Paper deep dive
From Conceptual EneMonIE to an Executable Backend: A Modular Architecture for Energy Information Extractionand Grounded Question Answering
Janik Lesage, Khadim Lô, and Jean-Sébastien Dessureault
Intelligence
Status: not_run | Model: - | Prompt: - | Confidence: 0%
Entities (0)
Relation Signals (0)
No relation signals yet.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:The original EneMonIE proposal describes a Web-based semantic system for continuously monitoring energy trends through targeted information extraction from heterogeneous online media. This article presents EneMonIE-Backend, an implementation-derived redesign that converts that conceptual pipeline into a modular, executable, and application-independent backend. The architecture separates five concerns: data sources, acquisition and normalization, semantic processing, knowledge organization, and persistent storage. It operationalizes configurable connectors for RSS feeds, URL collections, arXiv, social-media JSON files, and open-data CSV files; it also introduces question-driven Web search and scraping. Relevant documents pass through energy categorization, named entity recognition, temporal and event extraction, source sensitive sentiment analysis, and extractive summarization. The resulting semantic representations support entity co-occurrence graphs, cross-document interlinking, cluster-level information fusion, and relational persistence in SQLite. Three grounded question-answering paths are exposed at the backend level: live Web evidence, the stored corpus, and selected configured resources. Answer generation uses a language model when configured and otherwise falls back to deterministic extractive output. Unlike the original proposal, the presented architecture deliberately excludes the user-interface and server-side application layers, allowing the semantic core to be embedded in different applications. A deterministic smoke validation confirms relevance filtering, persistence, clustering, evidence retrieval, and document deletion. The work provides a reproducible architectural basis for future evaluation on larger multilingual energy corpora.
Tags
Links
- Source: https://arxiv.org/abs/2610.17243v1
- Canonical: https://arxiv.org/abs/2610.17243v1
Trouble viewing inline? Open PDF directly →
Full Text
33,166 characters extracted from source content.
Expand or collapse full text
From Conceptual EneMonIE to an Executable Backend: A Modular Architecture for Energy Information Extraction and Grounded Question Answering Janik Lesage, Khadim Lô, and Jean-Sébastien Dessureault Département de mathématiques et d’informatique, Université du Québec à Trois-Rivières, Canada Abstract The original EneMonIE proposal describes a Web-based semantic system for continuously monitoring energy trends through targeted information extraction from het- erogeneous online media. This article presents EneMonIE- Backend, an implementation-derived redesign that con- verts that conceptual pipeline into a modular, executable, and application-independent backend. The architecture separates five concerns: data sources, acquisition and nor- malization, semantic processing, knowledge organization, and persistent storage. It operationalizes configurable connectors for RSS feeds, URL collections, arXiv, social- media JSON files, and open-data CSV files; it also intro- duces question-driven Web search and scraping. Relevant documents pass through energy categorization, named- entity recognition, temporal and event extraction, source- sensitive sentiment analysis, and extractive summariza- tion. The resulting semantic representations support entity co-occurrence graphs, cross-document interlinking, cluster-level information fusion, and relational persistence in SQLite. Three grounded question-answering paths are exposed at the backend level: live Web evidence, the stored corpus, and selected configured resources. Answer generation uses a language model when configured and otherwise falls back to deterministic extractive output. Unlike the original proposal, the presented architecture deliberately excludes the user-interface and server-side application layers, allowing the semantic core to be em- bedded in different applications. A deterministic smoke validation confirms relevance filtering, persistence, clus- tering, evidence retrieval, and document deletion. The work provides a reproducible architectural basis for future evaluation on larger multilingual energy corpora. Keywords: energy informatics; information extrac- tion; natural language processing; semantic architecture; grounded question answering; data fusion. 