PaperlayerView exact revision on arXiv

Source-linked AI summary

AskChem: Claim-Centered Infrastructure for Chemistry Literature Synthesis

Bing Yan, Gregory Wolfe, Stefano Martiniani, Kyunghyun Cho

arXiv:2607.28618v1cs.CLcs.AIcs.IRcs.LG

TL;DR

Chemistry literature questions often require assembling and verifying claims scattered across papers, while existing systems mainly return ranked document lists. AskChem instead centers search on provenance-carrying claims and, on cross-paper questions, achieves 100% resolvable DOIs versus 88.3% without retrieval.

  • Problem

    Existing chemistry literature-search systems primarily return ranked document lists, although many questions require assembling specific findings across multiple papers.

  • Method

    AskChem extracts atomic, typed claims grounded by DOIs and quotes or evidence locators, then organizes them through facets, evidence graphs, and agent-accessible interfaces.

  • Results

    100% resolvable DOIs versus 88.3% without retrieval, with the highest citation density among five settings at 18.1 verified DOIs per answer.

  • Takeaways & Limitations

    AskChem provides a practical foundation for chemistry literature synthesis by making findings verifiable against source quotes and DOIs and connected across papers.

  • Takeaways & Limitations

    The corpus covers only a fraction of chemistry, abstracts are shallower than full text, and generated claims, relations, and taxonomy placements can be wrong.

Abstract

from arXiv · show

Chemistry literature synthesis often requires assembling specific findings scattered across many publications, yet existing literature-search systems primarily return ranked document lists. As a result, scientists and AI agents need to locate relevant information, verify their provenance, and assemble cross-paper answers manually. We present AskChem, a claim-centered infrastructure for cross-paper chemistry search. AskChem changes the unit of retrieval from the paper to the provenance-carrying claim: each paper is converted into atomic, typed claims, each grounded by a source DOI and a verbatim quote or an explicit evidence locator. Over this shared claim store, AskChem exposes complementary structures for search and synthesis: a stabilized faceted taxonomy for hierarchical retrieval and browsing, an evidence graph linking claims through relations, and an exploratory living taxonomy that situates indexed papers under scientific principles. AskChem currently indexes 2.4M claims from 147K papers and provides a web interface, as well as REST, SDK, and MCP access for AI agents. On AskChem-Bench, grounding a GPT-5.5 reader in AskChem yields 100% resolvable DOIs, compared with 88.3% without retrieval, and the highest citation density among five tested systems. AskChem is live at https://askchem.org.

1 Introduction

AskChem reframes chemistry literature synthesis around provenance-carrying claims rather than documents, enabling retrieval and inspection of specific findings for cross-paper answers. It organizes a shared claim store through faceted search, an evidence graph, and an exploratory principle-centered taxonomy.

  • Motivation: Chemistry questions often require assembling specific claims scattered across many papers, while document retrieval leaves users to locate, verify, and combine evidence manually.Existing interfaces primarily return ranked document lists rather than claim-level findings.
  • Motivation: Document-centered retrieval does not directly expose evidence, domain organization, or inter-document relationships, and parametric-memory answers may produce plausible-looking fabricated citations.These limitations affect both scientists and AI agents using literature-search tools.
  • Claim-centered infrastructure: AskChem makes the provenance-carrying scientific claim the central retrieval object, representing each claim as an atomic, typed assertion grounded by a source DOI and verbatim quote or explicit evidence locator.Direct claim indexing lets users retrieve findings, inspect evidence, and assemble cross-paper answers without filtering whole documents first.
  • Shared claim-store structures: AskChem provides a stabilized faceted taxonomy for hierarchical browsing, an evidence graph linking claims through typed relations, and a Living Taxonomy organizing claims beneath scientific principles and related concepts.The facets support operational retrieval, the graph supports cross-paper navigation and conflict discovery, and the Living Taxonomy offers an evolving principle-centered corpus overview.
  • Contributions: 2.4M claims from 147K papers are deployed in AskChem’s live claim-centered chemistry literature-synthesis infrastructure.The contributions include provenance-carrying claim representation and complementary structures over the shared claim store.

