Natural Language Software Specifications (NLSS) involve the use of human language to define software requirements or directly program software functionality 1. This approach aims to make these specifications broadly accessible and understandable to diverse stakeholders 1. It is a common practice in both legacy and modern software systems to capture requirements using unstructured or semi-structured natural language 3. Fundamentally, NLSS seeks to bridge the chasm between informal language and rigorous, formal analysis techniques, thereby alleviating the complexities often associated with formal specification languages 3.
NLSS encompasses various approaches that leverage natural language for specifying software. Its key characteristics include Accessibility, making specifications understandable by individuals with varying levels of familiarity with formal specification languages, such as developers, managers, and customers 1. It promotes Intuitive Understanding by preserving the natural properties of human language for correct comprehension of texts 4. Unlike naturally occurring sublanguages, NLSS methods often feature a Constructed Nature, involving languages that are explicitly and consciously defined 4. Furthermore, NLSS approaches exhibit Varied Formality, spanning a spectrum from inherently ambiguous to as precise as formal logic 4.
The underlying goals and principles of NLSS are centered on improving communication, clarity, and automation within software development. A primary objective is Bridging the Formal-Informal Gap, connecting natural language requirement specifications with efficient formal analysis techniques like goal model analysis 3 or OCL (Object Constraint Language) verification 1. This approach leads to Enhanced Communication, facilitating better collaboration among all project stakeholders through easily understandable documentation 1. NLSS also enables Stakeholder Validation, allowing non-technical individuals, such as customers, to verify that specifications accurately capture desired system behavior 1. By minimizing the need for specialized skills often required by purely formal approaches, which are considered burdensome for analysts, NLSS contributes to Reduced Development Burden 3. In certain advanced forms, it even supports Direct Executability, allowing for programming directly through natural language sentences, where each sentence can unambiguously compile into a procedure call in a high-level programming language 2.
To address the challenges of ambiguity and inconsistency inherent in general natural language, several established methodologies have emerged for developing and managing NLSS. These methodologies vary in their level of formality and application, each offering distinct advantages and disadvantages.
Controlled Natural Languages (CNLs) are constructed languages based on a specific natural language, made more restrictive in terms of lexicon, syntax, and/or semantics 4. Their goal is to improve communication, assist translation, or provide natural representations for formal notations 4. Examples include Basic English and Attempto Controlled English 4. Tools like the KeY tool exemplify CNLs by automatically translating formal OCL specifications into natural language, enabling parallel editing of both 1.
Structured English uses a restricted subset of natural language coupled with specific structural rules to minimize ambiguity and enhance clarity 2. This approach is often associated with programming languages that feature English-like syntax, such as COBOL or SQL 2.
User Stories are short, simple, and user-centric narratives designed to capture a feature's functionality from the end-user's perspective 5. They commonly adhere to the format: "As a [type of user], I want [some goal] so that [some reason]" 5.
Use Cases provide detailed, structured descriptions of interactions between "actors" (users or other systems) and the system to achieve a specific goal 5. They typically include components such as a title, goal, actors, preconditions, a main flow of events, and alternative flows 5.
Natural Language Programming (NLP, as a programming paradigm) is an ontology-assisted programming method where natural language sentences directly define program functionality 2. In this paradigm, each sentence unambiguously compiles into a procedure call in an underlying high-level language, resulting in a structured document that functions as a computer program itself 2.
Each NLSS methodology presents a unique balance of benefits and drawbacks, influencing its suitability for various project requirements and contexts. The aim remains to leverage the intuitive nature of human language while mitigating its inherent ambiguities for software specification.
