Language Detection

As Diana is from Canada, she needs to be bilingual so it is important that when asked a question or told something that the language is detected.  If the text comes from Twitter then the tweet will contain the language, however if it comes from another source say Slack or text on a console, that language is not provided.  This is where language detection providers are so important.

POC #1: Detection Via LanguageDetection Nuget Package

Initially Dain and Diana used the LanguageDetection Nuget:

Install-Package LanguageDetection

This package is described as "Detect the language of a text using naive a Bayesian filter with generated language profiles from Wikipedia abstract xml, 99% over precision for 53 languages. Original author: Nakatani Shuyo."

POC #2: Cognitive Services

After some growth and chatting with Mark Stiles a second provider was created for Cognitive Services based on this article Quickstart: Identify language from text with the Translator Text REST API (C#).

There are many other ways to detect language. DAIN and DIANA use a provider model so you have the option of using different providers. You may also use multiple providers weighting the reliability. You can then apply that weighting to determine which language is most probable.  This is important especially if you are given a single statement as some languages are too close to determine.

For example Portuguese and Spanish may be detected as the same if you simply read one response of Si however if you look at previous responses the Spanish or Portuguese ones would be quite different. That is why the language the language provider would be passed what we think the prev.

So that is the other improvement that is necessary. You need to look at previous statements. The problem is if someone changes languages mid conversation. For example if you were reading a speech given in parliament then it is possible that one sentence or paragraph is in English then the same paragraph repeated in french.  The same could be said if the politician is tweeting. They may tweet in English then follow it up with the same in french.

The language context is of importance. You need to use context of prior statements but need to understand that just because the previous statement is in one language does not mean the next one is in the same language. Just some things to ponder.

Language becomes more important as you determine how you will do Natural Language Processing. There are 3 approaches when dealing with language:

  1. Detect Language, Translation and then Natural Language Processing. Although the simpler way to do it and actually the way DAIN and DIANA worked for the POC, you do lose some of the local nuances the native language contains. You may also suffer some performance loss as you have to translate both the request and the responses back.
  2. Detect Language and then Language Specific NLP and Language Specific processing of response. This is more complex so it will take more work however you do gain a more fluid conversation. The other disadvantage is maintaining multiple NLP and as the volume of data grows it will be multiplied by number of langauges.
  3. Detect Language and then Language Specific NLP and then shared processing of response.

In the case of DAIN and DIANA, we use approach 3 because our Reputation Engine results are stored in English however any links to resources do have a language flag so that factor can be used to determine the best answer for the asker.

For your application you may find approach 1 or 2 is better. The base DAIN does allow you to downgrade to shared NLP, it also allows you to have Language Specific processing rather than a shared one.

If you are interested in language detection and adding that to your applications reach out and let's discuss. I will be providing a basic QuickStart to SitecoreDain subscribers soon. If you cannot wait email me at chris.williams@readwatchcreate.com and as a subscriber, I can release an early version to get you started now.

Comments

Post a Comment

Popular posts from this blog

Natural Language Processing

Another key concept to DAIN and DIANA is Natural Language Processing. There are a variety of libraries and algorithms that can convert a sentence or series of statements into their various figures of speech.  DAIN and DIANA use the provider model the same way we did for Language Detection. This allows you to swap out your NLP as well as use multiple with the same idea of providing weighting to determine which provider would be considered more important. Weighting can be a simple number weighting or could include simple or more complex rules like this provider works better for English than it does for french or if this sentence contains these keywords then use this provider over that one. Natural Language Processing Steps There are various phases in Natural Language Processing. This article explains the basic ones.  Some libraries will expose the various steps while others will group them together or have a single method to do all of them: Splitting: This is the proces...
Read more

Providers and Pipeline, Oh Why!!!

I have mentioned to a few people that DAIN and DIANA are built with providers and pipelines. The pipelines. Some providers are weighted while others run conditionally based on rules.  The question I get is often is why not just choose an algorithm and stick with it for simplicity.  There are a few reasons for that: There are 3 reasons to have different providers the first is Bias and the second is Context. Bias In the case of Natural Language Processing, you have heard a lot in the news about bias in AI and ML. Some of that is due to data which is the subject of another article but other times it is that the algorithm has been trained with data that is not specific to your domain and thus can produce bias that is not specific to your data. That is why it is so important to to use multiple algorithms to keep the other in check. You can weight one algorithm higher so it is most likely to produce the result but there are cases where it differs greatly from other algorithms t...
Read more

The Universal Content Processing Engine

I am starting to learn more about Digital Asset Management and Content Processing and automatically processing content to gather Metadata both structured and unstructured.  Through this I have come up with a new self-learning engine called the Universal Content Processing Engine. Whereas currently I take advantage of statically programmed logic on how to process a tweet, make calls to given API based on static pipelines, I am limited to how I am coded to a certain extent although some learning has been happening.  Through the use of UCP Engine, I will start with a core set of knowledge and will build upon that.  I will then ask my maker for more information on how to process different types of data. Sometimes it would be a configuration of an existing component and other times it will be another provider dynamically added to my system. To start I will process content from 3 sources: FILE, STRING, URL. Through supervised learning I will learn about more sources and ...
Read more