How to Write an Informatics Paper

This short guide is intended to assist Informatics researchers to write scientific papers, whether these are for conferences, journals, dissertations or some other purpose. Although it is not itself a scientific paper, it is based on an hypothesis:

The key to successful paper writing is an explicit statement of both a scientific hypothesis and the evidence to support (or refute) it.
We will show how this key idea underpins the overall structure of the paper and determines the story it tells.

The Importance of Hypotheses

Informatics is an engineering science. Like other branches of both engineering and science it contributes to the advancement of knowledge by formulating hypotheses and evaluating them. It is not enough merely to describe some new technique or system; some claim about it must be first stated and then evaluated. This claim has the status of a scientific hypothesis; the evaluation provides the evidence that will support or refute it.

Of course, the whole story may be spread across several papers and several authors. For instance, the initial paper about a new idea may not contain all the evidence needed to support a hypothesis; further evidence may be provided in later papers.

In experimental research, hypotheses typically take one of the following two forms:

  1. Technique/system X automates task Y for the first time;
  2. Technique/system X automates task Y better, along some dimension, than each of its rivals;
where the dimensions are typically:
Behaviour: X has a higher success rate then Y or produces better quality outputs, e.g. shorter, easier to understand, more similar to human outputs, etc.
Coverage: X is applicable to a wider range of examples then Y.
Efficiency: X is faster or uses less space then Y.
Useability: Users find X easier to use than its rivals.

A paper may contain one or more hypotheses. However, it is a mistake to try to cover too many hypotheses in a single paper: it leads to confusion.

Explicit hypotheses are rarely stated in Informatics papers. This is a Bad Thing. It makes it hard for the reader to understand and assess the contribution of the paper. If the reader misidentifies the hypotheses then s/he is bound to find the evidence for it unconvincing. If the reader is also a referee or examiner s/he may reject the paper. Worse still, it may indicate that the author is unclear about the contribution of the paper. In the absence of a clear hypothesis it is impossible to know what evidence would support or refute it.

The symptoms of this malaise are commonplace: papers whose contribution is vague, ambiguous or absent; papers with a confused mixture of multiple, implicit hypotheses; evaluations that are inconclusive or non-existent; referee reports that appear harsh or inconsistent. One of the main purposes of this guide is to help to reverse this unhappy situation.

Theoretical papers are usually welcome exceptions: they usually contain both hypotheses and convincing evidence to support them. The hypotheses are the statements of theorems and the supporting evidence is their proofs. This may account for the relatively healthy state of theoretical research in Informatics compared with experimental research.

The Structure of Informatics Papers

There is a default structure for writing an experimental Informatics paper, whether this be for a conference or journal or as a dissertation. When reporting experimental work, you should use this structure unless you have a good reason not to. Theoretical papers have a different default structure, which I hope to include at a later date.

The main parts of an experimental Informatics paper should be as follows. Each part could be a section of a paper or chapter of a dissertation. To reduce clutter I will refer to sections and papers below. Some parts may need to be spread over several sections/chapters, e.g. if there is a lot of material to be covered or it naturally falls into disjoint subparts. Some parts, especially adjacent parts, may be merged into a single section/chapter, e.g., where there is not much to be said or two or more topics are interlinked. Parts marked with * are optional, but you should think hard before deciding to omit them; if there is something that should to be said you should say it.

Ideally, the title should summarise the hypothesis of the paper. The reader should be able to work out what the paper is about from the title alone. Cute, cryptic titles are fun, but unhelpful.

Similarly, the abstract should state the hypothesis and summarise the evidence that supports or refute it. Remember, most readers will not read beyond the abstract, so be sure to include the key points you want casual readers to take with them.

The main purpose of the introduction is to motivate the contribution of the paper and to place it in context. It should also restate the hypothesis and summarise the evidence. It traditionally ends with a short summary of the rest of the paper.

Literature Survey*:
The literature survey is a broad and shallow account of the field, which helps to place the contribution of the paper in context and which supports the motivating remarks in the introduction. Rather than a list of disconnected accounts of other people's work, you should try to organise it into a story: What are the rival approaches? What are the drawbacks of each? How has the battle between different approaches progressed? What are the major outstanding problems? (This is where you come in.)