2 Claim-Centered Representation

AskChem represents literature as provenance-bearing atomic claims rather than whole papers, enabling individual findings to be searched, grouped, linked, and verified. Shared claim identifiers connect faceted taxonomies, typed claim relations, and source metadata across the platform’s deployed corpus.

  • Claim representation: A Claim is an atomic, typed scientific assertion grounded by a source DOI and verbatim quote, with structured fields and extraction confidence.This representation makes the individual finding the searchable, groupable, linkable, and verifiable object.
  • Shared structures: Shared claim identifiers anchor Source records, taxonomy TreeNodes, and typed relation Edges to the same provenance-bearing objects.Sources include DOI, venue, year, citation count, and disambiguated authors; TreeNodes support faceted paths and Edges support claim relations.
  • Deployed index: 2.4M claims from 147K papers spanning 1925–2026 comprise the current live index, with 307K populated taxonomy nodes.The deployed corpus is stored in SQLite with FTS5 full-text search and a vector index.
  • Deployed index: FastAPI serves the web interface and agent-facing APIs for the deployed claim index.Figure 1 and Appendix B provide example claims in the interface and index record.

3 Claim Extraction and Evidence Graph · 4 Stabilized Faceted Taxonomy

AskChem extracts provenance-carrying claims through complementary abstract- and full-text pipelines, then links them in a typed evidence graph for cross-paper reasoning. A stabilized, corpus-derived faceted taxonomy provides persistent navigational views that support hybrid retrieval, browsing, and contextualization of the same claims.

  • 3 Claim Extraction and Evidence Graph: Two extraction pipelines process abstracts at scale and full-text PDFs, with the deeper pipeline capturing hypotheses, limitations, and surprising findings often absent from abstracts.Each extraction call returns schema-validated structured JSON with provenance fields, numeric ranges, and claim content.
  • 3 Claim Extraction and Evidence Graph: 100% of 2.4M claims are source-grounded with a claim type, source DOI, and verbatim quote.The checks establish traceability rather than full semantic correctness.
  • 3 Claim Extraction and Evidence Graph: AskChem adds relation extraction because cross-paper search requires identifying how findings support, extend, derive from, or contradict one another.The evidence graph provides a relational layer over the claim store rather than replacing search.
  • 3 Claim Extraction and Evidence Graph: 97.9% edge-type precision was achieved on 146 decidable edges, within a graph containing 171,342 typed edges.A domain expert verified 148 sampled edges; 143 were correct after excluding two undecidable cases.
  • 4 Stabilized Faceted Taxonomy: The taxonomy induces category paths while digesting papers, then stabilizes them through canonical routing, synonym normalization, and fuzzy clustering.The result is a persistent, corpus-derived set of L1/L2/L3 paths for production retrieval and browsing.
  • 4 Stabilized Faceted Taxonomy: Each claim can receive a 2–5 segment path across views covering reaction, substance, application, technique, mechanism, claim type, data, time, and authors.These operational views organize the same claim store rather than independent datasets.
  • 4 Stabilized Faceted Taxonomy: Hybrid /search combines FTS5 claim-text, paper-level, taxonomy-node, and dense-vector recall using reciprocal rank fusion.Returned claims retain view paths so clients can group results, expand categories, or move from claims into their hierarchies.
  • 4 Stabilized Faceted Taxonomy: Lexical and semantic retrieval find candidate evidence, while facets broaden, group, and contextualize it under stable chemistry concepts.Browse endpoints expose taxonomy nodes with claim counts and temporal overlays; the network view displays the evidence graph, while the living taxonomy organizes papers under principles, theories, models, and mechanisms.