| Methodology | Benefits | Drawbacks |
|---|---|---|
| Controlled Natural Languages (CNLs) | Improve comprehensibility for humans and can aid in machine translation 4. Provide intuitive representations for formal notations 4. Allow non-experts to understand formal specifications 1. Syntax-directed editors ensure syntactically and type-correct specifications 1. | Can be a "fuzzy term" with wide variety, and some may remain ambiguous or look more like programming languages 4. Translations may remain at the same abstraction level as original formal specifications, not offering truly informal explanations 1. Automatic formalization of arbitrary informal specifications is typically outside their scope 1. |
| Structured English | Enhances clarity and minimizes ambiguity in textual descriptions 2. Can directly inform programming constructs with English-like syntax 2. | Limited in expressiveness compared to full natural language. Requires adherence to strict structural rules. |
| User Stories | Customer-centric approach, aligning development with user needs and experiences 5. Highly flexible, allowing for easy modification and reprioritization in agile backlogs 5. Simple, concise, and use non-technical language, fostering clear communication across teams 5. | Lack detailed documentation, which can lead to ambiguity or incomplete understanding of feature specifics 5. Limited visibility of dependencies between stories 5. Can make effort estimation challenging due to their high-level nature 5. |
| Use Cases | Offer a clear, structured representation of system behavior, facilitating effective communication among diverse project roles 5. Enable thorough requirement validation and help identify potential gaps or issues early 5. Serve as a strong basis for creating comprehensive test cases and validating system functionality 5. | Can become complex and overwhelming to manage in large systems 5. Time-consuming to create and maintain due to their detailed nature, requiring significant effort in defining scenarios and interactions 5. May not capture all possible system interactions, potentially overlooking niche functionalities 5. |
| Natural Language Programming (NLP) | Human-readable and also machine-readable, providing a precise formal description for software 2. Allows software agents to execute tasks based on the natural language sentences 2. Can function as high-level, language-independent pseudo-code 2. | Requires the explicit definition of an underlying ontology of concepts 2. The process of unambiguously compiling natural language sentences into code can be complex to establish and maintain 2. |
| General Natural Language Requirements | Highly intuitive for humans to author and read, widely adopted for initial requirement capture 3. | Prone to significant ambiguities, inconsistencies, and incompleteness 3. Difficult to perform rigorous, formal analysis (e.g., for satisfiability or conflict detection) directly 3. |
While Natural Language Software Specifications (NLSS) offer intuitive communication among stakeholders, their inherent nature introduces significant challenges and limitations that can profoundly impact software development processes, quality, and project success. These difficulties arise from the characteristics of natural language itself, affecting various stages of the software lifecycle.
Natural language, despite its versatility, presents several intrinsic problems when used for formal software specifications:
Ambiguity and Vagueness: Natural language is inherently ambiguous, with words, phrases, and sentences often having multiple interpretations depending on context . This leads to misunderstandings, subjective interpretations, and makes it challenging to specify unique requirements 6. For instance, a requirement such as "The system should provide a fast response" is vague because "fast" is subjective and open to diverse interpretations by different stakeholders 6. The English language exacerbates this issue, with many words possessing numerous meanings; "round," for example, has 70 distinct interpretations, making unambiguous statement construction difficult 7. Readers often subconsciously disambiguate these ambiguities to their initial interpretation, frequently overlooking alternative possibilities 7.
Lack of Precision: NLSS frequently lack the exactness required for technical specifications, making it difficult for developers to determine precisely what needs to be built 6. Requirements can be overly broad or imprecise, leading to varied implementations of the same functionality 6.
Inconsistency: In requirements documentation involving multiple authors, inconsistencies are almost inevitable 6. This can manifest as different terminology being used for similar functionalities, such as referring to the same group as "users" in one section and "customers" in another, causing confusion and errors 6.
Incompleteness: Incomplete requirements fail to fully describe the intended software features and functionalities, leading to software that is unprepared for expected scenarios or handles them unexpectedly 7. Incomplete requirements are consistently cited as a primary reason for project failures 8.
Dependency on Authors' Writing Skills: The quality of requirements documentation is highly dependent on the authors' ability to write clearly, concisely, and unambiguously 6. Poorly written requirements can cause significant confusion and misinterpretation throughout the entire development process 6.
Contextual Understanding: Natural language relies heavily on context, including idiomatic expressions, cultural references, and domain-specific jargon . Machines struggle to achieve this level of understanding, which is crucial for accurate interpretation of specifications 9.
Linguistic Diversity: Human languages exhibit vast diversity in syntax, semantics, and structure . Developing NLSS for multiple languages, particularly those less commonly used or with limited available data, presents a significant challenge .
The intrinsic characteristics of NLSS pose substantial hurdles for quality assurance activities:
Verification and Validation (V&V): Ambiguous and imprecise requirements are inherently challenging to verify and validate . This creates significant obstacles for testing teams attempting to create test cases that accurately reflect the system's intended behavior 6. The outcome can be incomplete testing and an increased risk of software defects 6. The IEEE Std. 830-1998 specifically highlights verifiability as a key criterion for the acceptability of Software Requirements Specifications (SRS) 7.
