Turn a research workflow into inspectable evidence
After retrieval, three difficult jobs remain: explaining selection, understanding methods faithfully and identifying researchers correctly. This guide uses an original example, comparing methods for detecting duplicate software bug reports. It contains no invented paper, result, author profile or executed search. The intended outputs are a review protocol and evidence matrix, bilingual reading notes and source-linked author cards.
Identify the actual dsh-scholar repository and revision first. Similar names do not establish the same plugin, and a research workspace is not a universal database connector. Prepare the question, search ledger, legally accessible source material, output location and citation workflow. Keep machine preparation, human review and publication decisions separate. An installation command cannot supply the missing methodology.
Scenario one: define a protocol before drafting a review
For duplicate reports, specify whether inputs include only text or also stack traces, whether evaluation crosses projects, and whether the outcome is retrieval ranking or duplicate classification. Define the period, sources, eligibility rules and exclusion reasons before screening. Version the protocol. Papers cannot be compared indiscriminately merely because their titles mention duplicate reports.
Retain complete queries, retrieval dates, filters and pagination for each source. Preserve relationships between paper versions. Count records and independent studies using explicit units. Screen titles and abstracts, then verify eligible full text, retaining exclusion reasons and disagreement resolution. Extract the question, dataset, method, baseline, metrics, split design, limitations and evidence location. A missing field remains missing rather than being completed from the model's expectations.
For example, a candidate described as using a temporal split requires inspection of the actual training and test boundary. Another using random partitioning cannot automatically be called better or worse. Dataset composition, metric definition, duplicate handling and baseline conditions affect interpretation. Separate comparable settings from incompatible ones instead of assembling the largest numbers into a leaderboard. These are extraction requirements, not claims about real studies.
Calibrate the rules on a small set and revisit disputed decisions. Active-learning tools can prioritize screening, but do not replace inclusion decisions or a justified stopping rule. ASReview is a documented public project worth evaluating for that supporting job; its repository identifies Apache-2.0. Check the actual version, formats and recovery requirements before adoption. It was not run here and is not presented as a verified SkillPI installation entry.
PRISMA 2020 provides reporting checklists and flow-diagram references, not certification that every relevant study was found. Without a defensible reference universe, do not claim a general recall percentage. A hit proportion against a known reference set must disclose that set, denominator, scope and limitations. Record supplementary citation chasing and how newly found studies affect the synthesis.
Before prose, connect each proposed claim to evidence. State a bounded conclusion, then its support, contradictions and applicable conditions. A research gap is a candidate interpretation within the coverage achieved. Missing results may reflect query vocabulary, language restrictions or unavailable full text rather than an unexplored scientific problem. The deliverable is a protocol, screening record, extraction matrix, claim map and explicit unresolved questions.
Scenario two: preserve meaning in bilingual close reading
Check whether the PDF yields reliable text. Columns, scanned pages, notes and broken lines can alter reading order. Preserve the original and compare extracted text with page images where needed. Start with one methods paragraph, one table and one equation before processing an entire document. A file that opens successfully is not evidence of complete content or layout.
Create a glossary for terms such as duplicate report, candidate retrieval and ranking metric. Record the context of an approved translation and keep usage consistent, documenting genuine exceptions. Preserve negation, uncertainty and conditions. A statement that a method may improve an outcome must not become a guarantee, and an association must not become a causal assertion.
Check equation symbols, subscripts, denominators, ranges and numbering against the original definitions. For tables, check headers, units, notes and missing-value symbols; for figures, check legends, axes and references in the prose. Mark unreadable material for review instead of substituting a familiar formula. Separate translated author text from commentary so readers can distinguish source claims from explanations and questions.
Organize reading notes by question, method, input, evaluation, finding and limitation, with page or section references. Verify a strong abstract claim against the relevant experimental setup and discussion. Label a criticism as unverified when appropriate. For HTML or PDF delivery, inspect beginning, middle and end plus dense table and formula pages. Check fonts, wrapping, links and images. If only a text comparison is produced, describe that accurately rather than promising lossless original layout.
Acceptance requires more than fluent translation. A note should lead back to its source passage, terminology should be consistent and important limitations should remain visible. Keep abstract-only notes when full text is unavailable; do not fabricate a complete page-by-page reading. Translation assists understanding while citation accuracy and scholarly judgment remain the researcher's responsibility.
Scenario three: disambiguate scholars before mapping relationships
A name is not a unique identifier. Begin with an author on a verified publication and compare stable records, public homepages, institutional pages and identifier links. Identical spellings may refer to different people and affiliations change over time. Retain source and access date. Preserve conflicting information instead of choosing whichever identity best fits the desired story.
For the example topic, construct a timeline of relevant publications, coauthors and public projects. A coauthorship edge establishes a recorded publication relationship, not a private relationship, current willingness to collaborate or quality ranking. Database coverage and author disambiguation differ. An absent edge does not establish that two people never collaborated.
For collaboration candidates, separate topic fit from identity verification. Explain the relevant verified work and outstanding uncertainty. Reviewer suggestions must also satisfy the destination venue's eligibility and conflict rules; a generated name is not automatically an eligible reviewer. This guide does not contact scholars or produce unsupported judgments about people.
Deliver source-linked author cards, a topic timeline and explicit candidate reasons without collecting irrelevant private information. Refresh by comparing changed sources and identifier relationships, rather than replacing a profile blindly when a new paper appears. A profile is a bounded research lead, not a comprehensive evaluation of a person.
Skill roles and a practical request
Notion Research & Documentation fits optional synthesis of already reviewed notes in authorized Notion pages. It requires connected tools, readable sources and a named writable destination. It can organize evidence into a comparison or brief. It is not an external academic searcher, translation-layout engine or author-identity certification tool. Its output does not establish full-text reading or completion of a systematic review. Pinned source review, scanning and observed execution are distinct signals.
A bounded request is: “For duplicate bug-report detection, propose a review protocol and extraction fields first. Use only supplied or legally accessible material, distinguishing abstract and full-text evidence. Prepare a close-reading sample preserving equations, tables and uncertainty, with source locations. Author cards must use verified public identities and sources. Return a matrix and open questions before drafting. Do not invent studies, contact people or publish.” Review each artifact before scaling the task.
Troubleshooting and FAQ
If screening counts disagree, inspect counting units, duplicates and versions. If translation skips text, inspect column extraction and page order. If an equation changes meaning, return to the page image. If names collide, verify identifiers and works. When literature is added halfway through writing, update the search record, matrix and affected claims rather than appending a citation without reconsidering the argument. Under a tight deadline, narrow the declared scope instead of describing incomplete work as exhaustive.
Can automation prove the review is complete? It can document actual queries and checks against a stated reference set, not guarantee universal coverage. Can it identify a gap? It can suggest a hypothesis that needs evidence and alternative explanations. Can a translation replace reading? It is an aid requiring source checks for important claims and technical details. Can a graph choose reviewers? It supplies leads; eligibility and conflicts need separate verification.
Sources and continuing review
References are PRISMA 2020, the ASReview public project and DBLP's search API documentation, with both original-language articles retained in the editorial record. The example is an original teaching design, with no fabricated result, recall rate, translation run or personal profile. Deliver versions, sources, actual review scope and unresolved issues. Revisit affected claims when literature or tools change.