Paper deep dive
Bangla Sentence Function Classification: Corpus Development, Model Benchmarking, and Interpretability
Swapnil Kundu Argha, Abdullah Al Shafi, Rowzatul Zannat, Shoumik Barman Polok, Abdul Muntakim, Jannatul Ferdousi, M. A. Moyeen
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 96%
Last extracted: 9/20/2026, 4:02:41 AM
Summary
This paper addresses the lack of benchmark resources for Bangla sentence function classification by introducing a novel corpus of 10,000 manually annotated Bangla sentences across four functional categories: declarative, interrogative, imperative, and exclamatory. The study benchmarks various feature representations (BoW, TF-IDF, Word2Vec) and classical machine learning classifiers, alongside two heterogeneous ensemble models (Single-Level and Double-Level Ensembles). Results indicate that TF-IDF outperforms Word2Vec, and the Double-Level Ensemble (DLE) model with TF-IDF achieves the highest performance with 0.95 accuracy and macro-F1. The work also includes interpretability analysis using LIME and establishes strong baselines for future research.
Entities (13)
Relation Signals (9)
Bangla Sentence Function Corpus → haslabel → Interrogative
confidence 99% · manually annotated into four functional categories, namely declarative, interrogative, imperative, and exclamatory
Bangla Sentence Function Corpus → haslabel → Imperative
confidence 99% · manually annotated into four functional categories, namely declarative, interrogative, imperative, and exclamatory
Bangla Sentence Function Corpus → haslabel → Exclamatory
confidence 99% · manually annotated into four functional categories, namely declarative, interrogative, imperative, and exclamatory
Bangla Sentence Function Corpus → haslabel → Declarative
confidence 99% · manually annotated into four functional categories, namely declarative, interrogative, imperative, and exclamatory
Double-Level Ensemble → achievesmetric → 0.95
confidence 95% · The DLE model with TF-IDF features achieves the best performance with accuracy and macro-F1 of 0.95
LIME → usedfor → Interpretability
confidence 95% · LIME-based interpretability provides insights into model predictions
Double-Level Ensemble → usesfeature → TF-IDF
confidence 95% · The DLE model with TF-IDF features achieves the best performance
Bangla Sentence Function Corpus → →
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Automatic sentence function identification is important for many downstream natural language processing (NLP) applications such as dialogue systems, text-to-speech synthesis, and machine translation. However, benchmark resources for Bangla sentence function classification remain limited. To mitigate this gap, this paper introduces a corpus of 10,000 Bangla sentences, manually annotated into four functional categories, namely declarative, interrogative, imperative, and exclamatory. The corpus is nearly balanced across the four classes, with high annotation reliability reflected by a Fleiss\' Kappa of 0.82. Furthermore, we evaluate multiple feature representations, including Bag-of-Words (BoW), TF-IDF, and Word2Vec, with several classical machine learning classifiers. In addition, two heterogeneous ensemble models, namely Single-Level Ensemble (SLE) and Double-Level Ensemble (DLE), are utilized to improve classification performance. Experimental results show that TF-IDF consistently outperforms Word2Vec, likely due to its ability to emphasize discriminative lexical cues associated with sentence functions, particularly given the relatively small corpus used to train Word2Vec. The DLE model with TF-IDF features achieves the best performance with accuracy and macro-F1 of 0.95, demonstrating the effectiveness of sparse lexical representations and heterogeneous ensemble learning for this task. Further cross-validation confirms the robustness of the approach, while LIME-based interpretability provides insights into model predictions. The developed corpus and model benchmarking establish strong baselines for Bangla sentence function classification.
Tags
Links
- Source: https://arxiv.org/abs/2609.13869v1
- Canonical: https://arxiv.org/abs/2609.13869v1
Trouble viewing inline? Open PDF directly →
Full Text
28,071 characters extracted from source content.
