Guidelines for giving a talk

Major Principles:

  1. The primary purpose of your talk should be to show what is interesting about what you have done. People who find the ideas interesting can then read your paper for details. Do not think of the talk as a substitute for reading the paper (although it should be a substitute for a quick skimming of the paper).
    If you concentrate on the low-level details, you may well find yourself talking to only 5-10% of the audience.
  2. The secondary purpose of your talk should be make your paper more accessible to those who found it interesting. Help the audience visualize what you describe in the paper, and try to bolster their intuitions about what you are doing and why.
    To these ends, your presentation should not be your paper cutdown to fit into the allowed time. Instead, you should take advantage of the different media available in the talk to present information that is difficult to explain in the paper. For example, a complex structure or chart can be more easily understood when the presenter uses pointing to identify the various elements and their interrelationships during his narrative.
    Transparency overlays can dramatically illustrate a variety of differences or changes (they can also get out-of-hand). Judicious use of color can also enhance comprehension of complex structures or charts.
  3. In planning your presentation, a major tactical concern should be to keep the memory load imposed on the audience within reasonable bounds. Remember that although you immediately recognize the various abbreviations, vocabulary, symbols and formulas used in your talk, most of the audience will be much slower to identify their meanings and import. If the audience has to spend too much time making such associations, they will get behind in following your presentation. Also, the audience may lose track of your higher level ideas attempting to leave room to remember your notation.
  4. Provide a balance between what, why, how, and how well. At a previous meeting, many people came away from a group of talks complaining that they knew what the success percentage was, but that they had little idea of what the author was actually trying to do, and no idea of what methods and techniques had been used to accomplish the results. This situation also raised questions about whether the author had gotten good results, or simply good numbers.
    NB: this is not an invitation to short-change the how well portion of the presentation.


  1. The use of abbreviations on transparencies is often necessitated by the limited amount of space available. To make it easier for the audience to follow the presentation, you can point at the abbreviation while speaking its full name (or a variant). For example, if you use the abbreviation "ATIS", you could say "Air Travel Planning" (a variant which is more descriptive and more common than the official name - "Air Travel Information System").
    Remember that although the audience may be expected to know/understand the underlying term or name, a substantial portion of them do not use it as an abbreviation, and thus will not quickly recognize it.
  2. Do not expect the audience to remember much detail from a previous slide. For example, if comparing two approaches, do not summarize each approach on different slides and expect the audience to be able to see the differences. Presenting such comparisons is often the trickiest part of planning a presentation because there is so little room on a transparency (assuming that you are using a font large enough to read) -- you often want to show multiple comparisons/trends, but can't fit all the relevant items on a single slide, but wish to avoid the discontinuity of switching slides.
  3. Present numbers in context. ****
  4. When you present how well your approach worked, try to go beyond percent correct. Try to provide insight into what the numbers mean. Try to characterize what you got right and what you got wrong. This characterization is then useful in predicting whether or not your approach can be improved/extended to cover data you currently get wrong. If your approach is unlikely to be extendable to cover many of the cases you currently get wrong, can it identify those cases so that you could conceive of it handing them off to another system built to handle such "hard" cases.

Examples better than definitions

  1. Skipping lesser detail: Unless there was something controversal, innovative or otherwise noteworthy in your evaluation methodology, much of that detail can be left in your paper. For example, few in the audience will care whether you reserved 10% or 20% of the corpus for test data.

No Theorems, unless revealing