NewsReX: Neural News Analysis Library

Updated 9 July 2026

NewsReX is an open-source, efficiency-first library for neural news recommendation that standardizes experimentation and reduces training costs using Keras 3 and JAX.
It integrates multiple neural models and extends to news extraction, bias analysis, and content retrieval, supporting a modular news-intelligence workflow.
Performance benchmarks show improved speed and reproducibility across models like NRMS, NAML, and LSTUR, with potential for multilingual and LLM-integrated extensions.

NewsReX is an open-source, efficiency-first library for neural news recommendation built on Keras 3 and the JAX backend, introduced to address a fragmented experimental ecosystem characterized by heterogeneous codebases, inconsistent evaluation setups, and resource-intensive models (Azevedo et al., 29 Aug 2025). In the related literature considered alongside it, the name is also used as the target platform for broader news-intelligence workflows, including relation extraction, entity-centric exploration, bias navigation, retrieval, and harvesting; this suggests a broader conceptual role for NewsReX as a modular news-analysis stack rather than only a recommendation benchmark (Hong et al., 2023, Bose, 17 May 2026).

1. Research identity and problem setting

The primary NewsReX system is situated in neural news recommendation, where reproducibility and efficiency are recurrent concerns. The library centralizes canonical neural news recommendation baselines, standardizes preprocessing and experimentation, and uses JAX-optimized training and evaluation to reduce wall-clock cost on both datacenter accelerators and consumer GPUs such as an 8GB RTX 3060 Ti (Azevedo et al., 29 Aug 2025). The motivating problem is not only predictive accuracy, but also the cost of reproducing baselines whose implementations are often scattered, incompletely specified, or sensitive to overlooked training choices such as validation splits, negative sampling, batch size, and epoch count (Azevedo et al., 29 Aug 2025).

That problem setting aligns with broader news recommendation constraints identified in survey work: news items are extremely short-lived, requests can exceed 100 per second, response targets can fall under 100 ms, and systems must simultaneously handle cold start, concept drift, session dynamics, source heterogeneity, and bias/fairness concerns (Raza et al., 2020). Socially driven ranking work further stresses recency and latency: importance must often be estimated at publication time, before post-publication behavioral signals have accumulated, and ranking quality depends on identifying rare high-impact items early (Moniz et al., 2015).

A further strand of related work treats NewsReX not as a fixed recommender alone but as an extensible platform. Technical briefs associated with relation extraction, bias analysis, knowledge-graph exploration, and Google News harvesting repeatedly use NewsReX as the integration target for LLM services, retrieval stacks, and structured analytics, thereby broadening its interpreted scope from recommender library to research framework for operational news processing (Hong et al., 2023, Wang et al., 2024, Herrero-Solana, 13 Feb 2026).

2. Core software architecture and model stack

NewsReX operationalizes Keras 3 with the JAX backend to obtain JIT compilation, XLA lowering, batch vectorization through jax.vmap, and precomputation or caching of user and news representations during evaluation (Azevedo et al., 29 Aug 2025). The first epoch incurs a warmup compilation cost, while later epochs reuse compiled graphs and therefore run substantially faster (Azevedo et al., 29 Aug 2025). The library uses Hydra configuration, a modular BaseModel, and a unified interface across models, datasets, and experiments, with processed artifacts such as dataset splits, vocabularies, and embedding weights cached via pickle to reduce I/O overhead (Azevedo et al., 29 Aug 2025).

NewsReX currently ships three foundational neural news recommendation models. All follow a two-encoder pattern in which a news encoder produces an article embedding and a user encoder aggregates clicked-news embeddings. A common scoring function is $s(u,c)=u^\top e_c$ , with predicted click probability $\hat{y}=\sigma(s(u,c))$ (Azevedo et al., 29 Aug 2025).