Expand or collapse full text
Bangla Sentence Function Classification: Corpus Development, Model Benchmarking, and Interpretability Swapnil Kundu Argha ∗ , Abdullah Al Shafi † , Rowzatul Zannat ‡ , Shoumik Barman Polok § , Abdul Muntakim ¶ , Jannatul Ferdousi ∥ , M.A. Moyeen ∗ Department of CSE, Khulna University of Engineering & Technology, Bangladesh ∗†§∗ Institute of ICT, Khulna University of Engineering & Technology, Bangladesh †∗ Department of CSE, Daffodil International University, Bangladesh ‡ Department of Computer Science, Kennesaw State University, United States ¶∥ swapnilkundu01@gmail.com ∗ , abdullah@iict.kuet.ac.bd † , w.rzrowza@gmail.com ‡ , polokbarman874@gmail.com § , amuntaki@students.kennesaw.edu ¶ , jannatulferdousi3110@gmail.com ∥ , moyeen.kuet@gmail.com ∗ Abstract—Automatic sentence function identifica- tion is important for many downstream natural lan- guage processing (NLP) applications such as dialogue systems, text-to-speech synthesis, and machine trans- lation. However, benchmark resources for Bangla sentence function classification remain limited. To mitigate this gap, this paper introduces a corpus of 10,000 Bangla sentences, manually annotated into four functional categories, namely declarative, interroga- tive, imperative, and exclamatory. The corpus is nearly balanced across the four classes, with high annotation reliability reflected by a Fleiss’ Kappa of 0.82. Further- more, we evaluate multiple feature representations, in- cluding Bag-of-Words (BoW), TF-IDF, and Word2Vec, with several classical machine learning classifiers. In addition, two heterogeneous ensemble models, namely Single-Level Ensemble (SLE) and Double-Level Ensem- ble (DLE), are utilized to improve classification perfor- mance. Experimental results show that TF-IDF consis- tently outperforms Word2Vec, likely due to its ability to emphasize discriminative lexical cues associated with sentence functions, particularly given the relatively small corpus used to train Word2Vec. The DLE model with TF-IDF features achieves the best performance with accuracy and macro-F1 of 0.95, demonstrating the effectiveness of sparse lexical representations and heterogeneous ensemble learning for this task. Fur- ther cross-validation confirms the robustness of the approach, while LIME-based interpretability provides insights into model predictions. The developed corpus and model benchmarking establish strong baselines for Bangla sentence function classification. Index Terms—sentence classification, ensemble learning, word embeddings, interpretability I. Introduction With about 250 million native speakers, mostly in Bangladesh and West Bengal of India, Bangla is one of the most frequently spoken languages in the world. Compared to languages like English or Chinese, Bangla is still relatively deficient in automated language processing resources [ 1]. But automated systems require solutions to understand the functional expression of large amount of textual data as digital Bangla content has rapidly expanded its presence through various platforms [2]. Identifying sentence functions requires an evaluation of whether a sentence functions as declarative, interrogative, imperative, or exclamatory. The different functional cat- egories allow a speaker to deliver his intended expres- sion while controlling how listeners will understand the information delivered. For example, the declarative“েস বািড়েত যায় (He goes home)”expresses a statement, while the interrogative“িভতের আসেত পাির? (May I come in?)” functions to obtain permission. Similarly, the imperative “েগালমাল কিরও না। (Don't make a fuss.)”delivers a command, while the exclamatory“ৈশেবর স্মৃিত কত মধ ু র! (How sweet childhood memories are!)”expresses strong feelings. Reliable functional classification of a sentence has broad implications across various downstream natural language processing (NLP) applications. Dialogue systems [ 3] rely on the identification of the