How to write Zotero research notes you can use in a draft

Attach a source finding to the paper you will cite, record a locator that takes you back to the exact passage, and label your interpretation separately. When you start drafting, the note should let you check a sentence against its source without rereading the entire paper.

Attach the finding to its source

With the paper selected in Zotero, use New Note, then Add Child Note, or the Add option in the Notes tab. Zotero documents these ways to add notes and says notes save automatically. A child note keeps this finding with the paper you will cite.

Start with a sentence the source supports. “Transformers use six layers and can see all words” is too broad: it turns one model specification into a rule for every Transformer and blurs the difference between encoder input and later decoder output positions. Split it into findings you can locate in the paper.

Note 1: What the decoder mask does

Decoder mask: the prediction cannot see later output positions

• Finding: The decoder self-attention masks subsequent positions. The shifted output embeddings limit a prediction to earlier known outputs.
• Locator: Attention Is All You Need, arXiv:1706.03762v7, section 3.1, Decoder paragraph; corroborating mechanism in section 3.2.3, final bullet.
• Source: Version 7, section 3.1
• Interpretation: When explaining the architecture, separate the encoder from the decoder. A statement about attending across the encoder input should not be transferred to future decoder output positions.
• Boundary: This note describes the architecture in this paper. It does not establish how every later model handles context.
• Search terms: decoder-mask future-position

The finding is a paraphrase, not a quotation. The interpretation tells you how to avoid a drafting mistake; it is your reasoning about the source, not a sentence to attribute to the authors.

Note 2: Which model has six layers

Six layers in this model specification

• Finding: Section 3.1 specifies six identical layers for the encoder and six for the decoder.
• Locator: Attention Is All You Need, arXiv:1706.03762v7, section 3.1, Encoder and Decoder paragraphs.
• Source: Version 7, section 3.1
• Interpretation: Keep the layer count attached to this paper and version when taking notes on implementation details.
• Boundary: Six layers is not a definition of all Transformers.
• Search terms: layer-count model-specification

Use the version and section to return to the specification. A note title or PDF viewer’s page index can omit that context. If you work from a PDF, record its printed page as well; do not guess one from the viewer. If the HTML anchor stops working, find the named section in the specified version.

You can download the two completed notes and adapt their field labels in your own library. Keep each Finding short enough to check against its Locator. Put an unanswered question in a separate line or note rather than making it sound like a result. Use quotation marks only for wording you have checked against the source; the findings above are paraphrases.

Keep an open question from becoming a claim

Suppose you are using the second note to explain an unfamiliar model. The source finding establishes the layer count in this paper. It does not answer whether a later implementation uses the same count. Put that question in a separate entry: “Does the implementation I am describing use the six-layer configuration in this paper? Find its own architecture specification before stating a count.”

You can keep that question beside the finding while you read another source. Leave its status as a question until you have a locator for the answer. Replacing the question with “All Transformers have six layers” would go beyond the note’s finding even though the original paper remains attached.

The first note has a similar boundary. “The decoder cannot attend to later output positions” identifies a component and a direction. “The model cannot see anything later” drops both details and invites a different reading. Keep “decoder” and “output positions” in the sentence you draft, then check section 3.1 again if you need a broader explanation.

An interpretation can still be useful. The advice to separate encoder and decoder descriptions tells you how to write the next paragraph. Labeling it lets you keep that advice without presenting it as a quotation or an experimental result from the authors.

Retrieve the note, then write

Zotero documents library-wide note search and Ctrl/Cmd+F in the note editor. Give each note a distinctive term you would plausibly remember while drafting. In the downloadable Markdown, a text search for decoder-mask matches the first note only; layer-count matches the second only. That is a check of the supplied file’s text, not a record of a Zotero search.

After finding a note, follow its source link and locator before using the claim. A draft might say: “In the architecture described by Vaswani and colleagues, the decoder mask prevents a position from attending to later output positions. The six-layer count describes the authors’ model specification.” Cite the parent paper for those claims. The note’s Interpretation can guide your explanation, but it should not become a claim that the authors made.

For a comparison involving several papers, a standalone note may be easier to own than a child note under one paper. Zotero’s Related tab can link that note to records in the same library; the relationship is reciprocal, but it does not automatically extend to other related items. Keep each source’s finding and locator distinct before writing the comparison.

Source note: The two completed examples and the one-match-per-term retrieval check come from the supplied Markdown file. Zotero actions above are documented instructions, not an independently observed Zotero interface test. Paper claims refer to the linked arXiv version, first submitted June 12, 2017, and revised August 2, 2023.