1 Introduction Energy systems generate a continuous stream of heteroge- neous information: news articles, research papers, public datasets, policy announcements, organizational reports, and social-media posts. The value of these materials lies not only in their individual content but also in their relationships. A power-grid investment reported by a util- ity, discussed by the public, and evaluated in a research preprint may represent a single underlying development. Manually monitoring such evidence is costly, slow, and difficult to reproduce. The conceptual EneMonIE architecture was originally proposed by Dilek Küçük [Küçük, 2022]. The present work implements, restructures, and extends that proposal as an application-independent backend. The original conceptual design combines text categorization, named-entity recog- nition (NER), temporal-expression and event extraction, sentiment analysis, social-network construction, media interlinking, information fusion, summarization, a central database, and Web-based retrieval and visualization. It also emphasizes modularity, bilingual processing, and the reuse of domain ontologies and lexical resources. The original proposal is deliberately broad. It speci- fies the semantic functions that the system should pro- vide, but it does not prescribe an executable control flow, concrete storage schema, connector contract, question- answering strategy, or graceful degradation mechanism. These implementation decisions become essential when the architecture is turned into software. They determine how duplicate documents are handled, how relevance is en- forced before expensive analysis, how semantic objects are represented, and how evidence is exposed to downstream consumers. This article presents EneMonIE-Backend, a backend architecture derived from a working Python implemen- tation of the EneMonIEKüçük [2022] concept. The central design decision is to isolate the semantic and data-management core from any particular application. FastAPI endpoints, a desktop client, an HTML dashboard, and Web visualization are therefore outside the boundary of the architecture described here. The backend can in- stead be invoked by a command-line process, a scheduler, a test harness, or a future service layer. The main contributions are as follows: 1) an exe- cutable five-layer decomposition of the original EneMonIE concept. 2) normalized ingestion through configurable source adapters and stable document identifiers. 3) an ex- plicit semantic-processing sequence with an early energy- relevance gate. 4) operational definitions for entity graphs, media similarity, clustering, and fusion. 5) a normalized SQLite schema connecting documents, semantic anno- tations, entity edges, and clusters. 6) three evidence- 1 arXiv:2610.17243v1 [cs.CL] 12 Oct 2026 grounded question-answering paths with a local fallback when no external language model is configured. 7) an application-independent boundary that separates reusable domain logic from presentation and transport concerns. 2 Background and Design Motivation 2.1From semantic monitoring to an exe- cutable pipeline EneMonIE belongs to the broader family of continuous knowledge-acquisition systems. Never-Ending Language Learning (NELL) emphasized persistent learning from Web content [Carlson et al., 2010]; DeepDive demon- strated large-scale knowledge-base construction using sta- tistical learning and inference [Niu et al., 2012]; and Knowledge Vault combined Web extraction with proba- bilistic knowledge fusion [Dong et al., 2014]. EneMonIE narrows this general objective to the energy domain, where entities, events, dates, technologies, organizations, and public opinions must be related across sources. Several components of the original design are well- established NLP tasks. NER over informal text is challeng- ing because social media frequently departs from conven- tional spelling and syntax [Ritter et al., 2011]. Temporal normalization links textual expressions to comparable dates [Strötgen and Gertz, 2010], while event extraction structures occurrences that may have policy, operational, environmental, or technological consequences [Ritter et al., 2012]. Sentiment analysis provides a coarse represen- tation of expressed opinion [Pang and Lee, 2008], and multi-document or social-media summarization reduces repeated evidence to concise statements [Chakrabarti and Punera, 2011]. Energy ontologies further provide domain vocabulary and conceptual structure [Küçük, 2015]. An executable architecture must coordinate these tasks and define their data contracts. In particular, down- stream clustering should not operate on raw text alone when normalized entities, event types, and dates are avail- able. Similarly, question answering should not bypass the semantic pipeline when live Web pages are collected. EneMonIE-Backend addresses these concerns by using one normalized document model and one processing function for both scheduled ingestion and question-driven evidence collection. 2.2 Design requirements The redesign follows six requirements: 1) Modularity. Source-specific acquisition and semantic analyzers must be replaceable behind stable interfaces. 2) Early filtering. Non-energy documents must be rejected before semantic extraction and persistence. 3) Traceability. Every de- rived object must remain connected to its source document and, where available, its original URL. 4) Determinism by default. Core extraction, storage, ranking, and fall- back answering must operate without a remote language model. 