Download as txt, pdf, or txt
Download as txt, pdf, or txt
You are on page 1of 1

given a problem definition, what should be the problem goal. a target category.

do
we enhance our understanding using gradient descent process. Is practicing
regularly
a gradient descent process.
Intelligence might be statistical learning through time. As a fours years old migth
not understand the danger of pushing someone into water, however, the same child
might
understand it when he or she will become 7 or 8 years old. As the child observe the
world more he becomes more aware of it. So Machine learning using statistic might
be the right path.
Now, the issue is what should we learn and how? Defining the goal for a human being
is an easy task but the same is not true for computer system. So far we are busy on
defining different set of problems and trying to solve
them, but one of the important issue suppose to be how a system could define its
own goal given a problem definition?

When I tried to solve the problem of doctors's note to ICD10 assignment, initially
I did not have the exact idea on how to proceed. Therefore, I used a (X,y) mapping
where
X is the doctors text explaining about the patient condition and y are the possible
ICD10 codes. After the experiment, I realized that if I try to solve the problem
in this way then I am going to need a lot of annotated data which is expensive to
acquire. I also realized I could not apply the technique of like self supervision,
because I still need the annotated codes. After thinking about the problem for long
enough I realized that what if I could just detect the key text component
which are related to specific types of ICD10 codes. So, every ICD10 codes represent
a disease, each of these disease happens becuase of the presence of a set of signs,
symptomps and conditions. These attributes are mentioned in the doctors note.
We can collect ICD10 disease related conditions, sign, symptomps from Medical
knowledge bases like UMLS and form a subset of knowledge base. So we reduces the
problem into a
finding out a data structure for each significant text components in doctors note.
We tried to find a text component from doctors note and started assigning various
types of attributes
such as whether it is a symptomps, medical conditions, negated, related set of
ICD10 codes and probabilities of association. In this way, using a small dataset I
got much better precision
and recall.

Finding a informative data structure for a given text input is not new, For
instance, given two input text related to protiens, we can try to find out the
predicate that
describes whether they treat each other or cause each other.

I would like to extend this idea that can be useful for artifical conversation or
chatbot for instance in google home or alexa applications. For instance,
given two sentences as input we can find out whether they are related or not using
BERT or RNN like models. Now, instead of finding out binary or multi class
relationship based output,
I like to try to find out an object structure, that is going to generate various
meta semantic information regarding the sentences along with the relationship
between object structure.
These object structure may contains predicate information, subject object, how they
related using predicate head,
the adjective modifying the predicate head, whether the predicate negated. These
types of discrete meta information is useful for computers to utilize in various
problems.

You might also like