Model	Encoder characteristics	Reported MIND-small averages
NRMS	Multi-head self-attention for news and user encoders; additive attention for aggregation	NDCG@10 40.58, NDCG@5 34.28, AUC 65.63, MRR 30.98
NAML	Attentive multi-view learning over title, abstract, and categories	NDCG@10 41.65, NDCG@5 35.34, AUC 66.53, MRR 31.91
LSTUR	GRU-based short-term modeling plus long-term user ID embedding, with two fusion strategies	NDCG@10 40.43, NDCG@5 33.99, AUC 65.98, MRR 30.74

For title-based encoders, the paper specifies additive attention over word embeddings $E_T=[e_{w_1},\dots,e_{w_L}]$ , with scores $s_i=v_n^\top \tanh(W_1 e_{w_i}+b_1)$ , normalized weights $a_i=\exp(s_i)/\sum_j \exp(s_j)$ , and article embedding $e_n=\sum_i a_i e_{w_i}$ (Azevedo et al., 29 Aug 2025). For user encoding, attention over clicked-news embeddings $E_H=[e_{n_1},\dots,e_{n_K}]$ can be written as $\mathrm{Attention}(Q_u,K_H,V_H)=\mathrm{softmax}((Q_uK_H^\top)/\sqrt{d_k})V_H$ (Azevedo et al., 29 Aug 2025). The data pipeline supports MIND-style schemas with a news table containing news_id, title, optional abstract, and optional category/subcategory, and a behaviors table containing user_id, timestamp, history, and impression lists with clicked versus non-clicked labels (Azevedo et al., 29 Aug 2025).

Training in NewsReX follows impression-level learning with one positive and multiple negatives. The library describes several objective functions used in news recommendation research: binary cross-entropy over candidates, softmax cross-entropy over an impression, sampled softmax or NCE, and optional BPR for implicit ranking (Azevedo et al., 29 Aug 2025). In the reported ablations, the “Random” negative-sampling strategy shuffles positions within the impression, whereas “Unshuffled” places the positive first; both draw negatives uniformly from the impression’s non-clicked candidates (Azevedo et al., 29 Aug 2025).

The reported implementation defaults are FP32, batch size 16, learning rate $10^{-4}$ , and five epochs on MIND-small, with JAX-optimized implementations of AUC, MRR, NDCG@5, and NDCG@10 (Azevedo et al., 29 Aug 2025). Benchmark timings against NewsRecLib on MIND-small show 9.38 versus 3.74 minutes per epoch for NRMS, 11.66 versus 7.82 for NAML, and 10.73 versus 4.55 for LSTUR, with total experiment times dropping from 66.2 to 39.0 minutes for NRMS and from 73.5 to 46.3 minutes for LSTUR (Azevedo et al., 29 Aug 2025). The paper also reports that Microsoft Recommenders baselines were much slower, at roughly 60 minutes per epoch in the authors’ setup (Azevedo et al., 29 Aug 2025).

The ablation results are unusually central to NewsReX’s contribution. Random validation outperforms chronological validation consistently; for NRMS, NDCG@10 rises from 38.64 to 40.04 and AUC from 63.52 to 65.16 under random validation (Azevedo et al., 29 Aug 2025). Unshuffled negative sampling slightly improves metrics over random shuffling; for NAML, NDCG@10 is 41.76 versus 41.35 and AUC is 66.97 versus 66.45 (Azevedo et al., 29 Aug 2025). Smaller batches tend to perform better, with batch size 16 giving NRMS NDCG@10 40.83 and MRR 31.45, compared with 39.29 and 30.07 at batch size 256 (Azevedo et al., 29 Aug 2025). NRMS also plateaus quickly, with 10–20 epochs reported as sufficient (Azevedo et al., 29 Aug 2025).

Related recommendation literature places these engineering results in a wider methodological context. The survey on news recommender systems distinguishes pointwise CTR losses, pairwise BPR losses, diversity-aware reranking, and bandit-style exploration, and emphasizes that evaluation should extend beyond ranking accuracy to include coverage, novelty, serendipity, calibration, fairness, and societal effects such as filter bubbles and polarization (Raza et al., 2020). Socially driven recommendation work adds a complementary perspective: importance can be modeled as the number of tweets in the two days following publication, utility-based resampling such as SMOTEr or under-sampling is crucial because rare high-popularity items dominate front-page utility, and rank combination with official recommendations can improve early-hour performance (Moniz et al., 2015). Event Extraction-based News Recommendation extends the representation problem further by replacing noisy body-level word encodings with event types, roles, and arguments; in its reported experiments, EENR achieved AUC 0.8764, MRR 0.6491, NDCG@5 0.6629, and NDCG@10 0.6808, outperforming NAML on the reported Chinese securities-app dataset (Han et al., 2021).

