Indian Knowledge System Unit 5 Notes – AKTU B.Tech 1st Year

Indian Knowledge System Unit 5 Notes cover Contemporary Relevance, AI and Applications of IKS. The unit examines how knowledge is preserved, represented, searched, analysed and responsibly adapted.

These AAS104/AAS204 notes apply to AKTU B.Tech 1st Year, Semester I & II, common to CSE, EE, ECE, CE, ME, BT, CH, TX and AG streams.

Unit Overview

The unit connects manuscripts and living traditions with digital preservation, intellectual property, Sanskrit language processing, AI evaluation and sustainable engineering.

A polished scan, fluent answer or traditional-looking product is insufficient by itself. A trustworthy application needs provenance, permission, interpretation and measured performance.

The Sanskrit-corpus case is deliberately tiny and constructed. Its dictionary-lookup procedure demonstrates reproducible evaluation, not a trained AI model or research benchmark.

Topics Covered

Learning Outcomes

• Distinguish digitisation and continuing digital preservation.
• Explain metadata, versioning, checksums and recovery tests.
• Distinguish TKDL, the IKS Division and patent authorities.
• Separate ownership, access, permission and intellectual property.
• Explain Panini’s computational relevance without equating grammar with AI.
• Describe OCR, normalisation, sandhi splitting, morphology, parsing and retrieval.
• Calculate error rates, precision, recall, F1 and coverage.
• Identify hallucination, leakage, domain shift and review needs.
• Develop a testable IKS-inspired engineering proposal.

Essential Vocabulary

Core question: Can a reader trace the source, check permission, reproduce the result and see its limits?

5.1 Contemporary Applications: A Trustworthy Chain

Four connected activities need distinction:

A high-resolution image may be unsearchable. Searchable text may contain errors. Correct words may be misinterpreted. A sound interpretation may still fail in an unsuitable engineering application.

Define a Manageable Question

“Build AI for all IKS” is too broad.

A reviewable first task is: Find a selected concept in a permission-cleared corpus and show the supporting passage.

Define success before implementation. Archive completeness, retrieval usefulness and cooling performance require different checks.

Thousands of attractive thumbnails without folio order, identifiers or rights information may be less useful than a smaller documented collection.

5.2 Digitisation of Manuscripts

Digitisation creates digital representations of physical material.

Preserve identifiers, folio order, visible condition and the relationship between image and object.

Fragile material should not be forced flat for appearance. Digital capture does not replace care of the original.

Capture Record

Record:

• Object identifier and photographed side.
• Capture date and operator.
• Relevant settings and processing.
• Sequence and orientation.
• Focus, readable detail and completeness.

An image alone does not establish authorship, date, authenticity, ownership or publication permission.

Keep the master capture distinct from access copies and enhanced views. Contrast adjustment may help reading, but should not silently replace original evidence.

5.3 Metadata and Linked Resources

Metadata should answer: What is this? Where did it come from? How do its parts relate? What use is permitted?

Rights and access details provide further administrative context.

Distinguish a cataloguer-supplied title from one present in the object. Preserve uncertainty about dates and attributions.

Constructed Record

• Identifier: NG-IKS-DEMO-001.
• Type: Teaching example.
• Language: Sanskrit.
• Representation: Roman transliteration.
• Source status: Constructed sentence, not a manuscript quotation.
• Version: 1.0.
• Annotation convention: Selected noun stems and verb roots.
• Access: Educational demonstration.

Controlled values improve search consistency. “Sanskrit”, “Skt” and unexplained abbreviations should not become disconnected categories without documented mapping.

Validate identifiers, expected uniqueness, sequence and file links. Structural validation does not establish that a disputed attribution is historically correct.

5.4 Digital Preservation Over Time

Digitisation is a capture event. Preservation continues through equipment failure, accidental changes and format obsolescence.

Checksums and Fixity

A checksum is computed from file bytes. Comparing it with a trusted earlier value can detect change.

• A mismatch calls for investigation.
• A match supports byte consistency with the reference.
• A match does not prove historical authenticity or correct initial capture.