Traceability: The inherent over-flexibility of natural language often prevents effective partitioning of requirements, making it difficult to trace specific requirements to software code or test cases 7. This also complicates understanding the impact of changes across the specification, even though traceability is recognized as an important activity in requirements engineering (RE) 8.
As software projects increase in size and complexity, managing and maintaining requirements documented in non-restricted natural language becomes increasingly arduous 6. The sheer volume of requirements complicates efforts to ensure consistency and clarity, potentially leading to overlooked requirements or critical misinterpretations that may only be discovered late in the development cycle 6.
The challenges and limitations of NLSS manifest in several practical pitfalls and theoretical barriers:
Software Project Failures: A significant number of software failures originate from inadequate, inconsistent, incorrect, imprecise, or ambiguous requirements 7. Surveys indicate that only a small percentage of projects are deemed successful, with many experiencing cost overruns and late deliveries due to these fundamental requirements issues 7. The Standish Group's CHAOS report identifies "incomplete requirements" and "changing requirements and specifications" as primary factors contributing to project failures .
Costly Error Resolution: Errors detected in NLSS are much more difficult, time-consuming, and expensive to resolve if they are discovered in later phases of software development 7.
Legal and Contractual Risks: When software requirements serve as contractual documents, the ambiguity and vagueness inherent in natural language can lead to disputes between clients and developers, making it difficult to ascertain whether the delivered software meets the agreed-upon criteria 6.
Inadequate Basis for Design and Contracts: Natural language is considered a poor basis for both software design and contractual agreements due primarily to its inherent ambiguity and informality 7. Its over-flexibility allows for the same concept to be expressed in numerous ways, demanding deep contextual understanding for accurate interpretation, which can be an error-prone process 7.
Computational Limitations: Advanced Natural Language Processing (NLP) models, while promising for addressing some of these challenges, require substantial computational resources 9. Furthermore, their effectiveness relies heavily on the quality and quantity of training data, which can be a significant barrier for less-resourced languages or highly specialized domains 9.
Human Intelligence Requirement: Automatic disambiguation by software tools is often not possible without human intelligence, as machines cannot perfectly understand or automatically resolve potential ambiguities in natural language without human oversight or prior knowledge 7.
The aforementioned challenges posed by NLSS have profound consequences across the entire software development lifecycle:
Compromised Software Quality: Inconsistent, incorrect, and ambiguous requirements directly compromise the quality of the resulting software 7. This can lead to an increase in defects and a final system that does not align with user expectations 6.
Inefficient Development Processes: Difficulties in verifying and validating NLSS result in incomplete or incorrect implementations that are costly and time-consuming to rectify later in the development process 6. Poor communication, often exacerbated by NLSS ambiguity, contributes to process failures where budget, time, or resources are significantly overrun 7.
Project Failure and Cost Overruns: Inadequate requirements are a primary contributor to project failures, characterized by significant cost overruns and delayed deliveries . The lack of user involvement, frequently stemming from unclear requirements, has historically been a reason for project failure, though adequate addressing of this can transform it into a success factor 7.
Mismatch Between Specification and System Behavior: Over-idealization in requirements can lead to inconsistencies between the specified system and its actual behavior, potentially resulting in poor performance or even hazardous outcomes 7.
The field of Natural Language Software Specifications (NLSS) is experiencing a profound transformation, driven by rapid advancements in Artificial Intelligence (AI), Machine Learning (ML), and Natural Language Processing (NLP). These technologies are increasingly integral to automating and enhancing various stages of the software requirements engineering lifecycle, moving beyond traditional manual methods to more efficient and accurate AI-driven approaches. This section delves into these technological advancements, emerging trends, and their transformative impact on NLSS, highlighting significant breakthroughs and nascent industry adoptions.
AI and ML are revolutionizing NLSS by automating and improving critical processes, thereby boosting accuracy and efficiency across the specification lifecycle.