4. Extraction and enrichment extensions proposed for NewsReX

One explicit extension path is document-level information extraction. “Relation Extraction from News Articles (RENA): A Tool for Epidemic Surveillance” describes a browser-based React frontend and Python Flask inference service for end-to-end NER and relation extraction over English-language infectious-disease news, using two fine-tuned LLaMA2-13B derivatives—OpenOrca-Platypus2-13B and Mythical-Destroyer-V2-L2-13B—adapted with QLoRA and trained for three epochs on 300 synthetic articles produced by 1-shot prompting with gpt-3.5-turbo-16k (Hong et al., 2023). On 10 randomly selected EPIWATCH articles, Open-Orca-13B achieved NER $P=0.93$ , $\hat{y}=\sigma(s(u,c))$ 0, $\hat{y}=\sigma(s(u,c))$ 1 and RE $\hat{y}=\sigma(s(u,c))$ 2, $\hat{y}=\sigma(s(u,c))$ 3, $\hat{y}=\sigma(s(u,c))$ 4, while Mythical-Destroyer-13B achieved NER $\hat{y}=\sigma(s(u,c))$ 5, $\hat{y}=\sigma(s(u,c))$ 6, $\hat{y}=\sigma(s(u,c))$ 7 and RE $\hat{y}=\sigma(s(u,c))$ 8, $\hat{y}=\sigma(s(u,c))$ 9, $E_T=[e_{w_1},\dots,e_{w_L}]$ 0 (Hong et al., 2023). The associated technical brief proposes this LLM-first, document-level extraction strategy, together with React model selection, max-token control, JSON output, and annotated article views, as a direct design for NewsReX (Hong et al., 2023).

A second extension concerns feed enrichment and corpus normalization. The RSS-quality system based on Feedparser, Scrapy, BeautifulSoup, NLTK, and a one-vs-rest linear SVM over IPTC parent categories defines an item-quality score

$E_T=[e_{w_1},\dots,e_{w_L}]$ 1

and reports an average improvement from 39.98% to 95.62% over 1,598 articles from 110 RSS feeds (Hurtado, 2015). It extracts full content, keywords, bigrams, the largest image by area, author information, and predicted categories, increasing average content length from 388 to 3,211 characters, yielding 1,544 items with categories and 1,440 with authors after enrichment (Hurtado, 2015). In a NewsReX setting, these outputs become higher-quality item representations for downstream search, ranking, and recommendation (Hurtado, 2015).

A third extension concerns editorial support rather than reader-side recommendation. The headline recommender system based on Stanford CoreNLP, a curated database of approximately 90,000 news-domain keywords, and separate Gradient Boosting Trees for Facebook and Twitter shareability ranks terms by a weighted combination of local salience, global popularity, and a named-entity boost, and uses median thresholds of 3.7 for Facebook and 1.7 for Twitter alerts (Szymanski et al., 2017). The reported evaluation gives test MSE 41.8 for Facebook and 37.6 for Twitter, and in a sample of about 3,000 Irish Times articles, 64% of headlines already contained two or more of the top five recommended keywords, while 88% contained at least one (Szymanski et al., 2017). This suggests a plausible editorial-control layer for NewsReX in which recommendation is applied not only to article selection but also to headline composition and distribution strategy.