A backup recovers bytes; an archive also preserves identity, context, permissions and relationships.

During migration, identify the original and new versions, record tools and settings, and check appearance and content.

Test restoration into a separate location. Recover images with their metadata, compare checksums and open representative files.

Permission to view a preview does not automatically permit republication of a master.

5.5 CSIR-TKDL and Defensive Protection

The Traditional Knowledge Digital Library, associated with CSIR, makes documented traditional medicinal knowledge more usable for patent examination through structured representation, classification and language support.

Its central purpose is defensive protection: helping prevent inappropriate claims over existing knowledge.

It does not grant a new patent for every documented practice.

Prior Art

Prior art is earlier information relevant to assessing patentability. Documentation can reveal knowledge that an examiner might otherwise miss because of language or format barriers.

A familiar ingredient name does not alone settle every invention involving it. A claim must be compared with the actual earlier disclosure and applicable requirements.

5.6 IKS Division and Institutional Roles

The Indian Knowledge Systems Division of the Ministry of Education has a broader educational and research role involving interdisciplinary engagement and dissemination.

Institutional association does not automatically establish the correctness of every output.

Example: Water-Management Exhibition

Community members explain use and upkeep. Historians interpret records. Engineers assess the physical system. The digital team captures permitted material. The institution reviews access and attribution.

A proposal should state problem, sources, partners, permissions, method, output and evaluation.

5.7 IPR: Ownership, Expression and Invention

Intellectual Property Rights concern legally recognised interests in particular creations or inventions. Traditional knowledge may be collective and transmitted across generations, complicating simple individual-authorship assumptions.

Physical possession, permission to photograph and permission to publish are different matters.

Traditional Knowledge and Patents

The source discusses Section 3(p) of India’s Patents Act, concerning inventions that are in effect traditional knowledge or merely aggregate or duplicate known properties of traditionally known components.

The syllabus lesson is that documenting existing knowledge does not make it new. A new label or vague “AI-powered” description does not itself establish patentability.

Copyright and Permission

Copyright concerns protected expression rather than general ownership of an idea. An ancient underlying text and a modern original translation, commentary or image can raise different rights questions.

• Defensive protection: Prevent inappropriate rights over existing knowledge.
• Positive protection: Support authorised control or benefit through relevant mechanisms.

These are academic distinctions, not a legal determination for a particular collection.

5.8 Consent, Credit and Responsible Access

Digitisation can make locally shared knowledge globally searchable. Discuss exposure and intended use before capture.

Clarify recording, access, publication, commercial reuse, attribution and restrictions. One person may not have authority over collectively held or restricted knowledge.

Credit does not replace consent. Some contributors may prefer anonymity.

Human-browsing permission does not automatically include model training or redistribution.

Example: A craftsperson excludes a sensitive step from an interview. Publish only approved content. Do not infer the missing step with AI and attribute the invention to the contributor.

Preserve variation and disagreement rather than forcing artificial consensus.

5.9 Case Study: A Small Digital Collection

A college receives permission to document 20 folios, each with two sides.

Expected side images = 20 × 2 = 40.

The folder has 40 files, but 07b is missing and 08a duplicated. Correct file count does not establish completeness.

Re-capture 07b if authorised and feasible. Mark the duplicate and record the correction. Do not renumber unrelated files to hide the gap.

Storage Estimate

At an assumed average 50 MB per master:

40 × 50 = 2,000 MB = 2 GB, using decimal units.

Three copies require approximately 6 GB of image bytes, plus metadata, derivatives, logs and growth.

GB and GiB are different units. This is a capacity exercise, not a capture-setting recommendation.

Test recovery of an image and metadata. Report completed checks separately from planned ones.

5.10 Sanskrit Representation and Normalisation

Sanskrit is a language; Devanagari is a script. Sanskrit can appear in other scripts and Roman transliteration.

Diacritics preserve meaningful distinctions. Removing them can merge different forms.

Sanskrit’s inflection means one exact-form search can miss related forms. Sandhi and compounds also complicate whitespace tokenisation.