5 Exploratory Living Taxonomy

The exploratory living taxonomy organizes paper-grounded leaves by broader scientific ideas governing contributions, complementing the faceted taxonomy’s claim-focused organization. Its current tree contains 4,931 nodes, covers 1.1M claims across 361K paper placements, and can propose new branches when existing nodes are unsuitable.

  • Principle-centered organization: The living taxonomy organizes paper-grounded leaves under principles, theories, models, mechanisms, and phenomena.It asks which broader scientific idea governs a paper’s contribution.
  • Principle-centered organization: The stabilized faceted taxonomy asks what a claim is about, whereas the living taxonomy asks which broader scientific idea governs a paper’s contribution.
  • Principle-centered organization: 4,931 nodes comprise the current tree, covering 1.1M claims across 361K paper placements.
  • Principle-centered organization: An abstention mechanism proposes new branches when no existing node is suitable.

6 Cross-Paper Search and Demonstration · 7 Evaluation

AskChem supports claim-level cross-paper chemistry search across hybrid structures, exposing structured conditions, evidence, taxonomy paths, and source DOIs through human and agent interfaces. Evaluation finds complete provenance attachment, 97.9% evidence-graph edge-type precision, improved DOI resolution and citation density, and corpus-scale operation, while noting semantic and comparative limitations.

  • 6 Cross-Paper Search and Demonstration: Hybrid search returns claim-level results with structured conditions, verbatim evidence, taxonomy paths, and source DOIs across three complementary structures.Users can browse taxonomy hierarchies, group hits into evidence graphs, or situate papers in the living taxonomy; web, REST, SDK, and MCP interfaces retrieve the same claim identities.
  • 7 Evaluation: 100% of deployed-corpus claims are source-grounded with a claim type, source DOI, and verbatim quote.These checks establish traceability and help detect unsupported generated text, but they do not prove semantic correctness of every extracted claim.
  • 7 Evaluation: 97.9% edge-type precision is estimated by a domain-expert audit of the evidence graph.The stabilized faceted taxonomy is used in production for hybrid-search signals, browsing, grouping, and temporal overlays, although its retrieval gain and placement validity are not isolated.
  • 7 Evaluation: 100% of AskChem-grounded answers contain resolvable DOIs, versus 88.3% without retrieval, with 18.1 verified DOIs per answer—the highest citation density among five settings.AskChem-Bench contains 30 cross-paper questions covering condition aggregation, temporal tracking, and contradiction surfacing; AskChem also achieves the best mean relevance and highest recent high-impact-work coverage.
  • 7 Evaluation: AskChem eliminates DOI hallucination on the benchmark, while Edison Scientific supplies substantially more citation-linked quantitative detail and a slightly higher on-topic rate.AskChem instead offers claim-level, open-data, interactive-browsing, and agent-tool access.
  • 7 Evaluation: 2.4M claims, 307K taxonomy nodes, and 171K evidence edges are jointly queried at corpus scale through the REST schema shared by the web interface, SDK, and MCP server.This deployed service addresses the evaluation question of whether AskChem operates at corpus scale.

8 Comparison with Existing Systems

AskChem differs from established chemistry databases and scholarly search engines by organizing provenance-carrying narrative claims rather than primarily returning structured entities or ranked papers. Unlike answer-generating assistants, it exposes a persistent claim store for reuse.

  • Databases and search: AskChem exposes faceted structure over provenance-carrying narrative claims, whereas Reaxys, SciFinder, PubChem, and ChEMBL provide curated or structured chemical data.Semantic Scholar and Google Scholar primarily return ranked papers.
  • Scientific NLP and assistants: AskChem builds on scientific information and relation extraction, taxonomy induction, retrieval-augmented generation, sentence embeddings, and rank fusion.These components connect claim extraction, graph construction, hierarchical organization, and retrieval to established scientific NLP methods.
  • Scientific NLP and assistants: Unlike answer-generating assistants benchmarked in Section 7, AskChem exposes a persistent claim store for reuse.The passage contrasts AskChem’s infrastructure with assistants that generate answers.

