Photon – a live demo of a natural language interface to databases

&quot;What is the price of Ginger Beer?&quot;<p>It couldn&#x27;t translate that into SQL.<p>&quot;What is the price in dollars of Ginger Beer?&quot;<p>&gt; SELECT Catalog_Contents.price_in_dollars FROM Catalog_Contents WHERE Catalog_Contents.price_in_dollars = &quot;Ginger Beer&quot;<p>Nope.<p>&quot;What is the price in dollars of catalog entry name Ginger Beer?&quot;<p>&gt; SELECT Catalog_Contents.price_in_dollars FROM Catalog_Contents WHERE Catalog_Contents.catalog_entry_name = &quot;Ginger Beer&quot;<p>Cool! You have to be more specific than I was hoping, but this is still pretty neat.

There’s a lot missing here. During the brief and unhappy period of my life where I worked in this area, we had quite a lot of luck just generating semantics based on Wordnet in the domain in question. So here you can’t successfully ask for “French wines” even though we know what a country is and that French is a correct adjectival form. Same with things like “oldest wine”, that’s an easy to derive superlative based on info you already have. We got some mileage out of this old fashioned tree based system at the core, with fuzzier machine learning stuff at the edges.

&gt; What singer sang in the most stadiums?<p><pre><code> SELECT singer.Name FROM singer JOIN singer_in_concert ON singer.Singer_ID = singer_in_concert.Singer_ID GROUP BY singer.Singer_ID ORDER BY COUNT(*) DESC LIMIT 1 </code></pre> It is close... sort of? It figured out it needed to join, group, and order, but it only drew the relation to the concert, not the venue. Correctness seems a huge challenge here. Even knowing SQL, I feel I&#x27;m double checking my results at times. But I can see how this might be incredibly useful someday for Salesforce if there&#x27;s confidence in the results.

&gt; how many teachers older than thirty?<p>&gt; SELECT COUNT(*) FROM teacher WHERE teacher.Age &gt; &quot;thirty&quot;<p>Not a bad idea. A good idea, maybe. Implementation needs some work.

The Covid database contains cumulative figures, so if you ask &quot;How many deaths in ...?&quot; you get the naive query with the sum of the Deaths column for that country, which is wrong. Actually I wonder how to explain it. I cheated and asked &quot;how many deaths in ... on July 14?&quot; but got the wrong query, with July 14 as Province_or_State no matter how I rephrased the date.

“Select all elements from catalog” did not work...

&gt; All confirmed covid deaths<p>&gt; SELECT covid_19_july_data.Deaths FROM covid_19_july_data WHERE covid_19_july_data.Confirmed = &quot;covid&quot;

I&#x27;ve seen a talk about this task, it&#x27;s supposed to be hard to get enough training data for regular DL approaches.

In my opinion this doesn&#x27;t make sense.<p>SQL is a tight, unambiguous language, that&#x27;s why it exists.<p>This is like a legal document written in spoken English. It&#x27;s only all fine when it works.<p>Part of writing SQL is also understanding the underling data. This won&#x27;t address this issue.<p>This is also not replicable. Language changes in context and time.

see also <a href="https:&#x2F;&#x2F;news.ycombinator.com&#x2F;item?id=24283687" rel="nofollow">https:&#x2F;&#x2F;news.ycombinator.com&#x2F;item?id=24283687</a>

Doesn&#x27;t do anything related to language

list all company<p>&gt; Please check the results in the table. Did I get it right?<p>yes<p>&gt; Great!<p>list all designation<p>&gt; Sorry, &#x27;designation&#x27; is confusing to me

meh &quot;is there any area column in addresses table&quot; did not work seems straight forward to me ?