One significant advancement is the automated analysis of software specifications. AI and ML techniques are extensively utilized to identify inherent issues such as ambiguities, inconsistencies, and incompleteness commonly found in natural language requirements 10. ML algorithms learn from historical data to reduce errors and discern patterns within Software Requirement Specification (SRS) data 11. Tools like Visual Narrator, QuARS, QVScribe, Qualicen Requirements Scout, and IBM Engineering Requirements Quality Assistant leverage automated or semi-automated analysis to enhance the clarity and quality of specifications 10. Furthermore, supervised ML algorithms, including Support Vector Machines (SVM), Naive Bayes (NB), Decision Trees (DT), K-Nearest Neighbors (KNN), and Random Forests (RF), are frequently applied for requirements classification and prioritization 12. Predictive analytics are also employed to detect potential conflicts, ambiguities, and completeness gaps during validation, especially beneficial in large-scale projects 13.
Generative AI (GenAI), particularly through Large Language Models (LLMs), is emerging as a new paradigm for automating the generation, refinement, and analysis of requirements 13. These systems possess the ability to synthesize novel and coherent content, spanning text, code, and structured data 13. Frameworks now exist that generate software requirement specifications based on user input, by processing existing SRS data using NLP libraries and neural networks 11. LLMs facilitate a deeper understanding of context, enable content generation, and support cross-lingual reasoning, allowing for the creation of consistent and structured specification drafts 13.
The transformation of natural language specifications into more formal or structured representations is another critical area benefiting from AI and ML. NLP, in conjunction with AI, can automate the mapping of natural language requirements to formal representations 10. This includes the automatic generation of Unified Modeling Language (UML) diagrams directly from natural language requirements, effectively bridging the gap between conceptual requirements and formal software design models 10. Additionally, techniques are being developed to extract domain knowledge and system features from natural language text, which aids in rewriting requirements into a standardized form and establishing dependencies 12.
The integration of AI/ML with NLSS is leading to several evolving paradigms that enhance the overall requirements engineering process.
The evolution of Requirements Engineering (RE) has seen a significant shift from traditional manual processes to Model-Driven RE. This approach incorporates visual representations and formal specifications to improve clarity and traceability 13. Research increasingly focuses on "formal specification" and integrating syntactic analysis into structured and formalized contexts, particularly for software specifications and design 10. This also involves the integration of Computer-Aided Software Engineering (CASE) tools with NLP systems 10.
Automated requirement elicitation is advancing through AI-driven approaches that improve analysis capabilities and handle larger data volumes 13. NLP capabilities embedded within GenAI enhance stakeholder communication by generating clarifying questions and uncovering implicit requirements during the elicitation phase 13. There is also a growing interest in applying NLP to understand and process "user stories" within agile software development processes 10. The focus on "extraction" from specifications indicates a clear trend toward automatically obtaining valuable concepts from software requirements, which is essential for automating the software development process 10.
Recent trends indicate a pronounced focus on cutting-edge AI technologies, with "deep learning," "machine learning," and "learning systems" becoming central from 2022 onwards 10. This includes specialized techniques like zero-shot learning (ZSL) for classifying requirements without requiring labeled training data and advanced neural networks such as Bidirectional Long Short-Term Memory (Bi-LSTM) combined with self-attention mechanisms for improved classification accuracy 12.
Recent years have witnessed substantial breakthroughs, particularly with the rise of GenAI, alongside growing industry interest in NLSS.
Large Language Models (LLMs) are at the forefront of driving innovations in NLP and broader AI applications within NLSS 10. Transformer models, such as BERT and GPT, have significantly boosted NLP capabilities, enabling the processing of large data volumes and capturing complex relationships within text 12. GenAI models, exemplified by GPT-4, are characterized by their ability to generate novel and coherent content across various modalities, including text, images, and code 13. Research specifically investigates how LLMs can automate the generation, refinement, and analysis of requirements 13. These advancements also include the employment of models like FastText and BETO for processing non-English languages, expanding the global reach of NLSS tools 10.
A notable trend is the increasing interest in automatically generating Unified Modeling Language (UML) diagrams from natural language requirements, showcasing NLP's potential to formalize conceptual requirements into design models 10. Furthermore, ML techniques are widely applied for classifying requirements into functional (FR) and non-functional (NFR) categories, and further into specific NFR subcategories such as security and usability 12. The PROMISE dataset is frequently utilized for these classification studies 12.