5) Evidence grounding. Generated answers must be based on a bounded evidence set rather than unconstrained model knowledge, following the grounding motivation of retrieval-augmented generation [Lewis et al., 2020]. 6) Application independence. Transport and presentation layers must not own the semantic pipeline or the database model. 3 Proposed Architecture Figure 1 presents the complete backend architecture. The main pipeline consists of five layers, while a separate lower section illustrates question-driven execution paths that reuse the same processing and persistence services. More specifically, the upper pipeline is responsible for acquiring, normalizing, analyzing, organizing, and stor- ing energy-related documents, whereas the lower sec- tion presents three backend question-answering workflows. User-interface, web-server, and visualization components are intentionally excluded from the architectural design. 3.1 Layer 1: heterogeneous data sources The first layer represents both configured and on-demand inputs. Configured sources include RSS news or video feeds, lists of Web pages, arXiv queries, social-media records stored as JSON Lines, and open energy data represented as CSV. These inputs cover the principal media types identified in the original EneMonIE proposal while making each acquisition mechanism explicit. Question-driven Web retrieval is modeled separately because its lifecycle differs from scheduled polling. In- stead of collecting all items from a predeclared source, it transforms a question into an energy-oriented query, obtains candidate results, downloads a bounded number of pages, and submits those pages to the standard seman- tic pipeline. This design prevents the question-answering path from becoming an unfiltered shortcut around En- eMonIE processing. 3.2Layer 2: acquisition and normaliza- tion All configured connectors implement a common fetch operation and return a normalizedDocument. A document contains a stable identifier, source type, source name, title, text, URL, language, publication time, and arbitrary metadata. Source adapters hide format-specific details such as RSS XML, arXiv Atom records, HTML cleaning, JSONL decoding, and CSV row serialization. Normalization is followed by identifier-based dedupli- cation. Within a collection run, documents are indexed bydoc_id; repeated identifiers overwrite earlier occur- rences before semantic processing. At storage time, the same identifier becomes the relational primary key and en- ables idempotent upserts. This two-stage strategy reduces duplicate processing while supporting updated represen- tations of a previously observed source. The orchestrator exposes three operational modes. A 2 Figure 1: Implementation-derived architecture of EneMonIE-Backend. The architecture is conceptually adapted and substantially extended from the original EneMonIE proposal by Küçük [2022]. one-shot mode executes a complete collection and pro- cessing cycle; a watch mode repeats the cycle at a con- figurable interval; and question-oriented modes collect or retrieve only the evidence required for a request. Config- uration controls the database path, relevance and align- ment thresholds, source definitions, model selection, and Web-retrieval limits. 3.3 Layer 3: semantic processing 3.3.1 Energy relevance gate The categorizer tokenizes each document and scores on- tology categories with weighted keyword counts. For category c and document d, the implemented score is S(c, d) = X k∈K c w c,k n(k, d),(1) whereK c is the category keyword set,w c,k is the on- tology weight, andn(k, d) is the number of exact token occurrences. The highest-scoring category is selected. A document is rejected if no category matches or if its max- imum score falls below the configured relevance threshold. Current categories cover renewable energy, grid operations, policy and markets, conventional energy, and research and technology. Placing this gate before the extractors has two bene- fits. First, it makes the stored corpus domain-specific. Second, it avoids running later components on evidently unrelated material. The threshold is externalized because the appropriate precision–recall trade-off depends on the deployment corpus. 3.3.2 Parallel semantic annotations Relevant documents are passed to four analyzers: 1) NER combines energy gazetteers with rule-based patterns. En- tities are assigned labels such as energy organization, location, technology, company, feeder, power plant, or miscellaneous proper name. Canonical lowercase forms support comparison across documents. 2) Temporal ex- traction detects supported date formats and weekdays, then normalizes parseable values to ISO timestamps. 3) Event extraction uses domain trigger lexicons for out- ages, plant commissioning, policy changes, investments, accidents, and research breakthroughs. Confidence is in- creased when a sentence includes reporting cues, a normal- ized time, or sufficient context. 