This is an automated email from the ASF dual-hosted git repository.
aradzinski pushed a commit to branch master
in repository https://gitbox.apache.org/repos/asf/incubator-nlpcraft-website.git
The following commit(s) were added to refs/heads/master by this push:
new 6604cbf WIP.
6604cbf is described below
commit 6604cbfc5d4c56197d7c03d79c2b64420b683ef1
Author: Aaron Radzinzski <[email protected]>
AuthorDate: Thu Jul 2 12:32:40 2020 -0700
WIP.
---
examples/sql_model.html | 31 +++++++++++++++++++------------
1 file changed, 19 insertions(+), 12 deletions(-)
diff --git a/examples/sql_model.html b/examples/sql_model.html
index a4d3317..8d2412f 100644
--- a/examples/sql_model.html
+++ b/examples/sql_model.html
@@ -42,14 +42,15 @@ id: sql_model
<section id="background">
<h2 class="section-title">Background</h2>
<p>
- Many (if not most) of modern existing natural language-to-SQL
implementations use variations of
+ Many of modern existing natural language-to-SQL implementations
use variations of
deep learning approach where you first train the neural network on
a pre-created training set and
- then get to use the trained network to provide (infer)
probabilistic answers for the new sentences. Although the latest
+ then get to use the trained network to provide (infer)
probabilistic answers for the new input sentences. Although the latest
attempts to implement this approach are getting into 90%
percentile of accuracy - they remain largely
unusable for the vast swath of enterprise applications where
non-deterministic nature of such systems
- renders them useless. The apparent problem with non-deterministic
systems like that is that the user
- never knows whether a given answer is correct or incorrect, and
they can't tolerate the fact that
- in 5-10% of the cases the answer will be invalid. Try that for you
HR or revenue reporting system,
+ renders them practically useless. The apparent problem with
non-deterministic systems like that is that the user
+ never knows whether a given answer is correct or incorrect. In
these business and enterprise
+ use cases users can't tolerate the fact that
+ in 5-10% of the cases the answer will be invalid. Try that for
your HR or revenue reporting system,
prescriptive analytics systems, and so on...
</p>
<p>
@@ -57,18 +58,24 @@ id: sql_model
not pose a significant problem. We happily accept such imprecision
when asking for direction on our
mobile devices, unlock our phones using face or fingerprint
recognition, when performing sentiment
analysis or trying to detect faces of our friends across thousands
of photographs. Cost of retries,
- as well as the cost of initial errors, is insignificant in these
cases.
+ as well as the cost of initial errors, is insignificant in these
cases. The same cost, however, in many business
+ and enterprise applications can be too significant to tolerate.
</p>
<p>
- As you may have learned already, NLPCraft uses a <i>fully
deterministic approach</i> in an attempt to
- match the user input with a defined set of intents. If it finds
the matching intent - it guarantees that
- match for a given intent declaration. If no matching intent can be
found - it always returns the negative
+ As you may have learned by now, NLPCraft uses a <i>fully
deterministic approach</i> in an attempt to
+ match the user input against a defined set of intents. If it finds
the matching intent - it guarantees that
+ match for a given intent declaration. If no matching intent can be
found - it returns the negative
result without any ambiguity. In other words, if the answer is
given, it is guaranteed to be correct.
+ Another positive side-effect of this approach is the fact that
such matching logic is easily traceable, i.e.
+ the user can easily see why given user input was matched against a
certain intent (and not any other). Such
+ traceability of the comprehension logic (or explainability vs
"black box" approach resulting from deep
+ learning techniques) is often critical for many real-life business
applications.
</p>
<p>
- One of the downsides of such an approach in general is the need to
have a domain-specific semantic model
- for each SQL database (when building natural language interface to
SQL databases). Building such a
- model can be a non-trivial experience. That is where, luckily,
NLPCraft brings a lot of built-in tooling
+ On the flip side, one of the downsides of such an approach in
general is the need to have a detailed,
+ domain-specific semantic model
+ for each SQL database (when dealing with SQL databases). Building
such a
+ model can be a non-trivial and time consuming experience. That is
where NLPCraft brings a lot of built-in tooling
and machinery to simplify this task dramatically.
</p>
</section>