The scope of NLP applications in NLSS is expanding to include advanced models for automatic classification and the processing of non-English languages 10. Predictive analytics and pattern recognition techniques are being employed to analyze past project data, uncover common requirement patterns, and detect omissions, thereby contributing to enhanced requirement quality and traceability 13. While industrial adoption of GenAI in RE remains largely in nascent stages, with over 90% of studies corresponding to early-stage development, some production-level integrations exist. For instance, LLMs are facilitating change impact analysis and automated validation within ERP systems 13.
NLP is a foundational technology that underpins the evolution of NLSS, enabling machines to process, analyze, and understand human language.
NLP bridges the gap between human language and machine comprehension, allowing machines to process, analyze, understand, and generate human-like responses 10. This involves core capabilities such as Natural Language Understanding (NLU) and Natural Language Generation (NLG) 14. NLU encompasses phonological, morphological, lexical, syntactic, semantic, and discourse analysis, enabling machines to extract concepts, entities, and emotions from text 14. In software requirements engineering, NLP advancements involve analyzing lexical and syntactic attributes, mapping to formal representations, and extracting relevant domain knowledge 10.
The evolution of NLP has been significantly impacted by Deep Learning (DL) algorithms, which excel in tasks such as document classification, paraphrase identification, and text similarity 12. Pre-trained Language Models (PLMs), including BERT and GPT, have provided a tremendous boost to NLP, efficiently processing long texts and achieving deeper contextual understanding through transformer architectures 12. These models have facilitated advancements like neural language modeling, multitask learning, word embedding, sequence-to-sequence mapping, and the use of Convolutional Neural Networks (CNN), Recurrent Neural Networks (RNN), Long Short-Term Memory (LSTM), and Gated Recurrent Unit (GRU) in various NLP tasks 14.
Natural language, while expressive, inherently suffers from ambiguity, incompleteness, and inconsistencies, presenting significant challenges in requirements engineering 10. NLP approaches have been used since the 1980s to automate RE tasks, with a continuous focus on resolving these issues 12. Techniques like sentiment analysis, Part-of-Speech (POS) tagging, chunking, Named Entity Recognition (NER), and Semantic Role Labeling (SRL) contribute to extracting and refining information from raw text requirements 14. Automated and semi-automated disambiguation tools, powered by NLP, show promising results in minimizing ambiguity, though complete elimination remains a challenge . The ability of NLP to analyze the context and semantic meaning of words is crucial for enhancing the quality and understandability of specifications 14.
The table below summarizes some key technologies and their applications in NLSS:
| Technology/Method | Application in NLSS | Key Benefits |
|---|---|---|
| Supervised ML (SVM, NB, RF) | Requirements classification and prioritization | Improved organization, faster processing, reduced human error |
| Generative AI (LLMs, GPT-4) | Automated specification generation, refinement, analysis | Novel content creation, contextual understanding, structured drafts |
| NLP for formal transformation | Automatic UML diagram generation from text | Bridging conceptual requirements with formal design models |
| AI for Elicitation | Generating clarifying questions, uncovering implicit requirements | Enhanced stakeholder communication, more complete requirements |
| Deep Learning (Bi-LSTM, Transformers) | Advanced classification, contextual understanding of text | Higher accuracy, processing of complex relationships in large data |
| Predictive Analytics | Detecting conflicts, ambiguities, completeness gaps | Proactive quality assurance, risk mitigation in large projects |
| Multilingual NLP (FastText, BETO) | Processing non-English language specifications | Global reach, inclusivity for diverse development teams |
In conclusion, the landscape of Natural Language Software Specifications is undergoing a significant transformation, moving towards highly automated, intelligent, and efficient processes driven by advancements in AI, ML, and NLP. These technologies are enabling new paradigms in requirement analysis, generation, and transformation, offering substantial potential for future software development.
Natural Language Software Specifications (NLSS), underpinned by advancements in Natural Language Processing (NLP), are increasingly integral across diverse industries for managing and processing software requirements . NLP, a domain of artificial intelligence, enables machines to understand and generate human language, thereby enhancing automation precision and data accessibility . This technology is bridging the gap between human linguistic expression and machine comprehension in software requirements engineering (SRE), significantly boosting the efficiency and accuracy of software development processes . The NLP market's substantial projected growth, reaching $114.44 billion by 2029 and $201.49 billion by 2031, highlights its increasing relevance and broad adoption .
