Tips for Writing Technical Papers
Here are the notes from a presentation I gave at the Stanford InfoLab Friday
lunch, 1/27/06, with a few (not many) revisions when I reprised
the talk on 12/4/09. The presentation covered:
As a running (fictitious!) example, suppose you've designed and run
experiments with a new algorithm for external multipass
merge-sort. Your algorithm reduces the complexity from O(n log
n) to O(n), under the premise that it's acceptable to have
some bounded "unsortedness" in the result. You plan to write up the
results for submission to a major conference.
Note: This example was used throughout the live presentation
but I haven't followed through much in these notes. Thus, the notes
include several exercises for the reader.
Titles can be long and descriptive:
or short and sweet:
Here's a middle-of-the-road length, plus a cute name that sticks in
people's minds:
State the problem, your approach and solution, and the main
contributions of the paper. Include little if any background and
motivation. Be factual but comprehensive. The material in the abstract
should not be repeated later word for word in the paper.
(Exercise: Write an abstract for the multiway sort example.)
The Introduction is crucially important. By the time a referee has
finished the Introduction, he's probably made an initial decision
about whether to accept or reject the paper -- he'll read the rest of
the paper looking for evidence to support his decision. A casual
reader will continue on if the Introduction captivated him, and will
set the paper aside otherwise. Again, the Introduction is crucially
important.
Here is the Stanford
InfoLab's patented five-point structure for Introductions. Unless
there's a good argument against it, the Introduction should consist of
five paragraphs answering the following five questions:
- What is the problem?
- Why is it interesting and important?
- Why is it hard? (E.g., why do naive approaches fail?)
- Why hasn't it been solved before? (Or, what's wrong with
previous proposed solutions? How does mine differ?)
- What are the key components of my approach and results? Also include any specific limitations.
(Exercise: Answer these questions for the multiway sort example.)
Then have a final paragraph or subsection: "Summary of
Contributions". It should list the major contributions in bullet form,
mentioning in which sections they can be found. This material doubles
as an outline of the rest of the paper, saving space and eliminating
redundancy.
(Exercise: Write the bullet list for the multiway sort example.)
The perennial question: Should related work be covered near the
beginning of the paper or near the end?
Guideline #1: A clear new important technical
contribution should have been articulated by the time the reader
finishes page 3 (i.e., a quarter of the way through the paper).
Guideline #2: Every section of the paper should
tell a story. (Don't, however, fall into the common trap of telling
the entire story of how you arrived at your results. Just tell the
story of the results themselves.) The story should be linear, keeping
the reader engaged at every step and looking forward to the next
step. There should be no significant interruptions -- those can go in
the Appendix; see below.
Aside from these guidelines, which apply to every paper, the
structure of the body varies a lot depending on content. Important
components are:
We could have an entire treatise on this topic alone and I am surely
not the expert. Here are some random thoughts:
In general a short summarizing paragraph will do, and under no
circumstances should the paragraph simply repeat material from the
Abstract or Introduction. In some cases it's possible to now make the
original claims more concrete, e.g., by referring to quantitative
performance results.
This material is important -- part of the value of a paper is showing
how the work sets new research directions. I like bullet lists
here. (Actually I like them in general.) A couple of things to keep in
mind:
Don't forget them or you'll have people with hurt feelings.
Acknowledge anyone who contributed in any way: through discussions,
feedback on drafts, implementation, etc. If in doubt about whether to
include someone, include them.
Spend the effort to make all citations complete and consistent. Do
not just copy random inconsistent BibTex (or other) entries
from the web and call it a day. Check over your final bibliography
carefully and make sure every entry looks right.
Appendices should contain detailed proofs and algorithms
only. Appendices can be crucial for overlength papers, but are still
useful otherwise. Think of appendices as random-access substantiation
of underlying gory details. As a rule of thumb:
In general everyone writing papers is strongly encouraged to read the
short and very useful The Elements of
Style by Strunk and White. Here's a random list of pet
peeves.
- Just like a program, all "variables" (terminology and
notation) in the paper should be defined before being used, and should
be defined only once. (Exception: Sometimes after a long hiatus it's
useful to remind the reader of a definition.) Global definitions
should be grouped into the Preliminaries section; other definitions
should be given just before their first use.
- Do not use "etc." unless the remaining items are completely
obvious.
- Acceptable: We shall number the phases 1, 3, 5, 7, etc.
- Unacceptable: We measure performance factors such as volatility,
scalability, etc.
- Never say "for various reasons". (Example: We decided not to
consider the alternative, for various reasons.) Tell the reader
the reasons!
- Avoid nonreferential use of "this", "that", "these", "it", and
so on (Ullman pet peeve). Requiring explicit identification of what
"this" refers to enforces clarity of writing. Here is a typical
example of nonreferential "this": Our experiments test several
different environments and the algorithm does well in some but not all
of them. This is important because ...
(Exercise: The above rule is violated at least once in
this document. Find the violations.)
- Italics are for definitions or quotes, not for emphasis (Gries
pet peeve). Your writing should be constructed such that context alone
provides sufficient emphasis.
(Exercise: The above rule is violated at least once in this
document. Find the violations.)
- People frequently use "which" versus "that" incorrectly. "That"
is defining; "which" is nondefining. Examples of correct use:
- The algorithms that are easy to
implement all run in linear time.
- The algorithms, which are easy to
implement, all run in linear time.
- Always run a spelling checker on your final paper, no excuses.
- For drafts and technical reports use 11 point font, generous
spacing, 1" margins, and single-column format. There's no need to
torture your casual readers with the tiny fonts and tight spacing used
in conference proceedings these days.
- In drafts and final camera-ready, fonts in figures should be
no smaller than the font size in the body of the paper.
- Tables, figures, graphs, and algorithms should always be
placed on the top of a page or column, not in the body of the text
unless it is very small and fits into the flow of the paper.
- Every table, figure, graph, or algorithm should appear on the
same page as its first reference, or on the following page (LaTex
willing...).
- Before final submission or publication of your paper, print
it once and take a look -- you might be quite surprised how
different it looks on paper from how it looked on your screen (if you
even bothered to look at it after you ran Latex the last time...).