How do I write a good abstract?

Question

So, I'm sitting here, rewriting the abstract of a paper I hope to submit shortly. The abstract I had was old and not very good.

I have on good authority that an abstract is an important part of a paper, because it is often what makes a reader take a closer look at the paper.

So, how does one write a good abstract that will make people take a closer look? What are things to include and things not to include? What are common mistakes, if any, to watch out for?

By way of example, here the abstract I am rewriting, in its current form. This is for an applied statistics paper. Feel free to critique.

We describe and implement a method to select a Bayesian model for a collection of DNA sequences. This method assumes the DNA sequences are generated from one of a particular class of distribution models. These models capture long range correlation structure among the sites of the DNA sequences. We choose a model from the class by using a simulated annealing search algorithm with a scoring function based on the prior predictive distribution corresponding to the model. We apply this method to model human and mouse Recombination Signal Sequences (RSS). We use the posterior predictive distribution corresponding to the model to predict which of a larger group of sequences are RSS in the context of a cross-validation setup.

ADDENDUM: Please comment on whether there is a standard abstract length limit in your field, unless it is journal specific.

EDIT: Here is a second attempt at an abstract. This is based partly on the discussion at How to Write an Abstract, which is a nice discussion of the main points to think about. The results teaser at the end is a little unorthodox, but it is intended as an inducement for people to look further at the paper. @F'x, do you have any feedback?

Given a specialised set of DNA sequences, it is a biologically
interesting problem to predict which members of a larger set of DNA
sequences belongs to that set. In this paper we consider the
particular example sets of Recombination Signal Sequences (RSS).

Problems of this kind are commonly addressed in the biological
literature. However, we approach this problem by selecting a
Bayesian model for this specialised set. This is an approach that is
rarely used in this context, but as we show, can give good results.

We select our model from particular class of distribution models. These models capture long range correlation structure among the sites of the sequences. We choose a model from the class by using a simulated annealing search algorithm with a scoring function based on the prior predictive distribution of the model. We apply this method to model human and mouse Recombination Signal Sequences
(RSS). We use the posterior predictive distribution of the model to
predict which of a larger group of sequences are RSS in the context
of a cross-validation setup. We summarize the results of the
prediction in figure and tabular form, showing good results. Example
statistic: out of 700,000 candidate sequences, 30 to 50 are actually
RSS. The algorithm ranks these, in descending order, by how likely
they are to be RSS. It ranks almost all the RSS (90\%+) in the top 100.

Answer 1

One of the things that is too much overlooked is that an abstract is read by a more diverse crowd of people than the rest of your paper. So, you have three different goals:

1. Give a take-home message to people who aren't interested enough to read the full paper.
2. Convince the undecided to read it.
3. Make it easier for experts to find it.

Nowadays, #3 is not so much of a constraint: Google and other search engines allow for full-text searching. So, focus on #1 and #2: be concise, crystal-clear, attractive.

---

Enough with the generalities. I love to critique, so I'll give it a try: I'm not in your field, which I think makes me a good test reader for your abstract, but also makes it harder for me to actually suggest actual changes to it. Anyway…

I think it's decent: no catastrophe, but far from enthusiastic. My main comment would be that it lacks a clear statement of the broader issue you are trying to address. You could start with it, something like:

The past few years have seen large advances in the statistical modeling of DNA sequences, mostly based on genetic algorithms. In this work, we explore the efficiency of an alternative and simpler route, and show how to efficiently choose within a class of distribution models.

I have filled in with semi-random keywords/buzzwords, just to give you a sense of how to achieve it. The idea is to put your findings in perspective:

• Roughly, what subfield are you working in?
• What is the main trend in this field?
• How do you position your work with respect to this recent research?

Then trim down the rest of the text:

We showcase a method to select a Bayesian model for a collection of DNA sequences, generated from a specifically chosen distribution model capture long range correlation structure. We choose the model based its prior predictive distribution. Applying this method to model human and mouse Recombination Signal Sequences (RSS), we predict which of a larger group of sequences are RSS in the context of a cross-validation setup.

(I hope I did not lose or betray some of the meaning, it's hard when you don't get the finer points of the text… but you get the idea anyway!)