Writing the Best Paper (Vol 5) - How to Write a Survey Paper? Research Tutorials with Dr. Sourish

Survey papers succeed or fail on focus and comparison, not on how many papers get listed. The core message is that a strong survey paper picks one clear focus (from a defined set), stays contemporary within that focus, and builds a structured, multidimensional comparative analysis that clearly shows how studies reinforce, extend, or contradict each other—while ending with gaps and future directions. That matters because survey papers often become the entry point researchers use to find relevant work, which can translate into higher visibility and citations.

The session breaks survey papers into three primary types—then adds narrower variants. One broad category surveys the contemporary landscape: it maps recent research questions around a topic (e.g., within NLP, questions like sentiment analysis, document similarity, language modeling, or event extraction). A second broad category organizes by approach: it collects the different methods or paradigms used to address a specific research question (e.g., sentiment analysis on Twitter using different modeling paradigms). A third broad category organizes by evaluation method: it compares how the same research question has been tested, emphasizing the diversity of evaluation designs and setups. Beyond these, the talk describes narrower survey styles that focus on (1) variants of a research question across settings (Twitter vs. Facebook vs. Amazon reviews), (2) variants of an approach (e.g., different deep learning architectures for sentiment analysis), or (3) variants of an evaluation technique while keeping the evaluation style consistent.

A key warning follows: weak survey papers typically suffer from mixed focus, lack of contemporary coverage, and weak or missing critical comparison. The most common failure is “touching on everything,” where the writing becomes a reference dump rather than a coherent synthesis. Another major weakness is insufficient multidimensional structuring—meaning the survey doesn’t lay out a blueprint of the dimensions that define the chosen focus. For example, if the focus is research questions, the structure should include variations in formulation, assumptions, and scope (what is inside vs. outside the study boundary). If the focus is research approach or evaluation, the structure should include variations in originality, method details, evaluation setup (controlled environment, parameters, qualitative vs. quantitative procedures), and resulting findings.

Comparison itself must be explicit and categorized. The session frames comparative study as three types: reinforcement (supporting a prior claim), augmentation (extending it—either by adding complementary contributions or by extracting new insights from the same evidence), and contradiction (ranging from unfavorable judgment to evidence-backed contradiction, and finally to proposing an alternative that is argued to be better). The talk emphasizes that these comparison types should be crystal clear in the writing, not implied.

Practical “good survey” habits are then laid out: state the primary focus in the introduction; make each focus dimension its own section; ensure a substantial portion of the cited work appears across those sections (suggested 35–40%); include critical opinions during comparisons; and use summary tables to keep the synthesis readable. If the focus is approach or evaluation, the tables should capture study samples/data sets, control or evaluation environments, parameters, and outcomes. The conclusion should enumerate gaps and future directions tied directly to the chosen focus.

Finally, the session connects these ideas to Raxter (Raxter.io), describing how it can help manage the exponential growth of reading lists by filtering related papers based on selected sections (problem statement, approach, or key abstract ideas), attaching papers to specific survey sections, comparing sections from a collection, and using “key insights” to extract the most relevant dimensions of a paper quickly when time is limited.