4) Sentiment analysis is applied only to social-media documents. English and French lexicons produce a positive, negative, or neutral label from the difference between positive and negative token counts. 3 These analyzers are logically parallel: each receives the normalized document, although event extraction also con- sumes the temporal output to associate a normalized time. Their outputs are combined into aProcessedDocument. A frequency-based extractive summarizer selects a small number of high-scoring sentences while preserving their original order. Semantic resources are stored outside the processing classes as YAML files. The current resources include the energy ontology, entity gazetteers, event triggers, and English and French sentiment lexicons. This separation enables vocabulary updates without changing orchestra- tion code. 3.4 Layer 4: knowledge organization The knowledge-organization layer operates across pro- cessed documents. First, the social-graph builder creates one undirected edge for each pair of unique canonical entities that co-occurs in a document. Edge weight is the number of documents containing the pair. The result is a compact association graph that can be analyzed or visualized by downstream software. Second, the media interlinker estimates whether two documents discuss the same development. It computes Jaccard similarities over canonical entities (J E ), event types (J V ), normalized calendar dates (J T ), and lexical tokens (J L ): sim(d i , d j ) = 0.35J E + 0.25J V + 0.15J T + 0.25J L . (2) An undirected link is created when the score meets the configured alignment threshold. Connected components of this graph become document clusters. This approach is transparent and inexpensive, although its transitive closure can join documents whose pairwise similarity falls below the threshold; this limitation is discussed in Sec- tion 7.2. Finally, information fusion computes a cluster title, an extractive summary of the combined texts, the five most frequent entities, the five most frequent event types, and a confidence score. For multi-document clusters, confidence is the mean pairwise similarity expressed as a percentage. Singleton clusters receive a confidence of 100 by construction; this value indicates structural certainty rather than independent corroboration. 3.5 Layer 5: relational persistence SQLite is used as a self-contained central store. Table 1 summarizes the schema. Foreign keys connect all semantic annotations to their source documents, while cascade deletion removes dependent rows. Documents and entity edges are updated through idempotent upserts; cluster tables are rebuilt after batch interlinking so that their membership reflects the latest processed set. The database boundary is central to application inde- pendence. A future Web service, notebook, batch report, Table 1: Persistent data model. TablePurpose documentsSource content, metadata, category, sen- timent, and summary. entitiesSurface forms, labels, canonical forms, and character spans. temporalsTemporal mentions, normalized values, and spans. eventsEvent types, triggers, sentences, times, and confidence. entity_edgesWeighted co-occurrence relationships. clustersFused titles, summaries, confidence, enti- ties, and event types. cluster_membersMany-to-many links between clusters and documents. or visualization tool can read the same normalized state without modifying the semantic pipeline. 4 Grounded Question-Answering Paths EneMonIE-Backend extends the monitoring pipeline with three backend question-answering paths. Each returns a structured object containing the question, answer, evi- dence, model identifier, and scope. 4.1 Question-driven Web QA The Web path augments a question with energy terms when necessary, then attempts multiple search providers. Candidate URLs are scraped into normalized documents. Each page must pass the same energy categorization and semantic extraction used during scheduled ingestion. Ac- cepted pages become bounded evidence items containing a title, source, URL, category, score, summary, entities, events, and excerpt. Evidence may optionally be per- sisted, thereby allowing a question-driven discovery to enrich the monitored corpus. 4.2 Database QA The database path retrieves either the full stored corpus or an explicit list of document identifiers. It then applies lexical ranking over the title, summary, and text. For query token q and document d, the contribution is R(q, d) = 5n t (q, d)+3n s (q, d)+I[n b > 0] (1 + ln(1 + n b )) , (3) wheren t ,n s , andn b denote occurrences in the title, summary, and body. The highest-ranked documents form the evidence set. This transparent baseline is reproducible and avoids introducing an embedding index before its value has been measured. 4 4.3 Selected-resource QA The selected-resource path accepts indexes of configured sources, fetches only those sources, processes and stores relevant documents, and delegates answering to database QA constrained to the newly stored identifiers. It there- fore combines controlled acquisition with the consistent retrieval semantics of the central corpus. 