NLSS, through NLP, is transforming operations by applying advanced language models to requirements-related tasks across various sectors. The following table summarizes key applications:
| Industry | Key Application of NLSS (via NLP) | Example/Impact |
|---|---|---|
| Healthcare | Medical text summarization, patient monitoring, drug discovery | Hospitals condense EHRs for faster decision-making; John Snow Labs' Spark NLP for Healthcare accelerated research timelines by 40% 15; Acentra Health's "MedScribe" saved 11,000 nursing hours 15; Vanderbilt analyzed 2.8 million notes for phenotype correlations 16. |
| Finance and Banking | Fraud detection, financial report generation, customer service | Wells Fargo detected unusual language patterns in quarterly reports 16; JPMorgan's COIN analyzed 12,000 loan agreements annually, reducing work hours by 360,000 and errors by 66% 16; Bank of America's Erica handled 100 million+ requests, reducing call center volume by 30% 15; HSBC classified 100 million transactions daily for compliance, reducing false positives by 20% 16. |
| E-commerce and Retail | Personalized recommendations, AI chatbots, sentiment analysis | Amazon uses NLP for search intent 16; H&M and Sephora use bots for customer preferences 16; Stitch Fix increased customer retention by 30% through tailored recommendations 15. |
| Recruitment and HR | Resume parsing, employee feedback analysis, automated interviews | Johnson & Johnson processes 1.5 million resumes annually, saving 70% recruiter time 16; Microsoft's "Employee Voice" improved satisfaction by 24% 16; L'Oréal's "Mya" chatbot reduced time-to-hire by 40% 16; Unilever reduced hiring timelines from four months to four weeks 16. |
| Software Requirements Engineering (SRE) | Elicitation, specification, modeling, validation, IT support | NLP is progressively adopted in SRE tasks 10; 58.2% of practitioners use AI tools in SRE, with 69.1% reporting positive impact 17; Moveworks uses NLP to automate IT support for clients like Autodesk 15; Coherent Solutions developed an Amify NLP System for e-commerce insights 15. |
NLSS, by leveraging NLP, significantly influences project success through enhanced efficiency, reduced costs, and improved outcomes 18.
NLP-powered solutions contribute to faster management and reduced operational costs across industries 18. In healthcare, medical text summarization enables doctors to quickly understand patient cases, leading to faster decision-making 19. Automated report generation in finance reduces manual effort and accelerates analysis 18. Similarly, automated resume parsing speeds up candidate screening, cutting down time-to-hire 18. Notable examples include KPMG's Ignite platform, which reduced document processing time by 60% and improved audit accuracy by 40% 15. Access Holdings Plc significantly reduced code development time and accelerated chatbot deployment 15. Zurich Insurance reduced claim processing time by 90%, from 58 minutes to 5 minutes 16, and Allen & Overy saved $2.5 million and shortened due diligence by three weeks using NLP for legal contract review 16. In requirements engineering, AI is largely seen as a productivity multiplier that expedites routine tasks 17.
NLSS helps meet user expectations through personalization and enhanced service. Personalized product recommendations and targeted marketing improve customer satisfaction and boost sales 18. Chatbots powered by NLP deliver faster resolutions and higher customer satisfaction by accurately interpreting context and intent 18. Real-time sentiment analysis helps brands refine messaging based on customer perception 18. In healthcare, improved clinical documentation and patient monitoring enhance patient care and outcomes 18. American Express saw a 20% improvement in Net Promoter Score and a 15% reduction in customer churn due to NLP-driven customer service monitoring 15. The New York Times' "Project Feels" increased subscriber retention by 31% through content recommendation engines 16. User expectations are better met when NLSS clearly defines requirements for user-friendly systems and adequate resources 20.
NLSS profoundly influences how stakeholders communicate and collaborate within development projects and beyond.
NLP-powered chatbots provide fast, intelligent, 24/7 support, enhancing customer experience 19. Instant language translation facilitates seamless communication across diverse regions, supporting global collaboration 18. In healthcare, NLP structures physician notes into actionable insights, improving clarity in communications 18. In requirements engineering, NLP assists in summarizing stakeholder inputs and generating clarifying questions 17. Natural language is widely adopted for documenting requirements due to its ease of learning and adaptability, simplifying initial communication 21. NLP-driven business intelligence allows conversational queries, democratizing data access and fostering a data-driven culture 15.
