Leveraging capabilities of NLP in manufacturing industries

1) Pre-processing of various types of data

It involves various steps as described below…

a) Removal of short structure words

• When an address is written, the street is represented as St.
• Months represented as Jan., Feb., and so on…
• Measurements represented as cm, mm, km, in.

Most of the framework splits the sentences based on the dot or period symbol. Words like {i.e., pg., e.g.} creates hindrance for the sentence splitting and tokenizing.

b) Removal of stop words

Stop words have extremely little value in helping a document. Examples are – it, didn’t, can, can’t, able, only, onto, or etc. Removal of such words will not create a huge impact on the application.

c) Removal of punctuations and special characters

Punctuation marks, Special characters, domain name behaves like a noise in the text. They do not play a vital role in the information.

d) Conversion of alphabet cases

Some of the programming languages and machine learning frameworks are case sensitive. The case of the text converted as either upper or lower. In general, lower case is preferred.

e) Tokenization

The text document parsed and chunked as smaller units like paragraphs, sentences, etc. The tokenization process chops the sentences or the text stream into pieces of words called as tokens.

f) Define a grammar rule / identify speech

In this process, words are assigned with proper tags (noun, verb, and preposition). This helps us to build the grammatical relationship between words. Few words are difficult to relate to the industrial context. For example, words like BREAK, CRUSH, SWING, SITE, and PLATFORM tagged with ‘NN’. N-Grams of text is a set of co-occurring words within a given word window size. The word window size could change as a unigram, bigram, and trigram and so on.

g) Word Embedding

The way humans understand words, machine cannot understand the same way. Word embeddings help convert the word/sentences from human understandable format to machine understandable format.

Using word embeddings, all individual words are represented as real-valued vectors in a predefined vector space. 

• Find TF IDF and Cosine Similarity

Term Frequency and Inverse Document Frequency (TF -IDF). TF-IDF is a factual result of Term Frequency and Inverse Document Frequency. Term Frequency is a raw count of a word in a Document and Inverse Document Frequency is a proportion of how much data completes a word give in a document (obtained by isolating the all outnumber of archives by the number of reports containing the term, and after that taking the logarithm of that remainder). TF-IDF is equivalent to {ratio of the event of a term in a document by the number of terms in a document} *log {ratio of the complete number of documents by the quantity of document that contains the term}

• GloVe / Word2Vec

Both uses Neural Network to come up with word embedding. Word2Vec learns embeddings by relating target words to their context. However, it ignores whether some context words appear more often than others. GloVe builds word embeddings in a way that a combination of word vectors relates directly to the probability of these words’ co-occurrence in the corpus.

h) Stemming and Lemmatization

Inflected words are words derived from other words. Inflected Language uses Inflected words in speech-language. The amount of deviation of the derived word from the root word is the degree of inflection. The degree of inflection may be lower to higher. In NLP, Stemming and Lemmatization are part of Text Normalization or Word normalization. A word in a selected document is present in different ways.

i) Grouping sentences based on similarity

A pair of selected sentences could be identical or non-identical. The structure of the sentence is identified by finding…

• First, the length of the sentences is calculated
• Next, the longest common substring and cosine similarity is calculated
• Finally, the similarity between the sentences based on a context is identified and grouped

Pre-processing steps improve the quality of the sentence and reduce the payload sent to the ML Algorithms.

2) Building lexical dictionaries

Every industry will have its own set of terms or keywords. For example, there can be a spare part called CUP and this is different from the general term ‘cup’. Same way, ‘Jaguar’ is a car name as well as an animal name. Consider sentences:

• Removed the front plate and applied the grease.
• Ensure refrigerant is not leaking after replacing the belt.

To extract the cause of the machine breakdown and to find remedy, simple cleansing and pre-processing will not be enough. The text related to the context of the industry is important. Three steps are required to relate the sentences with the context.

a) Build industry specific dictionary

Every industry will have some specific keywords. It is necessary to collect, build and train those keywords.

• In industries, words like chamber, boiler, unit, stage, and conveyor are common.
• In the IT sector, words like offshore development center (ODC), software requirements specification (SRS) are common.

b) Build a generic dictionary

Globally available lexical databases (like WordNet) are used to build a generic dictionary.

c) Build a derived dictionary

Finally, build derived keywords based on the Industrial keywords should be performed. Example: building concrete, pouring acid, adding lubricant. In addition to that, complex patterns like “front plate was broken due to too much of dust”.

Pre-processing steps and building a lexical dictionary improves the quality of content and reduces the error rate. Finally, cleansed text passed through the optimizer.

3) Optimize the data (or resultant data) with ML algorithms

Optimizer step applies multiple ML algorithms to achieve better & accurate results. Refer above diagram to know some of important ML algorithms. However, 6 important optimizer ML algorithms with descriptions are given in below table…