Update README.md

README.md

@@ -9,4 +9,28 @@ tags:
 - '-code-mixed-sentiment-analysis'
 - sentiment
 - '-dataset'
----
+---
+📊 **Data Analysis**
+
+The **En-Bn-Code-Mixed-Two-Class-Sentiment-Dataset** underwent a detailed analysis to understand its linguistic diversity, sentiment distribution, and code-mixing characteristics. The following observations summarize the key insights obtained during dataset exploration and preprocessing.
+
+**Sentiment Distribution:**
+The dataset follows a binary sentiment structure consisting of positive and negative classes. Both classes are distributed nearly evenly, ensuring that the dataset remains balanced and suitable for supervised machine learning tasks. This balanced structure minimizes class bias and supports fair model evaluation during binary sentiment classification.
+
+**Text Length and Token Variation:**
+A broad range of review lengths was observed across all categories. English reviews tend to have shorter average token counts, while Bengali and code-mixed texts demonstrate increased token length due to the nature of translation and word expansion. The fully translated Bengali reviews (5ⁿ category) exhibit the highest average word count, while original English reviews (1ⁿ) remain the most concise. This variation provides valuable diversity for training models to handle different sentence complexities and structures.
+
+**Linguistic Composition and Code-Mixing Intensity:**
+The dataset effectively captures multiple degrees of language mixing through its four structured categories. The Original English subset maintains pure English syntax, the Selective POS Translated subset introduces light code-mixing via selective translation of adjectives, adverbs, and conjunctions, while the Selective + Roman Bengali subset intensifies the mixing by adding Roman-script Bengali words. Finally, the Fully Translated Bengali subset represents complete linguistic transformation. This gradual increase in code-mixing intensity makes the dataset ideal for studying multilingual interference and domain adaptation in sentiment analysis.
+
+**Lexical and POS-Level Variation:**
+The application of POS-based selective translation ensures a meaningful linguistic shift without distorting sentence semantics. Adjectives and adverbs—key sentiment-bearing components—are primarily translated, allowing for a balanced blend of English grammatical structure with Bengali emotional tone. This selective inclusion enhances sentiment preservation across languages.
+
+**Script Diversity and Transliteration Impact:**
+The Roman Bengali transliteration layer introduces significant script variation, simulating real-world code-mixed writing behavior commonly observed in South Asian digital communication. The transliteration process produces multiple spelling variations of the same Bengali word, which helps capture orthographic irregularities and boosts the robustness of downstream NLP models.
+
+**Vocabulary Expansion:**
+The dataset exhibits notable vocabulary growth due to the introduction of Bengali and Roman Bengali tokens. This expansion diversifies the linguistic space and challenges language models to learn bilingual lexical representations. Distinct vocabularies for English, Bengali, and Roman Bengali words were separately stored to aid further lexical analysis and embedding training.
+
+**Applicability in NLP Research:**
+The dataset’s balanced sentiment structure, systematic code-mixing variation, and multilingual token distribution make it suitable for a range of research areas including code-mixed sentiment analysis, language identification, cross-lingual embeddings, and domain adaptation. The presence of structured categories also enables comparative performance analysis between monolingual, partially mixed, and fully translated text data.