4.4 Answer synthesis and graceful degra- dation When a language-model API key is configured, evidence blocks are sent to the selected model with an instruction to answer only from the supplied evidence and to cite bracketed source numbers. When no model is available, or when synthesis fails, the backend returns an extractive evi- dence summary. The fallback is important architecturally: collection, semantic analysis, persistence, retrieval, and inspectable evidence remain available without external model access. The language model is intentionally placed after re- trieval rather than inside the semantic extractors. This limits its authority to presentation and synthesis, pre- serves deterministic structured annotations, and supports independent evaluation of retrieval quality and answer faithfulness. 5Relationship to the Original EneMonIE Design The original EneMonIE proposal [Küçük, 2022] defines the conceptual foundation, while the present architecture in- troduces the concrete implementation choices summarized in Table 2. The table distinguishes retained, concretized, and newly introduced elements. The new architecture pre- serves the semantic intent of EneMonIE while narrowing its boundary to reusable backend services. The exclusion of the application layer is not a reduction of semantic functionality. It is a boundary decision. The original retrieval and visualization goals can be imple- mented by one or more downstream consumers, while the backend remains testable without an HTTP server or graphical environment. 6 Implementation and Verification 6.1 Reference implementation The prototype is implemented in Python and organized around four main packages: ingestion, processing, ques- tion answering, and resources. TheEneMonIEPipeline class composes these packages and owns the central database. Lightweight data classes define documents, enti- ties, temporal expressions, events, sentiment results, pro- cessed documents, and clusters. Configuration is loaded from YAML, making sources and thresholds deployment- specific. 6.2Preliminary Evaluation: Methodol- ogy and Results An offline evaluation assessed relevance filtering, informa- tion extraction, evidence retrieval, and local processing latency. The experiment used a small synthetic fixture to verify implementation behavior; it does not constitute a representative benchmark of real-world energy informa- tion extraction. Dataset and protocol. The evaluation comprised seven English energy-relevant documents: six news-style records and one social-media record. Reference annota- tions covered ten entity mentions, seven sentence-event- type pairs, and five temporal expressions. The question set comprised six questions: five with annotated relevant documents and one deliberately unanswerable question. The relevance threshold was fixed at 1.5, and tempo- ral interpretation used a reference clock of 22 Septem- ber 2026 at 00:00:00. Evaluation used a temporary SQLite database, with document creation timestamps fixed according to input order. No network acquisition or language-model synthesis was invoked. Component evaluation applied the extractors directly to the reference- relevant documents, while pipeline evaluation assessed outputs after relevance filtering. Metrics. Precision, recall, andF 1 were calculated from pooled true-positive, false-positive, and false-negative counts. Entity matching required the exact character span and entity type; event matching required the document, sentence, and event type; temporal detection required the exact span. Normalized-date accuracy measured calendar- date agreement among correctly matched temporal spans. Retrieval recall@5 measured the fraction of annotated relevant documents returned within the first five results for each question. MRR@5 measured the reciprocal rank of the first relevant result, assigning zero when none was retrieved. Both metrics were averaged over the five answerable questions. The unanswerable question was assessed separately. 6:19 PM Quality results. All seven recovered relevant docu- ments passed the relevance gate, yielding relevance recall of 1.000 on this subset. Because the rerun contains no off- topic documents, it does not establish the filter’s ability to reject irrelevant inputs. Table 3 summarizes extraction performance. Entity extraction recovered all ten reference mentions, with pre- cision of 0.588 andF 1 of 0.741. All seven annotated event pairs and all five temporal spans were recovered without additional predictions in those tasks. Normalized-date accuracy was 5/5 = 1.000. Relative expressions, includ- ing “Monday” and “yesterday,” were normalized to 21 September 2026 under the reference clock. On the recovered synthetic cases, the pipeline achieved 5 Table 2: Relationship between the original proposal and the implementation-derived architecture. ConcernOriginal EneMonIE proposalImplementation-derived architecture SourcesNews, videos, social media, scholarly