What Makes Computers Understand Human Language?

Human language is a rich combination of grammar, meaning, sentiment, and context—a puzzle even for humans at times. Yet, machines are perfecting this with surprising fluency, performing tasks like translation, summarization, sentiment analysis, and whatnot.

How does Siri know when you’re asking for important information and not cracking a dad joke?

How does Google Translate decode and reassemble languages so seamlessly?

The answer lies in Natural Language Processing (NLP), the invisible engine behind these impressive advancements. In tech, few achievements are as groundbreaking as teaching computers to understand our language. Through NLP, computers transform raw text strings into meaningful information, unpacking all the logic and challenges that power our daily interactions.

NLP Pipeline Overview

NLP pipelines include several phases, and the specific steps can vary based on the task at hand. Below is a brief overview of a standard pipeline, highlighting how each phase works in the background:

STEP 1: INPUT PREPROCESSING

The purpose of preprocessing input is to prepare the text so a computer can understand and process it more effectively.

Let's preprocess the sentence:

"This year, Gracie wants to be better at sports."

• Tokenization

  Break the sentence into individual tokens (words).

["This", "year,", "Gracie", "wants", "to", "be", "better", "at", "sports."]

• Lowercasing

  Convert all words to lowercase for uniformity so that words like "Sports" and "sports" are treated the same.

["this", "year,", "gracie", "wants", "to", "be", "better", "at", "sports."]

• Removing Stop Words
  Common words like "the," "is," “be,” “to,” and "in," which don’t carry much meaning, are often ignored.

["year,", "gracie", "wants", "better", "sports."]

• Stemming and Lemmatization

  Stemming: Chopping off word endings to reduce them to their base form.

  "wants" → "want"

  Lemmatization: Finds the dictionary form of a word based on its meaning.

  "better" → "good"

Final Output: ["year", "gracie", "want", "good", "sports"]

STEP 2: UNDERSTANDING STRUCTURE (SYNTAX ANALYSIS)

To understand sentences, machines analyze their structure first.

• Part-of-Speech (POS) Tagging

  Allocate grammatical tags to each word:

  "year" = noun

  "gracie" = proper noun

  "want" = verb

  "good" = adjective

  "sports" = noun
  

  Tagged Output:[("year", noun), ("gracie", proper noun), ("want", verb),
                   ("good", adjective), ("sports", noun)]

• Parsing

  Examine the relationships between the words:

  Subject: "Gracie" (proper noun)

  Verb: "want"

  Object/Complement: "good sports" (adjective + noun)

  Modifier: "year" (provides temporal context)

  
  The parsing step helps the computer understand the sentence structure as

  "Gracie wants something (good sports), and this relates to the context of the year."

STEP 3: EXTRACTING MEANING (SEMANTIC ANALYSIS)

Now, the computer tries to understand the meaning of the preprocessed input:

[("year", noun), ("gracie", proper noun), ("want", verb), ("good", adjective), ("sports", noun)].

• Named Entity Recognition (NER)

  The computer identifies specific entities within the text:

  "Gracie" = Person

  "year" = Temporal Entity (time-related context)

  "sports" = Activity

  Output: {"Gracie": Person, "year": Time, "sports": Activity}

• Word Embeddings

  Words are converted into numerical representations (vectors) that capture their meaning and relationships:

  "want" might be represented as a vector close to "desire" or "wish."

  "good" might be close to "excellent" or "great."

  "sports" will be close to terms like "games" or "athletics."

  These embeddings allow the computer to understand subtle relationships:
  e.g., "good sports" = "great athletics".

  Popular word embedding approaches: Word2Vec, CBOW, Skip-gram, GloVe, etc.
  

• Contextual Understanding

  Using context, the computer understands how words relate to each other:

  "Gracie" (the subject) is performing the action ("want").

  "good sports" (adjective + noun) is the object of the desire.

  "year" adds temporal context, indicating this is a goal for the current year.
  

  Conclusion: The computer understands the sentence as:
  "Gracie has a desire to excel in sports, and this relates to the current year."

  Common models for semantic analysis: BERT, GPT, and other transformer-based models.

STEP 4: HIGHER-LEVEL PROCESSING

After basic understanding, the computer can perform specific tasks on the sentence "This year, Gracie wants to be better at sports."

• Sentiment Analysis

  Determining the tone of the text (positive, negative, or neutral).

  Example:

  Sentence: "Gracie wants to be better at sports."

  Sentiment: Positive

• Text Summarization

  Creating a shorter version of the input text while retaining its meaning.

  Example:

  Input: "This year, Gracie wants to be better at sports."

  Summary: "Gracie aims to improve in sports."
  

• Translation (Machine Translation)

  Converting text from one language to another using algorithms.

  Example:

  Input: "This year, Gracie wants to be better at sports."

  Translation (English to Spanish): "Este año, Gracie quiere ser mejor en deportes."
  

• Question Answering

  Finding answers to specific questions from the given text.

  Example:

  Text: "This year, Gracie wants to be better at sports."

  Question: "What does Gracie want to improve?"

  Answer: "Gracie wants to improve at sports."
  

These higher-level tasks allow the computer to take actions such as summarizing, translating, or answering questions based on the given input.

Libraries and Models for Exploration

• Libraries:

  • NLTK, Spacy, TextBlob, Gensim, Hugging Face Transformers, CoreNLP, etc.

  • These libraries provide pre-built functions and pipelines for tasks like tokenization, POS tagging, and text classification.

• Models:

  • BERT, GPT, T5, Word2Vec, GloVe, FastText, ELMo, RoBERTa, etc.

  • These advanced models enable deeper contextual understanding, making them essential for tasks like semantic analysis, translation, and question-answering.

We now understand that NLP breaks down language into smaller chunks, analyzes their structure and meaning, and uses advanced models to interpret their context. Through such efforts, the communication gap between humans and machines is bridged, allowing them to understand and interpret what you say and what you mean.