Unicode can represent an equivalent visible character in different ways. NFC normalisation can standardise equivalent encodings; it is not spelling correction or permission to discard linguistic distinctions.

Retain a faithful display form even if offering a simplified search index.

5.11 Panini and the Ashtadhyayi

Panini’s Ashtadhyayi is arranged in eight chapters and uses concise rules, technical vocabulary and grammatical conditions.

It belongs to a wider tradition of predecessors, commentary and interpretation. Compact rules may depend on context elsewhere in the system.

The analogy does not make the work a modern programming language, neural network or complete NLP model.

Implementations must define representation, applicability, interactions and validation. No language becomes universally ideal for AI because it has sophisticated grammar.

5.12 From Grammatical Rules to a Program

For a selected teaching example:

deva + indra → devendra, with a + i → e at the identified boundary.

The components and boundary are supplied. This does not demonstrate unrestricted automatic splitting.

Generation Versus Analysis

• Generation: Known components produce a surface form.
• Analysis: A surface form generates possible underlying components.

Reverse analysis can have several candidates.

A rule must distinguish word boundaries, suffix boundaries and arbitrary internal positions.

Test positive cases and negative cases. Replacing every a with e could pass one poorly chosen example while failing the actual task.

State scope instead of calling one implemented rule a complete grammar engine.

5.13 NLP Pipeline for Ancient Texts

Natural Language Processing includes recognition, segmentation, grammatical analysis, retrieval and meaning-oriented outputs.

Retain raw captures, OCR text, corrections and annotations as linked versions.

An OCR vowel error may prevent lexical matching, cause a wrong split and produce an incorrect parse. A fluent summary can conceal that chain.

Stages may interact: morphology can reject implausible splits, while context can help choose OCR alternatives.

Expert and custodial review belong in system design, including decisions about whether material may be processed.

5.14 OCR and Character Error Rate

Optical Character Recognition converts visible text into provisional machine-readable characters.

Blur, skew, bleed-through, fading and cropped margins create different errors. Clean-print performance does not guarantee manuscript performance.

Keep original captures and document enhancement. Excessive cleaning can erase faint strokes.

Mark uncertain readings rather than replacing gaps with confident guesses. Model confidence is not proof of correctness.

CER Calculation

CER = (substitutions + deletions + insertions) / reference character count.

For 100 reference characters, 3 substitutions, 1 deletion and 1 insertion:

CER = 5/100 = 5%.

Document character-counting conventions, especially combining marks and spaces.

A 5% CER does not mean every sentence is 95% correct in meaning. One changed name or numeral can have a large interpretive effect.

5.15 Sandhi Splitting and Segmentation

Sound changes at boundaries and variable spacing make space-only tokenisation insufficient for many Sanskrit texts.

Candidate-Checking Sequence

1. Propose boundaries.
2. Reverse possible sound changes.
3. Check lexical entries.
4. Examine grammatical compatibility.
5. Review sentence context.

Dictionary membership alone does not guarantee a plausible sentence analysis.

Retain original surface forms, alternatives and selected decisions.

Score a defined unit: complete segmentation, individual boundaries or recovered components. High boundary accuracy need not imply high full-sentence accuracy.

Classify errors as lexical gaps, wrong sound reversal, boundary errors, encoding mismatch or contextual ambiguity.

5.16 Morphology and Lemmatisation

Morphology identifies word structure and grammatical features. Lemmatisation maps forms to a documented citation or base form.

Lemma-based search can retrieve related forms, provided their analyses are correct.

One form may allow several analyses. Annotation disagreements can arise from readings, conventions, genuine ambiguity or errors.

Removing endings mechanically is insufficient. In the later case, pibati maps to pā under the supplied verb-root convention; this is not derived by simple suffix stripping.

5.17 Syntax and Dependency Analysis

Syntax concerns relationships among words. Rich morphology can help identify relations despite variable order, but context remains important.

The constructed sentence means “Rama eats a fruit.”

