Examples:

<table border="1" summary="This table lists the 32 most significant floods
of the 20th Century grouped into 6 types of floods: Regional, flash,
ice-jam, storm-surge, dam-failure, and mudflow. The majority of the
floods (20) are listed as Regional floods. There are between one and
four floods in each of the other 5 categories.  ">

The summary attribute includes the following information:

• A description of the point of the table, that it lists the 32 most significant floods of the 20th century. The number of floods is not included in the caption, as it is easy to see approximately how many there are by looking at the table. The summary is structured as a full sentence, unlike the caption, which is strucutred as a title.
• The grouping into 6 types of floods and a list of the groups.  This type of grouping information can be difficult to understand for a user navigating this markup structure of the table, and is best described in prose.  This information is obvious to visual users of the table.  It can be included in a caption or paragraph under some circumstances.  In a document where the author finds this information redundant, it is better to use summary, so that the information is easily available to users who cannot see the table.
• Highlights the trends that are visible in the table. "The majority of the floods (20) are listed as Regional floods. There are between one and
  four floods in each of the other 5 categories." This information is easy to see visually, but much more difficult to find by navigating the structure of the table.li>

<caption>Significant Floods of the 20th Century <br />
[M, million; B, billion]</caption>

The caption includes the title of the table, and a key to abbreviations used in the table for Millions and Billions. This information is needed by all users of the table, and so is in the visible caption.

Alternatively, the information in the summary of this table could have been placed in a details section of the caption.

<caption>Significant Floods of the 20th Century <br />
[M, million; B, billion]<details>This table lists the 32 most significant floods
of the 20th Century grouped into 6 types of floods: Regional, flash,
ice-jam, storm-surge, dam-failure, and mudflow. The majority of the
floods (20) are listed as Regional floods. There are between one and
four floods in each of the other 5 categories.</details></caption>

In addition, this table makes use of structural markup, to allow non-visual user agents to expose the table to their users.

<tbody>
<tr>
<th id="type">Flood Type</th>
<th id="map">Map no.</th>
<th id="date">Date</th>
<th id="area">Area or stream with flooding</th>
<th id="deaths">Reported deaths</th>
<th id="cost">Approximate cost (uninflated)</th>
<th id="comments">Comments</th>
</tr>

TH elements with ID attributes are used to associate cells with their related headers. 

<tr>
<th rowspan="5" id="regional" scope="row">Regional flood</th>
<td headers="regional map">1</td>
<td headers="regional date">Mar.-Apr. 1913</td>
<td headers="regional area">Ohio, statewide</td>
<td headers="regional deaths">467</td>
<td headers="regional cost">$143M</td>
<td headers="regional comments">Excessive regional rain.</td>
</tr>

The first cell in this table is also a TH.

<tr>
<td headers="regional map">2</td>
<td headers="regional date">Apr.-May 1927</td>
<td headers="regional area">Mississippi River from Missouri to Louisiana</td>
<td headers="regional deaths">unknown</td>
<td headers="regional cost">$230M</td>
<td headers="regional comments">Record discharge downstream from Cairo, Illinois.</td>
</tr>
<tr>
<td headers="regional map">3</td>
<td headers="regional">Mar. 1936</td>
<td headers="regional">New England</td>
<td headers="regional">150+</td>
<td headers="regional">$300M</td>
<td headers="regional">Excessive rainfall on snow.</td>
</tr>
<tr>
<td headers="regional map">4</td>
<td headers="regional">July 1951</td>
<td headers="regional">Kansas and Neosho River Basins in Kansas</td>
<td headers="regional">15</td>
<td headers="regional">$800M</td>
<td headers="regional">Excessive regional rain.</td>
</tr>
<tr>
<td headers="regional map">5</td>
<td headers="regional">Dec. 1964 - Jan. 1965</td>
<td headers="regional">Pacific Northwest</td>
<td headers="regional">47</td>
<td headers="regional">$430M</td>
<td headers="regional">Excessive rainfall on snow.</td>
</tr>
<tr>
<th rowspan="3" scope="row">Flash flood</th>
<td>21</td>
<td>June 14, 1903</td>
<td>Willow Creek in Oregon</td>
<td>225</td>
<td>unknown</td>
<td>City of Heppner, Oregon, destroyed.</td>
</tr>
<tr>
<td>22</td>
<td>June 9-10, 1972</td>
<td>Rapid City, South Dakota</td>
<td>237</td>
<td>$160M</td>
<td>15 inches of rain in 5 hours.</td>
</tr>
<tr>
<td>23</td>
<td>July 31, 1976</td>
<td>Big Thompson and Cache la Poudre Rivers in Colorado</td>
<td>144</td>
<td>$39M</td>
<td>Flash flood in canyon after excessive rainfall.</td>
</tr>
<tr>
<th scope="row">Ice-jam flood</th>
<td>25</td>
<td>May 1992</td>
<td>Yukon River Alaska</td>
<td>0</td>
<td>unknown</td>
<td>100-year flood on Yukon River.</td>
</tr>
<tr>
<th rowspan="2" scope="row">Storm-surge flood</th>
<td>26</td>
<td>Sept. 1990</td>
<td>Galveston, Texas</td>
<td>6,000+</td>
<td>unknown</td>
<td>Hurricane.</td>
</tr>
<tr>
<td>27</td>
<td>Sept. 1938</td>
<td>Northeast United States</td>
<td>494</td>
<td>$360M</td>
<td>Hurricane.</td>
</tr>
<tr>
<th rowspan="2" scope="row">Dam-failure flood</th>
<td>29</td>
<td>Feb. 2, 1972</td>
<td>Buffalo Creek in West Virginia</td>
<td>125</td>
<td>$60M</td>
<td>Dam failure after excessive rainfall.</td>
</tr>
<tr>
<td>30</td>
<td>June 5, 1976</td>
<td>Teton River in Idaho</td>
<td>11</td>
<td>$400M</td>
<td>Earthen dam breached.</td>
</tr>
<tr>
<th scope="row">Mudflow flood</th>
<td>32</td>
<td>May 18, 1980</td>
<td>Toutle and lower Cowlitz Rivers in Washington</td>
<td>60</td>
<td>unknown</td>
<td>Result of eruption of Mt. St. Helens.</td>
</tr>
</tbody>
</table>