type of query to generate ap- propriate responses. Machine translation [ 4] benefits from preserving the intent of the sentence in multiple languages to maintain communicative accuracy. Sentiment and emo- tion detection can benefit from recognizing exclamatory and other functional cues, while stance detection can use the sentence function to better capture the communicative intent of the speaker [ 5]. Text-to-speech synthesis uses functional cues to produce suitable intonation patterns [6]. In short, without accurate understanding of sentence- level communicative function, higher-level NLP systems may fail to capture speaker intent, leading to reduced effectiveness in real-world language understanding. The main contributions of this work are threefold. •We present a novel Bangla sentence function corpus with 10,000 manually annotated sentences across four functional categories. To ensure high-quality labeling of the corpus, a reliable multi-annotator annotation framework with agreement validation is designed. Furthermore, the corpus is released as a publicly arXiv:2609.13869v1 [cs.CL] 12 Sep 2026 available resource 1 to support reproducible research. •A comprehensive benchmarking study for sentence function classification is conducted on various feature representations and classical machine learning mod- els. We also benchmark two heterogeneous ensemble architectures to effectively combine diverse classifiers. •We further incorporate interpretability and error analysis, providing insights into model behavior and linguistic challenges. I. Literature Review Early Bangla text classification research relied mainly on traditional machine learning methods due to the lack of large-scale datasets. Bijoy et al. [1] used ensemble models such as Random Forest and XGBoost for sentence classification, achieving 96.39% accuracy, while Das et al. [7] applied Decision Tree models for structural sen- tence classification with 93.72% accuracy. Furthermore, [8] introduced a large Bangla news dataset and showed that Word2Vec embeddings outperformed TF-IDF-based approaches, achieving a 0.96 F1-Score. With the rise of deep learning, Sikder et al. [9] proposed a hybrid CNN-BiLSTM model that achieved 98% F1- Score, demonstrating the effectiveness of neural architec- tures over traditional methods. Recently, Al Shafi et al. [2] introduced BiST, a Bangla-English bilingual corpus for sentence structure and tense classification, with experi- ments showing the superiority of dual-encoder models over multilingual baselines. I. Proposed Method for Sentence Function Classification The complete workflow of our proposed method is shown in Fig.1, which includes the sequence of corpus creation, data preprocessing, feature extraction, model training, evaluation, and interpretation. A. Development of Bangla Sentence Function Corpus The construction of a high-quality corpus is essential for robust text classification. Fig.2presents the workflow of the corpus creation process, including data collection, preprocessing, annotation, and quality verification. Data Collection and Preprocessing:We prioritized collecting Bangla sentences that demonstrate different usage patterns through formal and informal usage across various online platforms which users can access without restrictions. Several popular Bengali newspapers serve as our primary sources that deliver high-quality formal content. To create a wider linguistic range, we also added Bengali blogs and online magazines and youtube comments together with digitized storybooks and novels which exist in the public domain. We have used automated 1 https://github.com/AbdullahRatulk/Bangla_Sentence_ Function_Classification_Corpus Start Corpus Development Raw Bangla text corpus collection and Labeling Preprocessing Tokenization, Stop-word removal, Punctuation Removal Feature Extraction Methods Word2Vec BOW TF-IDF CBOWSkipGram Ensemble Techniques Single Level Ensemble Double Level Ensemble Validation Methods K-Fold Stratified K-Fold Shuffle Split Stratified Shuffle Interpretability using LIME Comparative Analysis Final Results Classifier Training LSVC RF DT MNB Ridge Ada LR GB Figure 1: Overview of the proposed system. web scraping techniques through Python scripts that used Beautiful Soup 2 to collect the initial raw sentences. The collected text then underwent a rigorous prepro- cessing pipeline that includes removing HTML tags and non-Bengali characters, sentence segmentation using a Bengali sentence tokenizer, deduplication, and noise fil- tering to maintain focus on well-formed sentences. Start Final Dataset Data Collection Data Preprocessing Agreement Calculation (Fleiss' Kappa) Multi- Annotator Labeling Conflict Resolution Confidence Filtering Quality CheckFinal Dataset Figure 2: Workflow of the corpus development process. Annotation Guidelines:Based on standard linguis- tic theory, we define four mutually exclusive functional classes, namely declarative, interrogative, imperative, and exclamatory for annotators. A sentence was annotated as declarative if it conveyed a statement, fact, or opinion without expressing a query, command, or strong emotion. For example, ‘আিম আজ স্ক ু েল িগেয়িছলাম।(I went to school today.)’. In contrast, sentences were annotated as interrogative when they were intended to seek information or clarification, often characterized by the presence of interrogative markers such asিক, েকন, েকাথায়, কেব, and িকভােব. For instance, “ত ু িম েকাথায় যাচ্ছ? (Where are you going?)”. Sentences expressing commands, requests, suggestions, or instructions were categorized as imperative. ‘দরজা 2 https://pypi.org/project/beautifulsoup4/ বন্ধ কেরা।(Close the door.)’ would be an example of such a sentence. Meanwhile, sentences conveying strong emotions, such as surprise, joy, anger, or admiration, were labeled exclamatory. An example would be ‘আহা! কী স ু ন্দর দ ৃ শয্!(What a beautiful view!)’. Algorithm 1Corpus Annotation with Fleiss’ Kappa Require:SentencesS=s 1 ,...,s n Ensure:Final labelsF=f 1 ,...,f n 1:Define annotatorsP=a 1 ,a 2 ,a 3 and label setsF 2:for alls i anda j do 3:Annotator assignsf ij ∈F, 4:end for 5:for alls i do 6:Compute item agreementA i 7:end for 8: ̄ A= 1 n ∑ n i=1 A i 9:Compute expected agreementA e 10:Fleiss’ Kappa,κ= ̄ A−A e 1−A e 11:for alls i do 12:f i ←argmax f∈F ∑ 3 j=1 1(f ij =f) 13:ifno clear majority emergesthen 14:consensus discussion among annotators 15:end if 16:end for returnAnnotated corpus with final labelsF Corpus Annotation:The multi-annotator framework which describes the corpus annotation process operates according to the steps which Algorithm1specifies. Three annotators assess each sentence in the corpus by applying their designated function labels from the defined function label set in annotation guideline. The reliability of the an- notations is evaluated through inter-annotator agreement (IAA) assessment which uses Fleiss’ Kappa method. The final label for each sentence comes using majority voting. The final label assignment process requires annotators to hold a consensus discussion when a situation arises with- out a distinct majority. The resulting labels collectively form the finalized annotated corpus. Final Corpus:The final corpus was created from the original collection of approximately 24,000 Bangla sentences from which 10,000 sentences underwent prepro- cessing and annotation to establish a reliable benchmark for sentence function classification. B. Data Preprocessing The raw Bangla sentences undergo a series of cleaning and preprocessing steps before feature extraction to reduce noise while preserving linguistic cues essential for function classification. The cleaning phase applies sequential oper- ations, including the elimination of hashtags and special characters that do not carry linguistic meaning, space normalization, etc. Common stop words for Bangla 3 are 3 https://github.com/stopwords-iso/stopwords-bn