Grammatical case and semantic role are different. Voice, construction and lexical behaviour affect their relationship.

Display original tokens, identifiers, selected analyses, links and alternatives. Modern annotation schemes must explain how they interpret classical grammatical concepts.

5.18 Rule-Based, Learned and Hybrid Systems

A hybrid system might generate grammatical candidates using rules, then rank them contextually with a learned model.

Begin with a baseline. Complexity is useful only when it improves a meaningful result under fair comparison.

Data Partitions

Training fits or builds the method. Development selects settings and examines errors. Test evaluates the frozen procedure.

Data leakage includes near-duplicate passages across partitions, tuning on test answers and building a lookup table from test labels.

A random sentence split tests something different from holding out a work, author or handwriting style.

Domain shift occurs when deployment material differs—for example, clean prose versus damaged verse manuscripts.

5.19 Precision, Recall, F1 and Coverage

Suppose 15 predicted concept mentions contain 12 correct ones, while the reference contains 16 relevant mentions.

TP = 12; FP = 3; FN = 4.

Precision = TP/(TP + FP) = 12/15 = 80%.

Recall = TP/(TP + FN) = 12/16 = 75%.

F1 = 2TP/(2TP + FP + FN) = 24/31 ≈ 77.4%.

F1 is the harmonic mean, not the ordinary arithmetic average.

Match Metrics to Tasks

• OCR: CER or WER.
• Lemma assignment: exact-match accuracy and coverage.
• Mention extraction: precision, recall and F1.
• Parsing: correct relations under a stated scheme.
• Retrieval: relevant results and missed evidence.

WER

For a 20-word reference with 2 substitutions, 1 deletion and 1 insertion:

WER = 4/20 = 20%.

CER and WER count different units.

Coverage is the fraction answered. High accuracy among answered items can hide many abstentions.

Report sample size and relevant breakdowns. Attractive decimals do not establish statistical significance on tiny datasets.

5.20 AI Outputs Need Verifiable Evidence

A fluent answer can invent a reference, fill a damaged passage or merge different traditions.

Keep source text, proposed reading, translation and commentary distinct.

A real citation may still fail to support the claim. Read the relevant passage rather than merely checking that a title exists.

An unreadable plant name must not be silently replaced with a familiar plant and then converted into a treatment claim. Mark the gap and seek linguistic review.

Access rules must apply to search and generation. A restricted passage can be disclosed through an answer even if the original file is not downloadable.

5.21 IKS-Inspired Sustainable Engineering

Extract a principle rather than copying an exterior everywhere.

Assess demand, structure, contamination, safe access, storage, overflow and upkeep.

Constructed Runoff Estimate

Roof area = 500 m²; rainfall = 0.6 m; runoff coefficient = 0.8.

Gross collected runoff = 500 × 0.6 × 0.8 = 240 m³.

Usable yield can be lower after first-flush, conveyance and storage losses.

Compare alternatives delivering the same service, including cost, materials, energy, repair access and equitable use.

5.22 From Inspiration to Innovation

Define whether a shaded waiting-area prototype improves comfort while maintaining daylight, access and air movement.

Measure comparable times, weather and occupancy. Air temperature and surface temperature require different suitable measurements; one infrared surface reading does not establish surrounding air temperature.

Development Record

Source principle → local need → assumptions → prototype → measurement → revision.

Include unfavourable results and trade-offs. Better shade may require changes to daylight or maintenance arrangements.

Possible themes include water harvesting, passive cooling, repairable products and material efficiency.

5.23 Case Study: A Tiny Sanskrit Corpus

Task: Map pre-segmented Sanskrit surface forms to supplied bases using a frozen dictionary lookup.

Nouns use selected stems; verbs use selected roots. The task is not OCR, unrestricted sandhi splitting, parsing or translation.

These are teaching sentences, not ancient quotations. The glosses explain inputs; the lookup does not compute translation.

Frozen Lookup

S1–S3 supply build annotations. S4 checks processing. Freeze the seven-entry table before S5–S6.