9 Conclusion · Limitations · Ethics and Broader Impact

AskChem presents claim-centered infrastructure for verifiable, cross-paper chemistry literature search, connecting provenance-carrying claims for human users and AI agents. Its limitations include incomplete corpus coverage, potentially erroneous LLM outputs, and evaluation and taxonomy constraints; provenance and community flagging support responsible use rather than replacing primary-source reading.

  • 9 Conclusion: AskChem demonstrates a claim-centered alternative to document-only chemistry literature search.
  • 9 Conclusion: Provenance-carrying claims support verification against source quotes and DOIs, organization through stabilized facets, and cross-paper connections through evidence relations.The deployed system exposes the same claim objects to human users and AI agents.
  • Limitations: The corpus covers only a fraction of chemistry, while abstract extraction is shallower than full-text extraction.
  • Limitations: LLM-generated claims, relations, and taxonomy placements can be wrong, and AskChem-Bench measures groundedness on 30 questions rather than full factual accuracy or user utility.
  • Ethics and Broader Impact: Each claim includes a source DOI and verbatim quote so users can verify LLM-based extraction against the original paper.The interface also supports community flagging.
  • Ethics and Broader Impact: AskChem is intended to assist literature search and synthesis, not replace reading primary sources for critical decisions.It aims to ground AI-assisted chemistry workflows in verifiable citations and reduce citation fabrication.

A Benchmark and Evaluation Details

AskChem-Bench v1.1 evaluates cross-paper chemistry question answering through 30 questions spanning three tasks, using a standardized GPT-5.5 protocol and multiple evaluation metrics. The benchmark and materials are publicly reproducible through AskChem’s API, source release, and indexed dataset.

  • Question bank: AskChem-Bench v1.1 contains ten questions in each of three cross-paper tasks, for 30 questions total.The public benchmark endpoint releases the full question wording.
  • Protocol: GPT-5.5 answers all 30 questions in five settings, with AskChem using 3–4 keyword subqueries, hybrid search, and at most 40 diversified claims before synthesis.Paperclip uses the same rewriter and synthesizer over paper retrieval; Edison Scientific uses PaperQAfamily, and NotebookLM uses Deep Research.
  • Metrics: Evaluation reports DOI existence, citation density, grounded specificity, quantitative grounding, recent impact, and relevance.Citation density counts distinct verified DOIs per answer, while relevance uses judge scores from 3 direct to 0 irrelevant.
  • Reproducibility: The benchmark API provides full questions, methodology, aggregate and task-level results, and recorded index snapshots.Prompt templates and a rerun script accompany the source release, while indexed data is available through the AskChem Hugging Face dataset.

B Representation and Structural Construction

AskChem represents literature as provenance-carrying structured claims and connects them through faceted, evidential, and exploratory taxonomies. Validation gates enforce structured extraction and provenance, while expert auditing found 97.9% precision for typed evidence edges.

  • Claim representation: Each claim is atomic and typed, with a source DOI plus either a verbatim quote or explicit evidence locator.Structured full-paper claims lacking contiguous quotes carry a location and structured evidence locator.
  • Faceted taxonomy: The stabilized faceted taxonomy consolidates persistent L1/L2/L3 paths across reaction, substance, application, technique, and mechanism views.Claim-type, measurement, and time views provide complementary facets, and paths support browsing and taxonomy recall.
  • Evidence graph: 97.9% precision was achieved for typed evidence edges: 143 of 146 decidable edges had the correct relation type.The audit sampled 148 edges; two were excluded as undecidable.
  • Living Taxonomy: The Living Taxonomy contains 4,931 nodes, including 663 open proposed branches, and permits abstention when no explanatory host fits.It organizes principles, theories, models, mechanisms, and phenomena, but expert placement validation remains future work.
  • Validation: Validation checks schema parsing, required provenance, numeric and chemical fields, canonical faceted routing, and confidence-bearing evidence edges.Living Taxonomy placements can abstain rather than force an unsupported host.
Loading 2607.28618v1…