5. Search, exploration, retrieval, and harvesting blueprints

Several systems extend the NewsReX label into search and exploration. TimeMachine is an entity-centric search and visualization system for large news archives that processed over 12 million Portuguese articles, updated daily, and used a CRF-based NER model bootstrapped from seed annotations on 50,000 news items (Saleiro et al., 2016). It builds entity documents from sentences mentioning an entity, indexes names, aliases, roles, quotations, related entities, and temporal signals, and visualizes co-occurrence networks with SigmaJS and ForceAtlas2, including egocentric views and global time-interval graphs (Saleiro et al., 2016). The technical brief accompanying TimeMachine explicitly presents its entity-document indexing, pattern-based enrichment, and co-occurrence visualization as a template for NewsReX (Saleiro et al., 2016).

NCExplorer provides a different extension axis: OLAP-like roll-up and drill-down over news using external knowledge graphs for due diligence and risk management (Wang et al., 2024). It combines spaCy preprocessing, LDA, Universal Sentence Encoder or Sentence-BERT embeddings, DBpedia Spotlight-based entity linking, fact-based knowledge graphs such as DBpedia, Wikidata, and YAGO4, ontology-based hierarchies, Solr BM25 retrieval, and vector search in Qdrant (Wang et al., 2024). The system supports traversal across organization, industry, geography, ownership, and event-type dimensions, and the paper reports superiority over existing state-of-the-art news search methodologies in empirical studies with master-qualified Amazon Mechanical Turk evaluators (Wang et al., 2024). Within a NewsReX interpretation, this yields an exploration layer in which retrieval becomes semantically navigable rather than only relevance-ranked (Wang et al., 2024).

At infrastructure scale, Cloudpress 2.0 contributes a cloud-native retrieval architecture in which fetching, preprocessing, indexing, storing, retrieval, query expansion, and summarization exploit MapReduce and Hadoop (Raj et al., 2012). It uses dual Lucene indexes, one preprocessed and one with positional information, supports Boolean, wildcard, and proximity queries, stores IDF values, and reports that its parallel crawler was approximately twice as fast as Ubicrawler and Mercator in an 8-node setup, with each RSS feed contributing roughly 100–120 URLs (Raj et al., 2012). The same synthesis recommends elastic scaling, distributed storage, BM25 modernization, and WebGL-based visualization as implementation guidance for a large-scale NewsReX retrieval service (Raj et al., 2012).

Google News harvesting adds a complementary corpus-construction perspective. The IFMIF-DONES case study used Selenium WebDriver, randomized 1.5–4.0 second delays, User-Agent rotation, BeautifulSoup, local LLM extraction with Mistral-NeMo-12b via Ollama, and SCImago Media Rankings enrichment to collect across 81 valid region-language combinations in a five-stage pipeline (Herrero-Solana, 13 Feb 2026). The process involved 737 scripts, reduced 3,340 unique candidates to 1,482 validated records—a 56% reduction—and showed that 76% of Google News records were exclusive to that platform (Herrero-Solana, 13 Feb 2026). The paper also details the operational constraints that matter for a NewsReX harvester: a roughly 100-result cap per query, temporal instability requiring synchronic collection, academic PDFs and non-HTML resources, false positives, and pornographic content infiltrating results through black hat SEO techniques (Herrero-Solana, 13 Feb 2026).

A specialized retrieval extension is given by NewsRECON, which links images to relevant articles when reverse image search evidence is unavailable (Tonglet et al., 20 Jan 2026). It filters a 415,193-article pool down to 91,376 articles, uses a CLIP/ViT-L-14 bi-encoder over image embeddings and automatically generated “news captions,” then applies a location cross-encoder and an event cross-encoder over retrieved candidates or clusters (Tonglet et al., 20 Jan 2026). On TARA in the open-ended setting, NewsRECON reports GREAT $E_T=[e_{w_1},\dots,e_{w_L}]$ 2 59.1%, GREAT $E_T=[e_{w_1},\dots,e_{w_L}]$ 3 49.5%, and overall GREAT 54.3%; when top-3 NewsRECON articles are added as evidence to Qwen2.5VL-7B, overall GREAT rises to 60.4% (Tonglet et al., 20 Jan 2026). For a NewsReX-style platform, this implies that article retrieval can also function as evidence retrieval for multimodal contextualization tasks, not only for ranking or recommendation (Tonglet et al., 20 Jan 2026).

