *