Machine Learning Artificial Intelligence Natural Language Processing Semantic Web Ontology Knowledge Information Processing Digital Transformation Probabilistic Generative Model Deep Learning Autonomous AI Navigation of this blog
Structuralism and meta-information behind the structure
Structuralism is a theory of philosophy and social science based on the idea that human thought and culture are formed by “structure,” and was developed mainly in linguistics, philosophy, anthropology, and literary theory in the early to mid 20th century.
Its main characteristics include the following
- Emphasis on the “relations” of things, looking not at individual elements but at how they are connected.
- Structure is universal, with common patterns across different cultures and eras.
- Meaning does not reside in stand-alone elements, but is determined by how those elements function within the system.
Key thinkers include the following figures.
- Ferdinand de Saussure – advocated “semiotics” in linguistics. He believed that language is based on the relationship between “signifier” and “signified”.
- Claude Lévi-Strauss – Applied structuralism to anthropology, analyzing the “structure” of myths and social institutions.
- Roland Barthes – Advances the analysis of symbols in literature and culture, examining the “structure of narrative” and the “operation of meaning.
This structuralism analyzes not only the surface “structure” but also the “meta-information” behind it.
Meta-information refers to “information about information” rather than information itself. This is, for example, as follows.
- Linguistics perspective: not the words themselves, but the rules and patterns that govern the relationships between words.
- Cultural anthropological perspective: not the myths themselves, but how the myths are constructed and linked to the values of the society.
- From the viewpoint of information science: not the data itself, but how the data is classified, organized, retrieved, and given meaning.
Examples of meta-information include the following
- Meta information in language
- Grammatical rules: rules for how words and sentences are organized.
- Semantic networks: how words are related and form meaning.
- Meta-information in mythology and culture
- Lévi-Straussian “dichotomy”: underlying culture are oppositional structures such as “life/death,” “nature/culture,” and “man/woman.
- Semiotic analysis: For example, “white dress” is not just a piece of clothing, but has social meanings such as “purity” and “marriage.
- Meta-information in digital technology and AI
- Metadata: attributes of data (e.g., file name, creation date, tags, etc.)
- Knowledge graph: the connections between concepts behind the data.
- AI learning models: not just data, but algorithms that determine how that data is classified and interpreted.
Structuralism analyzes the world on the assumption that there is a universal structure, but since the 1960s, post-structuralism (Post-structuralism) has emerged.
The claims of this post-structuralism have been as follows,
- Structure is not fixed but fluid.
- Meaning changes with context and is not absolute.
- The “meta-information” itself is also structure-dependent, so that no one interpretation is fixed.
Representative thinkers include Jacques Derrida (who advocates “deconstruction” and deconstructs the structure of language and text) and Michel Foucault (who analyzes the relationship between power and knowledge and explores how the structure of knowledge changes).
In contemporary society, structuralist perspectives are also being applied to AI and information science.
- Search engine algorithms → search results are determined by relevant “meta information” and “context” rather than by keywords alone.
- SNS information recommendation → Analyzes user behavior patterns and optimizes feeds based on “data structure” rather than individual posts.
- Metaverse Web 3.0 → Identity and data in the digital space are given meaning through “meta-information.
Utilizing the perspectives of structuralism and meta-information, it is expected that insights into the “invisible structure” behind information flows and data will deepen and provide clues for the future of digital society.
Related AI Technologies
The “relationships” and “systematic structures” emphasized by structuralism are also important concepts in AI technologies, especially in the fields of natural language processing (NLP), knowledge graphs, data mining, and generative AI, where technologies for handling the structures themselves and the “meta-information” that lies behind them are developing. They are described below.
1. structuralist approach and AI technologies
(1) Language models and semiotics
Related technologies (Natural Language Processing (NLP), Transformer (GPT, BERT, etc.)): In structuralism, the meaning of language is considered to be determined by its relationship with context and structure, not by words alone. This view is reflected in the process of learning language models.
- Saussure’s semiotics: the relationship between “signifiers (sounds and letters) and signifiers (meanings)” determines the meaning of language.
- A similar concept in AI: in NLP, the meaning of a word is determined by the surrounding context through a vector representation of the word (Word2Vec, FastText, BERT).
- Transformer model: GPT (including ChatGPT) generates sentences while considering the “relationship” between words.
Specific examples
- Word2Vec / BERT → Understands the meaning of words in context and generates appropriate responses.
- ChatGPT → Learns patterns in language and builds a network of meanings.
- Structural Text Analysis → Analyzes the structure of sentences (context, relationships) to classify and organize content.
(2) Knowledge Graph and Relational Structure
Related technologies (Knowledge Graph, Graph Database (Neo4j, RDF)): Claude Lévi-Strauss viewed myths and social structures as “binaries” or “networks of relations. This idea has been carried over to the Knowledge Graph in the field of AI.
- What is a Knowledge Graph?
- It organizes data as “nodes (concepts)” and “edges (relationships)” to create a semantic network.
- Widely used in Google and Bing search engines, corporate data management, and medical and pharmaceutical research.
- Graph database (Neo4j, RDF)
- A database that explicitly describes relationships such as “A is a kind of B” or “C is related to D.”
- It contributes to organizing structural data and improving search accuracy.
Specific examples
- Google Knowledge Graph → Search engines understand “information connections” and provide related information.
- Medical care and drug discovery → Visualize the relationship between diseases, genes, and drugs with a knowledge graph.
- Corporate data management → Organize internal information (technical documents, specifications, development history) as a knowledge graph.
(2) AI technology for handling meta-information
(3) Metadata analysis and information recommendation
Related technologies (metadata processing, recommendation engines): Meta information is “information about information. This can be, for example, “video title,” “creator,” “tags,” “genre,” etc.
- Automatic metadata extraction (AI + NLP)
- AI extracts and classifies meta information from text and images.
- Examples: automatic tagging of YouTube, search engine page rank analysis.
- Recommendation Engine (Collaborative Filtering)
- Netflix and Amazon recommendation systems analyze user behavior and suggest “relevant content”.
Specific examples
- Spotify / Netflix recommendation algorithm → Recommends content that should be played next based on the user’s viewing history.
- Academic article metadata analysis (Google Scholar) → Analyzes article titles and citation relationships, and recommends related articles.
(4) Image/video metadata analysis
Related technologies (image recognition (CNN), computer vision (YOLO, OpenCV)): From a structuralist perspective, video and images can be viewed not in terms of “individual pixel information” but in terms of “meaningful structures and relationships.
- Structural analysis in images (CNN, OpenAI CLIP)
- Object recognition (e.g., “this is a car”, “this is a cat”).
- AI learns elemental relationships (spatial arrangement) in images.
- Video analysis and scene understanding (YOLO, Mask R-CNN)
- Analyzes relationships between objects and characters in videos.
Specific examples
- Automatic caption generation for YouTube → Converts video audio into text and stores it as meta-information.
- AI analysis of surveillance cameras → Analyzes the movement of people and detects abnormal behavior.
- Automatic driving (Tesla, Waymo) → Real-time analysis of meta-information on road conditions.
(5) Deconstruction by AI (post-structuralist approach)
Related technologies (Generative AI (GAN, Diffusion Models), Self-Organizing Maps (SOM)): In the poststructuralist perspective, “meaning is fluid, not fixed. This idea is associated with creative generative technologies of AI.
- Generative AI (StyleGAN, DALL-E, Midjourney)
- A technology that generates new images and text “without being bound by existing structures.
- It goes beyond fixed semantic frameworks to generate diverse interpretations and new patterns.
- Self-organizing maps (SOM)
- A technique to classify and visualize data without being bound by “prior categories.
- Example: Organizing unclassified news articles by topic.
Specific examples
- AI-based poetry generation (GPT-4, Bard) → Learning the structure of existing poems and creating new ones.
- Digital identity generation in Web 3.0 → Generate individual “digital personalities” from user activity data.
Applying the structuralism perspective to AI technologies, we can see that technologies for analyzing the relationship between “structure and meta-information” and creating new knowledge are developing in many fields, including language, knowledge, video, recommendation systems, and generative AI. Adding the post-structuralist perspective, we can also see a future in which AI “creates new meaning.
Mounting example
In this section, we describe examples of implementation of knowledge graphs, metadata analysis, and generative AI based on the structuralist approach.
1. Knowledge Graph Construction (Python + NetworkX + Neo4j)
Structuralism believes that meaning is created from “relationships” rather than single elements. Using a knowledge graph, it is possible to visualize the relationships among concepts.
Simple Implementation of Knowledge Graph
import networkx as nx
import matplotlib.pyplot as plt
# Creating Graphs
G = nx.DiGraph()
# Add a node (concept)
G.add_nodes_from([“language”, ‘symbols’, ‘structure’, ‘culture’, ”knowledge”])
# Add relationships (edges)
edges = [(“Language”, ‘Symbol’.), (“Language”, ”Structure”), (“Structure”, ‘Culture’.), (“Culture”, ”Knowledge”)]
G.add_edges_from(edges)
# drawing
plt.figure(figsize=(5, 5))
nx.draw(G, with_labels=True, node_color='lightblue', edge_color='gray', node_size=2000, font_size=12)
plt.show()Result: Visualization of the relationship between language, symbols, and culture as a graph
Knowledge graph construction using Neo4j
from neo4j import GraphDatabase
# Neo4j connection
uri = "bolt://localhost:7687"
username = "neo4j"
password = "password"
driver = GraphDatabase.driver(uri, auth=(username, password))
def create_knowledge_graph(tx):
query = """
MERGE (lang:Concept {name: 'Language'})
MERGE (sign:Concept {name: 'symbol'})
MERGE (struct:Concept {name: 'structure'})
MERGE (culture:Concept {name: 'culture'})
MERGE (knowledge:Concept {name: 'knowledge'})
MERGE (lang)-[:RELATES_TO]->(sign)
MERGE (lang)-[:RELATES_TO]->(struct)
MERGE (struct)-[:RELATES_TO]->(culture)
MERGE (culture)-[:RELATES_TO]->(knowledge)
"""
tx.run(query)
with driver.session() as session:
session.write_transaction(create_knowledge_graph)
driver.close()Result: Constructed a knowledge graph in Neo4j and managed the relationships between concepts in a database.
2. meta-information analysis (NLP + metadata extraction)
Meta information is “information about information” and can be automatically extracted from text using NLP.
Extract meta information from article titles
import spacy
nlp = spacy.load("en_core_web_sm")
text = "A Structural Approach to Deep Learning in Knowledge Representation"
doc = nlp(text)
# Extracting Meta Information
entities = [(ent.text, ent.label_) for ent in doc.ents]
keywords = [token.text for token in doc if token.is_alpha and not token.is_stop]
print("Meta information (unique expression):", entities)
print("Metadata (keywords):", keywords)Results:
Meta information (proper nouns): [] # (no proper nouns in this example)
Metadata (keywords): ['Structural', 'Approach', 'Deep', 'Learning', 'Knowledge', 'Representation']Applications: – Tagging of academic articles (inferring research fields from article titles)
- Tagging of academic papers (inferring research fields from paper titles)
- Classification of corporate documents (classification of design documents, technical documents, etc.)
3. structure transformation using generative AI (GPT-4 + DALL-E)
From a post-structuralist perspective, it is important to “disrupt existing structures and generate new meanings”.
Text generation using GPT
import openai
openai.api_key = "your_api_key"
prompt = "Write a poem about the relationship between AI and humans in light of Levi-Strauss' structuralism."
response = openai.ChatCompletion.create(
model="gpt-4",
messages=[{"role": "user", "content": prompt}]
)
print(response["choices"][0]["message"]["content"])Result: GPT produces “Structuralism-Based Poetry”.
Application:
- Automatic generation of philosophical texts
- Narrative creation with structural changes
4. image/video metadata analysis (OpenCV + YOLO)
Extract “structural meaning” from video images and treat it as meta-information.
Detect objects from images and extract metadata
import cv2
import torch
# Load YOLOv5 model
model = torch.hub.load('ultralytics/yolov5', 'yolov5s')
# Load and analyze images
image_path = "scene.jpg"
results = model(image_path)
# Display detection results
results.show()Result: Analyze object relationships from images and output metadata such as “where the person is”, “what is in the background”, etc.
Applications:
- Anomaly detection in surveillance cameras (e.g., behavior pattern analysis)
- Semantic analysis of movie scenes (emotion and situation classification for each scene)
reference book
Discuss structuralism, meta-information, and references for applying it to AI technology.
Foundations of Structuralism
Basic literature for understanding the concept of structuralism.
Claude Lévi-Strauss,(La Pensée Sauvage, 1962)
Analyzes the “structure” of symbols and culture to elucidate myths and social patterns.
Ferdinand de Saussure,(Cours de linguistique générale, 1916)
Considers language as a “system of signs” and defines the relationship between signifiers and signifieds.
Roland Barthes, Analyse structurale des récits (1966)
Analyzes the universal structure behind literary works and links semiotics and narratology.
The Order of Things: An Archaeology of the Human Sciences (Les Mots et les Choses, 1966)
Proposes the concept of “episteme” (framework of knowledge) and analyzes the structure of knowledge in different periods.
Jacques Derrida, Of Grammatology (De la grammatologie, 1967)
Criticized structuralism and proposed the concept of “deconstruction. Shows the fluidity of meaning and the polysemy of meta-information.
Technologies related to meta-information and knowledge graphs
Technical books on meta-information, knowledge graphs, and semantic AI.
Building Knowledge Graphs: A Practitioner’s Guide
Christopher Manning, Hinrich Schütze, and Prabhakar Raghavan, Introduction to Information Retrieval (2008)
Explains the basic theory of information retrieval and the importance of metadata. A must-read as a basic theory of search engines.
George Lakoff, Women, Fire, and Dangerous Things: What Categories Reveal About the Mind (1987)
Analyzes how the concept of categories is formed from the perspective of cognitive science. Useful for knowledge graph construction.
“Knowledge Graphs and Semantic Web”
François Chollet, Deep Learning with Python (2nd Edition, 2021)
Describes how to implement knowledge representation and information processing using deep learning.
Technologies for Generative AI and Structural Transformation
Literature on techniques for creating new meaning using generative AI.
Neural Networks for Signal Processing
Sebastian Raschka and Yuxi Liu, “Machine Learning with PyTorch and Scikit-Learn” (2022)
Provides a wealth of examples of natural language processing (NLP) implementations using Python.
Noam Chomsky, “Syntactic Structures” (1957)
Proposed the concept of generative grammar and provided the theory underlying AI’s language understanding model.
The Principles of Deep Learning Theory
Applications: video analysis and knowledge integration
A technical book on video, image, and knowledge integration as an application of meta-information analysis.
Richard Szeliski, Computer Vision: Algorithms and Applications (2021).
Describes methods for image analysis and metadata extraction.
Joseph Redmon, YOLO: Real-Time Object Detection (2016).
Describes real-time object detection techniques using YOLO (You Only Look Once).
Ian Goodfellow, Yoshua Bengio, Aaron Courville, “Deep Learning” (2016)
A basic book on AI-based understanding of information structures, knowledge representation, and meaning generation.
コメント