arti- cles, organizational sites, and open data. Concrete adapters for RSS, URL lists, arXiv, JSONL, and CSV, plus question-driven Web search and scrap- ing. FilteringOntology-based text categorization.Weighted keyword scoring with a configurable rele- vance threshold before downstream processing. Information extractionNER, temporal expressions, and energy- event extraction. Explicit rule- and gazetteer-based modules returning typed objects with spans, canonical values, times, and confidence. SentimentOpinion analysis for social-media text.Source-sensitive English/French lexicon analysis ap- plied only to social-media documents. Interlinking and fusionAlign related media, fuse information, and summarize it. Weighted Jaccard similarity, threshold graph, connected-component clusters, mean-similarity confi- dence, and extractive fusion. Knowledge structureSocial-network construction and a central relational database. Entity co-occurrence edges and a seven-table SQLite schema with foreign keys and cascade deletion. Information accessServer-side application and Web re- trieval/visualization. Three backend QA paths with structured evidence; UI, server, and visualization layers are excluded. Model dependencyNot specified.Optional language-model synthesis with deterministic extractive fallback. Table 3: Extraction results on seven recovered synthetic documents. TaskNPRF 1 Entities10 0.588 1.000 0.741 Events7 1.000 1.000 1.000 Temporal spans 5 1.000 1.000 1.000 an entityF 1 of 0.741, an eventF 1 of 1.000, and a temporal- span F 1 of 1.000. Recall@5 and MRR@5 both reached 1.000. Every an- swerable question retrieved all annotated relevant docu- ments within the first five results, with a relevant docu- ment ranked first. The deliberately unanswerable question returned no evidence. These outcomes demonstrate re- trieval behavior on the supplied questions, without estab- lishing general answer correctness or reliable abstention across other queries. Runtime. On Windows 11 with Python 3.13.14, one warmup call per document preceded five measured process- ing calls per document, producing 35 timing observations. Median processing latency was approximately 0.183,ms. These measurements exclude initialization, warmup, stor- age, retrieval, and language-model calls. They concern very short synthetic documents and do not estimate ap- plication throughput. Interpretation and reproducibility. The evaluation script and its JSON and Markdown reports record con- figuration, predictions, extraction counts, query rankings, runtime scope, and reference provenance. The results support the intended behavior of the event, temporal, and retrieval components on the evaluated fixtures. The small synthetic corpus and reconstructed annotations limit broader conclusions. Evaluation on independently annotated real documents with development/test separa- tion remains necessary. Sentiment quality, interlinking, network reliability, and generated-answer quality were outside the scope of this experiment. 4:38 PM 7 Discussion 7.1 Architectural benefits The redesign provides three practical benefits. First, it establishes a single semantic path for scheduled monitor- ing and question-driven discovery. As a result, evidence collected for a question is filtered and represented consis- tently with the rest of the corpus. Second, it separates deterministic semantic state from optional generative syn- thesis. Database rows, event confidence, entity canonical- ization, and cluster membership can be inspected inde- pendently of a model-generated answer. Third, it gives downstream applications a stable relational and object- level contract instead of coupling them to source formats. The architecture is also suitable for incremental mod- ernization. Rule-based NER can be replaced by a trans- former model; lexical ranking can be replaced by hybrid sparse-dense retrieval; SQLite can be migrated to a server database; and connected-component clustering can be replaced by a time-aware or learned method. Because each responsibility is localized, these changes need not alter all consumers simultaneously. 6 7.2 Limitations The current system remains a research prototype and therefore presents several limitations. First, it relies heav- ily on lexical methods. Weighted keywords, gazetteers, and trigger lists are interpretable, but they provide limited recall for paraphrases, spelling variations, and emerging terminology. Additionally, the temporal component sup- ports only a bounded set of explicit date patterns and does not resolve complex relative expressions or event-specific temporal scopes. Another limitation concerns sentiment analysis. Lexicon-based counts do not adequately capture nega- tion, sarcasm, opinion targets, or aspect-specific senti- ment. The system also relies on transitive clustering, which may cause connected components to group chains of documents even when the first and last documents are only weakly related. Moreover, confidence assigned to singleton clusters should not be interpreted as evidence of corroboration. 