NewsReX-oriented research also extends into interpretive analysis rather than only retrieval. NewsLens formulates media-bias analysis as structured navigation rather than outlet labeling, using a five-agent pipeline composed of a Fact Verifier, Progressive Framing Analyst, Conservative Framing Analyst, Propaganda Detector, and Neutral Summarizer (Bose, 17 May 2026). The system computes a Perspective Divergence Score

$E_T=[e_{w_1},\dots,e_{w_L}]$ 4

and a Manipulation Index in $E_T=[e_{w_1},\dots,e_{w_L}]$ 5, and on 15 articles across Kashmir, Gaza, Climate Policy, and Ukraine reports that center outlets showed the highest mean PDS—0.907 for Qwen2.5-3B on the center group in the reported run—while conservative-framing outlets showed the highest mean MI, approximately 0.600 across models and groups (Bose, 17 May 2026). The associated adaptation brief maps these agents directly onto a proposed NewsReX architecture, with typed shared state, deterministic prompts, and omission maps as the principal output (Bose, 17 May 2026).

Another adjacent capability is notability detection. “Identifying Notable News Stories” defines notable events using Wikipedia’s Current Events Portal, represents candidate stories as triples with time and location metadata, and learns to rank candidates against anchor events using lexical, structural, and entity-based features (Saravanou et al., 2020). On the reported test set, RF $E_T=[e_{w_1},\dots,e_{w_L}]$ 6 achieves MAP 0.56, P@5 0.47, P@10 0.32, NDCG@5 0.64, NDCG@10 0.61, and MRR 0.75; in the binary setting RF $E_T=[e_{w_1},\dots,e_{w_L}]$ 7 reaches MRR 0.79, while adding entity features yields a +0.07 absolute MRR gain over RF $E_T=[e_{w_1},\dots,e_{w_L}]$ 8 (Saravanou et al., 2020). This gives NewsReX a principled route to prioritization when the task is not click prediction but alignment with trusted notable-event anchors (Saravanou et al., 2020).

Influence mapping offers a third analytical layer. NewsCompare combines a polite parallel crawler implemented with Spring Boot and jsoup, Lucene-based “more-like-this” comparison, tri-gram matching, and hyperlink graph analysis to detect similarities and possible allegiances among Romanian news sites (Pop et al., 2019). The broader Romanian web graph reported in the paper has diameter about 15, clustering coefficient about 0.362, average distance about 7.8, and average degree about 3.871, while the news-only subgraph has diameter about 8, clustering coefficient about 0.543, average distance about 3.08, and average degree about 1.687 (Pop et al., 2019). The underlying logic—clusters of sites with strong links to each other and posting similar content—can be transplanted into a NewsReX ecosystem monitor concerned with syndication, agenda propagation, or coordinated behavior (Pop et al., 2019).

The principal limitations of NewsReX as a library are explicit. Its current model coverage is centered on NRMS, NAML, and LSTUR; public validation is on MIND, with multilingual validation supplemented by a proprietary Japanese Nikkei dataset; and future work is said to include more recent news-recommendation models and LLM-integrated news encoders, broader multilingual datasets, privacy-preserving modules, and live testing visualizations (Azevedo et al., 29 Aug 2025). The extension blueprints add further constraints: RENA is English-only and trained on 300 synthetic articles (Hong et al., 2023); NewsLens is based on $E_T=[e_{w_1},\dots,e_{w_L}]$ 9 articles and reports no statistically significant between-group differences under Mann–Whitney testing (Bose, 17 May 2026); Google News harvesting requires substantial methodological investment and robust filtering (Herrero-Solana, 13 Feb 2026); and the news recommendation survey emphasizes persistent risks of filter bubbles, echo chambers, polarization, privacy loss, and content-quality failures (Raza et al., 2020). Taken together, these limitations indicate that NewsReX is best understood not as a closed finished system, but as a reproducible computational nucleus around which richer extraction, retrieval, exploration, and governance components can be assembled.