DataScience

Question 1

Mesh TensorFlow

Answer 1

Mesh TensorFlow (mtf) is a “language” for “distributed deep learning”

Supports:
Data-parallel training (batch-splitting)

The parameters of the model do not fit on one device - e.g. a 5-billion-parameter language model.

An example is so large that the activations do not fit on one device. - e.g. large 3D image model(experimental/unet.py).

Lower-latency parallel inference (at batch size 1).

https://github.com/tensorflow/mesh

Question 2

T5

Answer 2

Text to text transfer transformer

https://www.machinecurve.com/index.php/question/what-is-the-t5-transformer-and-how-does-it-work/

Question 3

Transfer learning

Answer 3

pre-training a model on “large unlabeled text data” with a “self-supervised task”,

After that, the model can be fine-tuned on smaller labeled datasets, often resulting in (far) better performance than training on the labeled data alone

such as language modeling or filling in missing words.

Question 4

General Language Understanding Evaluation (GLUE)

Answer 4

GLUE benchmark is a collection of nine natural language understanding tasks,

single-sentence tasks CoLA and SST-2,
similarity and paraphrasing tasks MRPC, STS-B and QP
natural language inference tasks MNLI, QNLI, RTE and WNLI.

Question 5

ROUGE-N

Answer 5

ROUGE-N measures the number of matching ‘n-grams’ between our model-generated text and a ‘reference’.

Question 6

Fine-tuning

Answer 6

“unfreezing” the pre trained model (or part of it), and “re-training” it on the new data with a “very low learning rate”.

Question 7

checkpoint

Answer 7

A checkpoint is an intermediate “dump of a model’s entire internal state (its weights, current learning rate, etc.)”

Question 8

natural language inference

Answer 8

Natural language inference is the task of “determining” whether a “hypothesis” is true (entailment), false (contradiction), or undetermined (neutral) given a “premise”.

Question 9

How to automatically compute weights?

Answer 9

Construct a simple neural network

network = (scores * weights) -> Activation function -> 0/1 -> grid search to update weights

Question 10

NER models

Answer 10

spacy
med7
deep pavlov
flair
allen nlp polycot
biobert

Question 11

extract important information from text

Answer 11

1. using structural integrity of the document (pdf, ms doc, speaker conversation)
2. identify entities and extract surrounding text
3. remove noise like greetings
4. remove tables and paragraphs with anchor points/ place holders

Question 12

Thins to research before implementing a pipeline

Answer 12

1. understand the code requirements (size of the libraries used, how much memory your code consumes while computation, difficult libraries to install, safety)
2. Run a demo on an instance before finalising the requirements
3. understand your platform (was, gcp, databricks, azure etc)
4. the data model should have proper updates
5. ALWAYS start of with a pipeline flow (process flow and dataflow)
   6.

Question 13

list of dict to pandas df

Answer 13

df = pd.DataFrame(dict)

Question 14

STT connection error while loading models

Answer 14

1. download the folder
2. set local path to model fodler

#Q&A
model = BertForQuestionAnswering.from_pretrained(r"C:\Users\sd22768\Downloads\bert_uncased_large") #'bert-large-uncased-whole-word-masking-finetuned-squad')

 #Extractive sumamriser
custom_config = AutoConfig.from_pretrained(r"C:\Users\sd22768\Downloads\bert_uncased_large")
custom_config.output_hidden_states=True
custom_tokenizer = AutoTokenizer.from_pretrained(r"C:\Users\sd22768\Downloads\bert_uncased_large")
custom_model = AutoModel.from_pretrained(r"C:\Users\sd22768\Downloads\bert_uncased_large", config=custom_config)

from summarizer import Summarizer

body = “text”
summ_model = Summarizer(custom_model=custom_model, custom_tokenizer=custom_tokenizer)
summ_model(body)

Question 15

filter dataframe rows based on column values

Answer 15

df.loc[df.col_name == condition]

Question 16

decode ascii froms tring

Answer 16

s.decode(“ascii”)

Question 17

apply function to pandas column

Answer 17

df[“col name”] = df[“col name].apply(lambda x : x.decode(“ascii”))

Question 18

question tagger

Answer 18

1. nltk nps_chat

2. stanza

Question 19

question tagger

Answer 19

1. nltk nps_chat (fast)

2. stanza (dead slow)

Question 20

keyword extraction

Answer 20

1. kBERT

Question 21

keyphrase extraction

Answer 21

1. kBERT

Question 22

replace string

Answer 22

string.replace(“?”, “word”)

Question 23

replace string at the index level

Answer 23

string.replace(“?”,”word”,2)

Question 24

syntax

Answer 24

structure of words in a sentence

Question 25

semantic

Answer 25

the meaning of a word, phrase, or text.

Question 26

Haystack

Answer 26

Open source framework for building search system

Question 27

call assessment -

Answer 27

phrase- filter with use, on filtered sentence apply semantic search
embedding - try new things like beeeert, use etc
keyword search - pos, expansion using word2vec,fastex, trained models and pruning using masked bert and score
multiple silimalirty scores - embedding score, aspect score, number score
final match

Question 28

TOPIC SEARCH

Answer 28

keyword search - exact keyword match, lemma and search, noun-noun/keyword pos pattern search
-> Identify duplicates

Question 29

keyword expansion

Answer 29

embedding search- fastex/use/bert and cosine similarity search
reduce computation byslecting nouns and other patters

Question 30

Zero-shot learning

Answer 30

Bart/Roberta -> NLI model for topic classfication without keywords

Question 31

Taxonomy

Answer 31

hierarchical ontology for med affairs

Question 32

sentiment search

Answer 32

BERT ABSA model - mind blowing performance

Question 33

emerging themes

Answer 33

kendal/slope eta/time series decomposition using stas library

Question 34

q&a

Answer 34

haystack pipeline based implementation