removed with a critical exception made for wh-question words such asকী(what),েকমন(how),কখন(when),েকাথায় (where), as these words are essential for identifying inter- rogative and/or exclamatory sentences. All punctuation marks are removed to ensure that punctuation does not serve as a direct cue for sentence function, forcing the classification model to rely on lexical and structural cues instead. A Bangla-aware custom tokenizer splits text on whitespace and returns only non-empty tokens after all cleaning steps, with the same tokenizer used across all feature extractors for fair comparison. In addition, the four function labels are encoded into integer IDs from 0 to 3 for model training. C. Feature Extraction We employ multiple feature extraction techniques to represent the preprocessed text as numerical vectors, namely Bag of Words (BoW), TF-IDF, and Word2Vec. BoW [7] counts the frequency of each word or n-gram in a document while ignoring word order and grammar, implemented with CountVectorizer using configurable n- gram ranges. On the other hand, TF-IDF [1] weights terms by their frequency in a document and their rarity across the corpus according to the mathematical formula- tion where TF represents term frequency in a document and IDF represents inverse document frequency across the corpus. For example, the termকী(what) appearing in an interrogative document receives high TF in that document and moderate IDF, resulting in high TF-IDF weight that marks it as a useful feature for identifying interrogatives. Hidden Output Context Words Predicted Target Backpropagation Input Hidden Target Word (a) CBOW 0 0 0 1 0 0 0 1 0 0 0 0 0 0 1 Actual Target Actual Target (াস) (াস) (দাও) (দাও) (পািন) 0 0 1 0 0 0 0 0 0 1 0 0 0 0 1 Output Context Words 0 0 1 0 0 0 0 0 0 1 0 0 0 0 1 (পািন) Backpropagation Input (b) Skip-Gram আমােক এক াস পািন দাও Figure 3: Comparison of (a) CBOW and (b) Skip-Gram architectures where CBOW predicts the target word from surrounding context words while Skip-Gram predicts con- text words from the target word. In contrast, Word2Vec [8] is a two-layer neural network trained on the corpus to capture semantic relationships using two different architectures as shown in Fig. 3: (i) Continuous Bag of Words (CBOW) which predicts the target word from its surrounding context and performs well for frequent words, and (i) Skip-gram which predicts (a) Single-Level Ensemble (SLE) architecture(b) Double-Level Ensemble (DLE) architecture Figure 4: Architecture of the two heterogeneous ensemble models: (a) SLE where multiple classical classifiers feed into a Voting Classifier for final prediction; (b) DLE with two levels of voting classifiers leading to final prediction. context words from the target word and tends to capture rare words and semantic relationships better. D. Classification We evaluate multiple classifiers spanning classical ML and heterogeneous ensemble approaches. Classical ML Classifiers:The classical classifiers trained using the extracted features include Decision Tree (DT), Multinomial Naive Bayes (MNB), Logistic Regres- sion (LR), Linear Support Vector Classifier (LSVC), Ridge classifier, Gradient Boosting (GB), AdaBoost (Ada), and Random Forest (RF). Heterogeneous Ensemble Classifiers:We design two heterogeneous ensemble architectures, namely Single Level Ensemble (SLE) and Double Level Ensemble (DLE), to leverage the complementary strengths of different clas- sical classification algorithms. Fig. 4presents a detailed overview of both architectures. SLE encompasses all classical ML classifiers operating on fixed feature vectors. These classifiers then feed into a voting classifier to produce the final classification output. On the other hand, DLE implements a more sophisti- cated hierarchical voting mechanism with two distinct aggregation levels. In the first level, the base classifiers are organized into two groups based on their underlying algorithmic