1

We are writting a demo app in which we make use of the Google's cloud Speech API and Natural Language API. We are using it for voice control in TV sets and STBs.

Overall I must say that the APIs are working excellent.

However we are wondering if it would be possible to customize the Natural Language API to be more specific in our industry.

More specifically we had problems with the salience score that is passed to specific entities.

E.g. Passing the following text data to the Google Cloud API: "Show me movies like the Zoolander."

Gives us the following results:

Entity name | Entity type | Metadata | Salience

movies | WORK_OF_ART | {} | 0.7970062 Zoolander | WORK_OF_ART | {'mid': '/m/01shy7', 'wikipedia_url': 'https://en.wikipedia.org/wiki/Zoolander'} | 0.20299383

The salience for the entity Zoolander is very low, even though it is the most important entity in the sentence...

Is there any way to improve that? Or is there any other Natural Language API that can be customized?

Cooli
  • 169
  • 3
  • 14

2 Answers2

0

I am evaluating Google Cloud Natural Language and I can't find anything to customize the context... so it seems that is not possible :(

4lberto
  • 505
  • 6
  • 14
0

There is not a way to improve the API results since Google uses their own data to train the the machine learning models, and there are no options to customize it.

The only way to have a custom model is to train it yourself, as for example using AutoML. But this requires a lot of effort.

You could file a feature request to "condition" the results based on a context (or I could do it for you), which could affect the salience scores somehow, instead of taking into account just the word relevance on the text.

Anyway, this might not be necessary because I see an important salience difference between the first and second result. Have you tried to use the salience scores differences instead of the salience absolute value? Or a score multiplier for specific "types" of results (WORK_OF_ART in this case)?

I am aware that this solution might work for some kind of sentences but not for others.

Rubén C.
  • 1,098
  • 6
  • 16