# Rasa NLU without Rasa Core

**URL:** <https://forum.rasa.com/t/rasa-nlu-without-rasa-core/2322>\
**Category:** Getting Started with Rasa\
**Tags:** confidence\
**Created:** [October 30, 2018, 9:03pm UTC](https://forum.rasa.com/t/rasa-nlu-without-rasa-core/2322 "2018-10-30T21:03:00Z")\
**Posts on this page:** 5\
**Page:** 1

<div class="post-metadata">

**Author:** ![ipandzic](https://dub1.discourse-cdn.com/flex013/user_avatar/forum.rasa.com/ipandzic/32/715_2.png) [@ipandzic](https://forum.rasa.com/u/ipandzic)\
**Post date:** [October 30, 2018, 9:03pm UTC](https://forum.rasa.com/t/rasa-nlu-without-rasa-core/2322/1 "2018-10-30T21:03:00Z")

</div>

Hello, I am building a German language chatbot with 29 intents, mostly dealing with questions relating to an entrance exam on a university. Currently I am using only Rasa and I am getting pretty low confidence results sometimes 0.2, 0.3 or 0.4, rarely 0.5 or higher. The questions in the training data are sometimes similar for different intents. I have around ten sentences per intent for the training data. I am using spacy. Would I get better results if I used tensorflow? Also, is there a way of specifying keywords?

I read in the docs that tensorflow gives a higher confidence that spacy:

[https://rasa.com/docs/nlu/fallback/](https://rasa.com/docs/nlu/fallback/)

Does that mean that these approaches are not easily comparable?

---

<div class="post-metadata">

**Author:** ![labeebee](https://dub1.discourse-cdn.com/flex013/user_avatar/forum.rasa.com/labeebee/32/762_2.png) [@labeebee](https://forum.rasa.com/u/labeebee)\
**Post date:** [October 31, 2018, 8:43am UTC](https://forum.rasa.com/t/rasa-nlu-without-rasa-core/2322/2 "2018-10-31T08:43:53Z")

</div>

Hi, I read from the docs that tensorflow is only recommended if the dataset is of more than 1000 data points. So, if you are to enrich the dataset with more data, tensorflow for sure would give better result.

---

<div class="post-metadata">

**Author:** ![akelad](https://dub1.discourse-cdn.com/flex013/user_avatar/forum.rasa.com/akelad/32/69_2.png) [@akelad](https://forum.rasa.com/u/akelad)\
**Post date:** [November 2, 2018, 3:09pm UTC](https://forum.rasa.com/t/rasa-nlu-without-rasa-core/2322/3 "2018-11-02T15:09:06Z")

</div>

yes how much training data do you have for that number of intents? you’d probably need even more than 1000 at that point

---

<div class="post-metadata">

**Author:** ![ipandzic](https://dub1.discourse-cdn.com/flex013/user_avatar/forum.rasa.com/ipandzic/32/715_2.png) [@ipandzic](https://forum.rasa.com/u/ipandzic)\
**Post date:** [November 2, 2018, 8:02pm UTC](https://forum.rasa.com/t/rasa-nlu-without-rasa-core/2322/4 "2018-11-02T20:02:30Z")

</div>

Thanks guys! I’ve got 30 intents, and each with at least 10 training sentences, some with more, so around 350 examples all in all. Is a data point the same thing as a sentence/training example?

---

<div class="post-metadata">

**Author:** ![swainwindows](https://dub1.discourse-cdn.com/flex013/user_avatar/forum.rasa.com/swainwindows/32/3758_2.png) [@swainwindows](https://forum.rasa.com/u/swainwindows)\
**Post date:** [August 23, 2019, 10:26am UTC](https://forum.rasa.com/t/rasa-nlu-without-rasa-core/2322/5 "2019-08-23T10:26:35Z")

</div>

Hello Ivan,

Is this the right way to configure NLU Pipeline for german language in Config.YML file ?

language: “de”

pipeline:

- name: “spell\_check\_component.SpellCheck”
- name: “nlp\_spacy”
- name: “tokenizer\_whitespace”
- name: “ner\_crf”
- name: “ner\_synonyms”
- name: “intent\_featurizer\_count\_vectors”
- name: “intent\_classifier\_tensorflow\_embedding”