A New Neural Search and Insights Platform for Navigating and Organizing AI Research

Marzieh Fadaee, Olga Gureenkova, Fernando Rejon Barrera, Carsten Schnober, Wouter Weerkamp, Jakub Zavrel

Abstract

To provide AI researchers with modern tools for dealing with the explosive growth of the research literature in their field, we introduce a new platform, AI Research Navigator, that combines classical keyword search with neural retrieval to discover and organize relevant literature. The system provides search at multiple levels of textual granularity, from sentences to aggregations across documents, both in natural language and through navigation in a domain-specific Knowledge Graph. We give an overview of the overall architecture of the system and of the components for document analysis, question answering, search, analytics, expert search, and recommendations.

The Challenge of the AI Information Flood

The volume of AI research is growing exponentially. A significant portion of this knowledge now appears on pre-print servers like arXiv and informal blogs rather than just traditional journals. This "information flood" makes it nearly impossible for researchers to stay current using general-purpose academic search engines like Google Scholar, which often lack the domain-specific "know-how" required for AI.

The exponential growth of AI research publications from 2011 to 2020, highlighting the need for specialized discovery tools.

These traditional tools typically focus on broad bibliographic data and don't fully leverage modern NLP techniques—like semantic search or automated question answering—to help users find specific facts or discover experts. The AI Research Navigator aims to bridge this gap by building a tool that understands the specific language and structure of AI research.

System Architecture and Methods

Developed by Zeta Alpha, the platform utilizes a multi-layered technical stack to provide deep insights:

  • Hybrid Search and Neural Retrieval: The system combines traditional keyword-based search (using heuristics like n-gram matching and citation counts) with vector-based neural search. It employs Sentence-BERT (SBERT) and SciBERT to generate embeddings, allowing for semantic retrieval based on meaning. To maintain speed at scale, HNSW (Hierarchical Navigable Small World) graphs are used for efficient nearest-neighbor lookups.
  • Multi-level Indexing: Unlike engines that only index abstracts, this system indexes at three levels of granularity: sentences (31M), "chunks" of ~10 sentences (4.2M), and full documents.
  • Knowledge Graph (KG): Built on Dgraph, the KG links people, concepts, and papers. A custom Named Entity Recognizer (NER) trained on AI concepts automatically populates the graph from new papers.
  • Question Answering (QA): The platform features both Extractive QA (finding specific text spans) and Abstractive QA (generating natural language summaries).
The platform's QA capabilities, showing both extractive and abstractive answers to a research query.
  • Analytics & Recommendations: The system provides visual trend analysis (e.g., comparing the popularity of frameworks like TensorFlow vs. PyTorch) and a modular recommendation engine that balances content similarity with user-specific tags.

The platform introduces several novel approaches to research organization:

  • Granularity is King: Indexing sentences and "chunks" allows the system to answer specific factual questions and find relevant snippets buried in long papers.
  • The "Expert Search" Shift: Instead of just searching for papers, the system uses a document-centric approach to identify people. By embedding a user's query and finding related papers, it can "vote" for authors, identifying experts in niche sub-fields.
  • Discovery Paired with Productivity: The authors argue that discovery must be paired with productivity tools. The platform allows users to tag papers, write notes, and organize research into projects, which feeds back into the recommendation engine to improve personalization.

Results and Scale

The system successfully manages a massive index including 31 million sentences and 740,000 citation records. In experimental benchmarks, the QA system provided relevant answers to approximately 70% of in-domain questions. Furthermore, the authors found that modern Transformer-based contextualized embeddings significantly outperform traditional methods like TF-IDF or LSI for the task of expert search.

Limitations and Future Directions

While powerful, the system faces challenges such as the "cold-start problem" for new papers, where the recommendation engine must rely heavily on content similarity rather than user behavior. Currently, the Knowledge Graph still requires some manual curation, though the goal is fully automated extraction of complex entity relations. Future work will also address factuality, bias, and enhanced collaboration features for research teams.

Summary updated

Abstract

To provide AI researchers with modern tools for dealing with the explosive growth of the research literature in their field, we introduce a new platform, AI Research Navigator, that combines classical keyword search with neural retrieval to discover and organize relevant literature. The system provides search at multiple levels of textual granularity, from sentences to aggregations across documents, both in natural language and through navigation in a domain-specific Knowledge Graph. We give an overview of the overall architecture of the system and of the components for document analysis, question answering, search, analytics, expert search, and recommendations.
BibTeX
@inproceedings{fadaee2020new,
  author = {Marzieh Fadaee and Olga Gureenkova and Fernando Rejon Barrera and Carsten Schnober and Wouter Weerkamp and Jakub Zavrel},
  title = {{A New Neural Search and Insights Platform for Navigating and Organizing AI Research}},
  booktitle = {SDP @ EMNLP},
  year = {2020},
  url = {https://openreview.net/forum?id=Dkhgi9K-uGj},
}
Figure