The background allows previous work to be stated in more technical detail, where this is necessary for a proper understanding of the contribution of the paper.

This part develops the underlying theory of technique(s) or system. Where appropriate, a mathematical style of definitions, lemmas, theorems, corollaries, remarks, may be used.

The techniques that underlie the implementation are formally specified. The requirements of the implementation are given.

Only the final state of the implementation should be described: not a blow by blow history of its development. However, each of the major design decisions should be identified and reasons given for the choices made. You should abstract away from the code and outline the overall structure of the system and the key algorithms in abstract form, e.g. using diagrams or formalised English. You can point to bits of actual code in the appendices, if necessary. A worked example is often helpful.

Evaluation is not testing. Testing is the process of debugging that ensures that the implementation meets the specification. This debugging process is not usually considered worthy of discussion unless either the bug or the debugging process is especially remarkable. The evaluation, on the other hand, is the gathering of evidence to support or refute the hypothesis. If the hypothesis is of type 1 then system X must be applied to a sufficient range and diversity of examples of task Y to convince the reader that it constitutes a general solution to this task. Descriptions of its behaviour, coverage, efficiency and useability must be presented. If the hypothesis is of type 2 then, in addition to this evidence, there must also be a comparison with rival systems along the chosen dimensions. There should also be a brief comparison along the unchosen dimensions, even if this is a negative result for system X; honesty in science is essential and negative results are also important.

A thorough evaluation usually requires large-scale experimentation, with system X being applied to many examples of task Y. To aid the reader's understanding, the result of these experiments are best presented graphically. To verify the hypothesis, the results must usually be statistically processed (Cohen's book "Empirical methods for artificial intelligence", MIT Press, 1995, is a good guide to statistical methods for Informatics researchers. Toby Walsh has also collected some useful resources on empirical methods in Informatics.). It can aid the reader's understanding of the processing of system X to give one or two worked examples before the results are presented in detail. Further details of the results can be presented in an appendix.

I make no apology for the length of this discussion of evaluation. Evaluation is the most important part of the paper as it provides the evidence for the hypothesis. It is also one of the most neglected parts: even being absent in many papers. If this guide succeeds in raising the profile of evaluation than half my battle will be won.

Related Work*:
A narrow but deep comparison is made between system X and its main rivals at their critical points of difference. Its purpose is to explain the differences in behaviour, coverage, efficiency or useability that were identified in the evaluation. Note that related work is different in purpose, position, breadth and depth from the literature survey; both are needed.

Further Work*:
Some unexplored avenues of the research and new directions that have been suggested by the research are identified and briefly developed. In particular, any research that would improve the evidence for the hypothesis or increase its strength or scope should be highlighted.

The conclusion should both summarise the research and discuss its significance. This includes a brief restatement of the hypothesis and the evidence for and against it. It should then recapitulate the original motivation and reassess the state of the field in the light of this new contribution.

The appendices should provide any information which would detract from the flow of the main body of the paper, but whose inclusion could assist the reader in understanding or assessing the research. Appendices might include any of the following: a glossary of technical terms, some technical background that only some readers may require, examples of program code, a trace of the program on one or more examples, more details of the examples evaluated and the experimental results, the full versions of proofs, an index.


This guide has both a descriptive and a normative role: descriptive of best practice in the presentation of Informatics research and normative in highlighting the importance of explicit hypotheses in such presentations. In particular, I have tried to show how the conventional structure of papers describing experimental research should be used to emphasise these hypotheses and their supporting (or refuting) evidence.

I believe that the neglect of explicit hypotheses has caused methodological problems in Informatics. At worst, it leads to work that fails to advance the state of the field. Systems are built with no clear idea of the contribution they will make to the advancement of knowledge. Such systems are described without convincing evaluation, since it is unclear what the purpose of evaluation would be. Readers of the research may not explicitly notice the absence of hypotheses, but will feel some vague unease about the contribution of the research, sometimes summarised in the comment "so what?". It is tragic to witness such talent, energy and opportunity being wasted in this way. I hope this guide can make some small contribution to preventing such waste in the future.

This guide is under development. I would be grateful for any comments on ways to correct, improve or extend it. More stuff on methodology can be found in my AI Research Methodologies course.