Human-AI Collaboration (HAIC) is prevalent in SRE, accounting for 54.4% of AI techniques used, indicating a preference for AI as a collaborative partner rather than a replacement 17. HAIC is particularly high in the analysis phase (60.5%) 17. AI can enhance stakeholder validation meetings by pre-identifying issues, enabling human participants to concentrate on decision-making 17. However, challenges persist as AI models often lack deep domain understanding and contextual awareness, hindering their ability to interpret unstated or implicit requirements 17. Current AI cannot replicate interpersonal skills, rapport-building, non-verbal cue interpretation, or political navigation crucial for effective stakeholder engagement in requirements engineering, underscoring that RE remains a fundamentally social process 17. Additionally, explaining requirements to AI systems may require more time investment to achieve proper understanding 17.
The adoption of NLSS, through NLP, presents both positive contributions and potential drawbacks for software quality.
NLP improves software requirements quality by analyzing lexical and syntactic attributes, automating mapping to formal representations, and refining specifications through domain knowledge extraction 10. It aids in detecting and resolving ambiguities and inconsistencies in software requirements 10. Automated analysis ensures consistency and reduces human error in documentation 19. NLP-driven systems can process large data volumes to identify patterns, suggest missing requirements, detect inconsistencies, and check for compliance 17. AI models clarify vague terms, prioritize features, and generate detailed technical specifications, leading to accurate and traceable documentation 17. Effective systems engineering, founded on well-defined requirements, aims to minimize defects, reduce risk, and enhance maintainability 20. Extracting goal models from natural language specifications allows for rigorous inspection regarding satisfiability, consistency, completeness, conflicts, and ambiguities 3.
Natural language is inherently susceptible to ambiguity, inaccuracy, and inconsistency, which can lead to misinterpretations in requirements if not carefully managed . Carelessly chosen words, weak phrases (e.g., "as a minimum"), and generalities (e.g., "large") can produce imprecise requirements 20. Informal language can result in "stream of consciousness" statements that are difficult to understand and maintain 20. Current AI models may produce inaccurate, incomplete, or generic outputs due to a lack of deep domain understanding and contextual nuances 17. AI's probabilistic nature and reliance on generalized training data can lead to contextually incorrect solutions 17. Incomplete training data can limit AI effectiveness, potentially causing the omission of critical requirements 17. Furthermore, improper use of AI tools, such as substituting Large Language Models (LLMs) for direct customer engagement, can undermine requirement quality 17.
To maximize the benefits and mitigate the challenges of NLSS adoption, adherence to best practices and consideration of regulatory aspects are crucial.
A disciplined and consistent approach to document design, statement formulation, and keyword selection is essential 20. Requirements should be organized logically to ensure clarity and avoid arbitrary grouping of information 20. Using simple, short declarative sentences, ideally one requirement per sentence, enhances readability and reduces confusion . Specification statements should be structured with elements such as localization, actor/owner, action, target/owned, and constraint 20. It is critical to employ formally defined words and avoid informal, relative, or ambiguous terms to ensure precision 20. Consistent use of imperative words like "shall" for prescriptions, "will" for descriptions, "must" and "must not" for constraints, and "should" for suggestions is recommended 20. Weak phrases like "as appropriate" or "if practical" should be avoided as they offer options regarding requirement satisfaction 20. Clear terminology definitions, possibly through glossaries, and standardized templates promote consistent documentation 21. Thorough reviews of documented requirements are necessary to identify linguistic pitfalls and ensure clarity 21. Effective requirements documentation also demands continuous training and improvement of writing skills 21.
Industry standards such as IEEE Standard 830-1993, MIL-STD-498, and NASA-STD-2100-91 provide frameworks for structuring software requirements documentation 20. Adapting Data Item Descriptions (DIDs) for project-specific needs is a critical design activity, with maintaining paragraph numbering integrity in DIDs being important for information correlation 20. Financial institutions utilize NLP to monitor regulatory publications and extract compliance requirements, though complex interpretations often still necessitate human judgment .
Regarding AI usage in SRE, responsible AI practices are emerging:
Concerns about privacy, security, legal risks, and the explainability of AI decisions pose significant challenges, particularly in highly regulated industries where human oversight is critical 17. Custom NLP models tailored to industry-specific needs offer significant advantages in automation, efficiency, and compliance 18.