The following table has an unusual layout which will not be obvious to many users who can see it. It might benefit from a visible description explaining to the way the table is laid out, something like "Characteristics are given in the second column, with the negative side in the left column and the positive side in the right column". In this case, it is best to describe the table in visible text. The text MUST be marked up in such a way that assistive technologies can determine programmatically that it is a description for the table.

There are a variety of ways to include this information, such as:

In prose, surrounding the table.  In this case, the author MUST link the prose to the table via the aria-describedby attribute

<p id="foo">In the following table, characteristics are given in the second
column, with the negative side in the left column and the positive
side in the right column.</p>
<table aria-describedby="foo">
 <caption>Characteristics with positive and negative sides</caption>
 <thead>
  <tr>
   <th id="n"> Negative
   <th> Characteristic
   <th> Positive
 <tbody>
  <tr>
   <td headers="n r1"> Sad
   <th id="r1"> Mood
   <td> Happy
  <tr>
   <td headers="n r2"> Failing
   <th id="r2"> Grade
   <td> Passing
</table>

In the table's caption

<table>
 <caption>
  <strong>Characteristics with positive and negative sides.</strong>
  <p>Characteristics are given in the second column, with the
  negative side in the left column and the positive side in the right
  column.</p>
 </caption>
 <thead>
  <tr>
   <th id="n"> Negative
   <th> Characteristic
   <th> Positive
 <tbody>
  <tr>
   <td headers="n r1"> Sad
   <th id="r1"> Mood
   <td> Happy
  <tr>
   <td headers="n r2"> Failing
   <th id="r2"> Grade
   <td> Passing
</table>

In the table's caption, in a details element

<table>
 <caption>
  <strong>Characteristics with positive and negative sides.</strong>
  <details>
   <dt>Help</dt>
   <dd>
    <p>Characteristics are given in the second column, with the
    negative side in the left column and the positive side in the right
    column.</p>
   </dd>
  </details>
 </caption>
 <thead>
  <tr>
   <th id="n"> Negative
   <th> Characteristic
   <th> Positive
 <tbody>
  <tr>
   <td headers="n r1"> Sad
   <th id="r1"> Mood
   <td> Happy
  <tr>
   <td headers="n r2"> Failing
   <th id="r2"> Grade
   <td> Passing
</table>

Next to the table, in the same figure

<figure>
 <dt>Characteristics with positive and negative sides</dt>
 <dd>
  <p>Characteristics are given in the second column, with the
  negative side in the left column and the positive side in the right
  column.</p>
  <table>
   <thead>
    <tr>
     <th id="n"> Negative
     <th> Characteristic
     <th> Positive
   <tbody>
    <tr>
     <td headers="n r1"> Sad
     <th id="r1"> Mood
     <td> Happy
    <tr>
     <td headers="n r2"> Failing
     <th id="r2"> Grade
     <td> Passing
  </table>
 </dd>
</figure>

Next to the table, in a figure's dt

<figure>
 <dt>
  <strong>Characteristics with positive and negative sides</strong>
  <p>Characteristics are given in the second column, with the
  negative side in the left column and the positive side in the right
  column.</p>
 </dt>
 <dd>
  <table>
   <thead>
    <tr>
     <th id="n"> Negative
     <th> Characteristic
     <th> Positive
   <tbody>
    <tr>
     <td headers="n r1"> Sad
     <th id="r1"> Mood
     <td> Happy
    <tr>
     <td headers="n r2"> Failing
     <th id="r2"> Grade
     <td> Passing
  </table>
 </dd>
</figure>

Authors may also use other techniques, or combinations of the above techniques, as appropriate.

The best option, of course, rather than writing a description explaining the way the table is laid out, is to adjust the table such that no explanation is needed.

<table>
 <caption>Characteristics with positive and negative sides</caption>
 <thead>
  <tr>
   <th> Characteristic
   <th> Negative
   <th> Positive
 <tbody>
  <tr>
   <th> Mood
   <td> Sad
   <td> Happy
  <tr>
   <th> Grade
   <td> Failing
   <td> Passing
</table>