Explainable AI Whiteboard Technical Series: Natural Language Processing

0:10 today in the tech whiteboard series

0:12 we talk about nlp or natural language

0:15 processing

0:16 and how it impacts networking and more

0:18 specifically

0:19 ai ops natural language processing gives

0:23 machines the ability to derive meaning

0:25 from human language

0:26 nlp is a combination of linguistics and

0:28 ai specifically

0:30 machine learning right here is where nlp

0:33 lies

0:35 let's take a look at a question you

0:36 might ask marvis our virtual network

0:38 assistant

0:40 nlp converts this question into more

0:42 general meaning that our models know how

0:44 to interpret

0:45 to provide you with actionable insights

0:47 about your network needs

0:49 let's take a look at what's really

0:50 happening here the first step in nlp is

0:53 to clean up the text

0:54 and convert the words into a form the

0:56 computer can understand

0:58 first stop words or unimportant

1:01 information like and

1:02 and the are removed the remaining text

1:05 is then split into smaller units like

1:07 words and phrases

1:09 a process called tokenization next

1:12 featurization occurs meaning each word

1:14 is transformed into a vector

1:17 vectors numerically capture the features

1:19 or information about a word in a way

1:21 that the computer can understand and

1:23 process

1:24 here's an example of vectorized words

1:27 more semantically similar words fall

1:29 closer together

1:31 this is a crucial concept that allows

1:33 nlp to be possible

1:36 vector representations of words can

1:37 extend past 3d

1:39 higher dimensional vectors can

1:41 numerically capture more meaning about a

1:43 word

1:44 while each word is represented by a

1:46 vector we need to come up with an

1:47 encoded vector representation for the

1:49 overall sentence

1:51 sentence encoded vectors are valuable

1:54 because they allow information about the

1:55 order of the words to be captured

1:57 because words can have varying meanings

1:59 depending on their context

2:00 or position in a sentence at this point

2:03 in the process

2:05 embedding models are used embedding

2:07 models map

2:08 categorical data such as words or

2:10 sentences into high dimensional vectors

2:12 which capture semantic meaning about the

2:14 text

2:15 embedding models are usually pre-trained

2:17 on a large amount of data outside of

2:19 your own

2:20 like wikipedia which harnesses the power

2:22 of transfer learning

2:24 or leveraging prior knowledge from one

2:26 domain and task

2:27 into a different domain and task an

2:30 example of a pre-trained embedding model

2:32 is word to vec

2:33 which is trained on all the word data in

2:34 wikipedia meaning it's able to embed

2:37 extra meaning about the semantics of

2:39 text

2:39 into vectors because it's learned from

2:41 so many examples

2:42 what words can mean in certain contexts

2:45 the embeddings can now be fed into a

2:47 machine learning model

2:49 the machine learning model learns how to

2:50 understand the meaning of unseen words

2:53 by comparing

2:54 the similarity between the input word

2:56 vectors and the word vectors whose

2:57 meanings are known from your training

2:59 data

3:00 you can make sure that your model is

3:02 able to recognize certain meanings by

3:04 including them in your training data set

3:07 words that are semantically similar will

3:09 be closer in multi-dimensional space

3:11 which is how the model learns how to

3:13 predict the meaning of unseen words

3:15 the closest vector with a known meaning

3:16 in the vector space is the predicted

3:18 meaning

3:20 as a result of decades of

3:21 troubleshooting top-tier networks at

3:23 juniper

3:23 we've created a high-level structured

3:25 set of training data born from decades

3:27 of in-depth networking knowledge

3:30 we take real customer questions and

3:32 annotate them to create our training

3:34 data set

3:35 annotation includes flagging the tokens

3:37 in the question as intense

3:39 intended actions like troubleshoot count

3:42 list

3:43 or entities information about the intent

3:45 like device name or time frame

3:48 the training data questions are also

3:50 made into vectors and sentence encoded

3:52 vectors

3:53 with information about the annotation

3:54 flags

3:56 so when unseen questions are asked our

3:58 model can predict the user's desired

4:00 intents and entities

4:01 based on how similar the unseen vectors

4:03 are to the known vectors which have been

4:05 trained

4:07 the benefits of nlp are clear resolving

4:10 network issues in minutes

4:11 not days just by asking a single

4:13 question as opposed to poking around the

4:16 network looking for clues

4:17 and the versatility of using that same

4:19 interface to perform tasks

4:21 such as firmware upgrades allows marvis

4:24 to become a virtual member of the it

4:26 operations team

4:28 in networking we use nlp to allow our

4:30 customers to interface with marvis

4:32 pushing ai ops to the next level

4:35 we hope this episode helped uncover some

4:37 of the magic and mystery behind our

4:39 ai-driven network solutions