7.3 Evaluation agenda Future work should evaluate each stage separately and end to end. Relevance filtering can be measured with precision, recall, andF 1 on a stratified energy/non-energy corpus. Entity, temporal, and event extraction require span- and type-level annotations. Interlinking should be evaluated using document-pair and cluster metrics, with attention to time-separated events sharing the same entities. Retrieval can be measured with recall@k, mean reciprocal rank, or normalized discounted cumulative gain. Grounded answer quality should include factual consistency, citation correctness, evidence sufficiency, and abstention when the corpus is incomplete. Scalability experiments should report ingestion through- put, per-document processing time, database growth, pair- wise clustering cost, and the effect of source failures. Be- cause the current interlinker compares all document pairs, approximate candidate generation or blocking by time, entity, and category will become necessary as the corpus grows. Finally, multilingual expansion should move beyond isolated English and French lexicons. Language identi- fication, multilingual entity normalization, cross-lingual clustering, and evaluation on comparable reports are nec- essary to fulfill the original multilingual ambition of En- eMonIE. 8 Conclusion This article introduced EneMonIE-Backend, an exe- cutable backend reinterpretation of the EneMonIE energy- monitoring architecture. The design retains the original semantic objectives while making acquisition, normaliza- tion, relevance filtering, extraction, interlinking, fusion, persistence, and question answering operational. Its five- layer decomposition is complemented by three evidence- grounded QA paths that reuse the same semantic pipeline. The deliberate exclusion of user-interface and server-side application components creates a reusable core that can support multiple future applications. The implementation demonstrates architectural feasibil- ity and passes a deterministic integration smoke test. It does not yet establish large-scale extraction or question- answering performance. The next scientific step is there- fore a component-wise and end-to-end evaluation on la- beled, multilingual energy corpora, followed by targeted replacement of rule-based modules where measured limi- tations justify greater model complexity. Reproducibility Statement The accompanying Overleaf project contains this manuscript, the architecture figure, and the BibTeX database. The described validation is deterministic and does not require an external language model. Live Web acquisition remains dependent on third-party source avail- ability and should be evaluated with recorded URLs and retrieval timestamps. References Andrew Carlson, Justin Betteridge, Bryan Kisiel, Burr Settles, Estevam R. Hruschka, and Tom M. Mitchell. To- ward an architecture for never-ending language learning. In Proceedings of the Twenty-Fourth AAAI Conference on Artificial Intelligence, 2010. Deepayan Chakrabarti and Kunal Punera. Event sum- marization using tweets. In Proceedings of the Inter- national AAAI Conference on Web and Social Media, volume 5, 2011. Xin Dong, Evgeniy Gabrilovich, Geremy Heitz, Wilko Horn, Ni Lao, Kevin Murphy, Thomas Strohmann, Shaohua Sun, and Wei Zhang. Knowledge vault: A web-scale approach to probabilistic knowledge fusion. In Proceedings of the 20th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pages 601–610, 2014. Dilek Küçük. A high-level electrical energy ontology with weighted attributes. Advanced Engineering Informatics, 29(3):513–522, 2015. Dilek Küçük. Monitoring energy trends through automatic information extraction. arXiv preprint arXiv:2201.01559, 2022. URLhttps://arxiv.org/ abs/2201.01559. Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen tau Yih, Tim Rocktäschel, Se- bastian Riedel, and Douwe Kiela. Retrieval-augmented generation for knowledge-intensive NLP tasks. In Ad- vances in Neural Information Processing Systems, vol- ume 33, pages 9459–9474, 2020. 7 Feng Niu, Che Zhang, Christopher Ré, and Jude W. Shav- lik. DeepDive: Web-scale knowledge-base construction using statistical learning and inference. In Proceedings of the Second International Workshop on Searching and Integrating New Web Data Sources, 2012. Bo Pang and Lillian Lee. Opinion mining and sentiment analysis. Foundations and Trends in Information Re- trieval, 2(1–2):1–135, 2008. Alan Ritter, Sam Clark, Oren Etzioni, et al. Named entity recognition in tweets: An experimental study. In Proceedings of the Conference on Empirical Methods in Natural Language Processing, pages 1524–1534, 2011. Alan Ritter, Oren Etzioni, and Sam Clark. Open domain event extraction from twitter. In Proceedings of the 18th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pages 1104–1112, 2012. Jannik Strötgen and Michael Gertz. HeidelTime: High quality rule-based extraction and normalization of tem- poral expressions. In Proceedings of the 5th Interna- tional Workshop on Semantic Evaluation, pages 321– 324, 2010. 8