Procedure

1. Apply Unicode NFC normalisation.
2. Split the already segmented sentence at whitespace.
3. Look up each exact token.
4. Return its supplied base when found.
5. Otherwise return UNKNOWN.
6. Compare with separately retained references.

Use ṃ consistently; converting another convention requires a documented rule.

No parameters are trained. The method cannot infer unseen roots or bases. Repeated familiar tokens mean the split demonstrates workflow rather than broad linguistic generalisation.

5.24 Case Outputs and Evaluation

The frozen lookup reproduces these outputs:

Reproduce the Counts

• Total tokens = 6.
• Answered = 4.
• Correct answered = 4.
• Unknown = 2.

Coverage = 4/6 ≈ 66.7%.

Accuracy among answered = 4/4 = 100%.

Overall exact-match accuracy, counting abstentions as incorrect = 4/6 ≈ 66.7%.

The 100% value applies only to four answered tokens. It does not establish complete understanding or six correct analyses.

Error Analysis

pustakaṃ and paṭhati are absent from the lexicon. These are coverage gaps, not OCR or segmentation failures, because inputs were clean and pre-segmented.

Adding the exposed answers and rescoring the same test measures familiarity with that test, not independent generalisation.

A next experiment can compare lookup and morphology tools on a larger permission-cleared corpus, with documented genre, conventions and held-out target.

Retain inputs, partitions, table, normalisation, outputs and formulas for reproduction.

5.25 Solved Questions and Quick Revision

Digitisation Versus Preservation

Digitisation creates a representation. Preservation maintains bytes, context and usability through metadata, copies, fixity, recovery and managed change.

TKDL Versus IKS Division

TKDL supports defensive prior-art examination. The IKS Division has a broader role in research, education and dissemination.

Why Is Panini Relevant to Computation?

Rules, conditions, technical terms and derivations support study of formal representation. Implementations still require explicit modelling and validation.

Why Report Coverage?

A system can answer only easy items. In the corpus case, 100% answered-item accuracy accompanies 66.7% coverage and 66.7% overall accuracy.

Unit Recall Table

Formula Strip

• CER = character edits / reference characters.
• WER = word edits / reference words.
• Precision = TP/(TP + FP).
• Recall = TP/(TP + FN).
• F1 = 2TP/(2TP + FP + FN).
• Coverage = answered / total.
• Answered-item accuracy = correct / answered.
• Overall accuracy = correct / total, under the stated abstention convention.
• Gross runoff = area × rainfall depth × coefficient.

Short-Answer Questions

1. List six manuscript metadata fields.
2. What does a checksum establish, and what does it not establish?
3. Explain defensive protection.
4. Distinguish language and script.
5. Explain deva + indra → devendra.
6. Distinguish morphology and syntax.
7. Calculate F1 for TP 12, FP 3 and FN 4.
8. Reproduce the corpus case’s coverage.
9. Explain leakage and domain shift.
10. Propose a measured engineering project.

Long-Answer Questions

1. Design a digitisation and preservation workflow with permission and recovery checks.
2. Compare TKDL, the IKS Division and relevant IPR distinctions.
3. Explain Panini’s computational relevance and the limits of the analogy.
4. Describe an ancient-text NLP pipeline and error propagation.
5. Explain OCR, sandhi, morphology and parsing with examples.
6. Analyse the tiny Sanskrit corpus, its outputs and evaluation limits.
7. Develop a responsible IKS-inspired engineering proposal.

Common Errors

• Assuming an online scan permits commercial reuse.
• Calling the full TKDL database unrestricted.
• Equating Sanskrit grammar with guaranteed AI performance.
• Treating fluent generation as verified translation.
• Erasing diacritics without documenting information loss.
• Tuning on test answers and claiming independence.
• Reporting 100% accuracy while hiding abstentions.
• Inferring thermal or water-quality benefits from images.

Frequently Asked Questions

Related B.Tech First Semester Study Materials

CSE 1st Year All Subject PYQs → CSE 1st Year All Subject Notes → CSE 1st Year All Subject Syllabus →