characteristics. Group 1 consists of MNB, LR, DT, and LSVC, whose outputs are combined into voting classifier 1 (VC1). Group 2 comprises the homogeneous ensemble methods including RF, GB, and AdaBoost, whose predictions are aggregated into voting classifier 2 (VC2). In the second level, the outputs from VC1 and VC2 are fed into a third voting classifier (VC3) that performs the final aggregation to produce the ultimate prediction. This hierarchical structure with two layers of voting enables the model to capture different levels of decision making, first within similar classifier types and then across the two broad categories of classifiers. It is to be noted that the voting classifiers use soft vot- ing, where the final prediction is obtained by averaging the predicted probabilities of all base classifiers and selecting the class with the highest score. Cross-Validation and Hyperparameter Tuning: Robust evaluation is ensured through multiple cross- validation (CV) strategies including (i) K-fold where data is split into K folds with the model training on (K- 1) folds and validating on 1 fold rotating K times, (i) stratified K-fold which maintains original class proportions in each fold, (i) shuffle split with multiple random train- validation splits, and (iv) stratified shuffle split combining random splitting with stratification. Grid search with 5-fold CV is used to optimize hyperparameters for each model with macro-F1 score as the objective metric, and after identifying the best parameters, models are refitted on the full training set. E. Interpretability using LIME The research employs Local Interpretable Model- agnostic Explanations (LIME) [ 10] to analyze the model’s prediction results. This local approximation method en- ables us to identify which words or phrases have the greatest impact on model’s prediction. IV. Experimental setup and Result analysis Table I: Class distribution and sentence length statistics for the Bangla Sentence Function Corpus. Function Total Length Bins (Tokens) Avg. Len 0-8 9-14 15+ Declarative 2,487 881 93267410.78 Interrogative 2,532 895 90273511.02 Imperative 2,511 845 83383311.29 Exclamatory 2,470 804 83383312.52 Total10,000 3,425 3,500 3,07511.40 Table I: Comparative analysis of various feature representation and classifiers based on performance evaluation. All the metrics except accuracy are macro-averaged. Metric Feature NB LSVC DT LR RF Ridge Ada GB SLE DLE Accuracy BOW 0.86 0.93 0.84 0.93 0.94 0.90 0.88 0.94 0.940.94 TF-IDF 0.87 0.94 0.85 0.94 0.94 0.92 0.89 0.94 0.940.95 CBOW 0.83 0.90 0.81 0.89 0.91 0.89 0.86 0.910.930.92 SkipGram 0.84 0.91 0.82 0.90 0.92 0.90 0.87 0.92 0.930.94 Precision BOW 0.86 0.93 0.84 0.94 0.94 0.91 0.88 0.94 0.940.95 TF-IDF 0.87 0.94 0.85 0.94 0.94 0.92 0.89 0.94 0.940.95 CBOW 0.83 0.90 0.81 0.89 0.91 0.89 0.86 0.910.940.93 SkipGram 0.84 0.91 0.82 0.90 0.92 0.90 0.87 0.92 0.930.94 Recall BOW 0.86 0.93 0.84 0.93 0.94 0.90 0.88 0.94 0.940.95 TF-IDF 0.87 0.94 0.85 0.94 0.94 0.92 0.89 0.94 0.940.95 CBOW 0.83 0.90 0.81 0.89 0.91 0.89 0.86 0.910.93 0.93 SkipGram 0.84 0.91 0.82 0.90 0.92 0.90 0.87 0.92 0.930.94 F1-Score BOW 0.86 0.93 0.84 0.93 0.94 0.90 0.88 0.94 0.940.94 TF-IDF 0.87 0.94 0.85 0.94 0.94 0.92 0.89 0.94 0.940.95 CBOW 0.83 0.90 0.81 0.89 0.91 0.89 0.86 0.910.940.93 SkipGram 0.84 0.91 0.82 0.90 0.92 0.90 0.87 0.92 0.930.94 Table I: Cross-validation performance evaluation of SLE and DLE models (mean ± std). Here, we set K = 5. CV techniquesFeatures SLEDLE Accuracy Macro-F1Accuracy Macro-F1 K-Fold BoW0.94±0.012 0.94±0.0120.95±0.009 0.95±0.009 TF-IDF 0.94±0.011 0.94±0.0110.95±0.008 0.95±0.008 CBOW 0.92±0.018 0.92±0.0180.93±0.015 0.93±0.015 SkipGram 0.93±0.015 0.93±0.0150.94±0.012 0.94±0.012 Stratified K-Fold BoW0.94±0.010 0.94±0.0100.95±0.007 0.95±0.007 TF-IDF 0.94±0.009 0.94±0.0090.95±0.006 0.95±0.006 CBOW 0.92±0.015 0.92±0.0150.93±0.012 0.93±0.012 SkipGram 0.93±0.012 0.93±0.0120.94±0.010 0.94±0.010 Shuffle Split BoW0.94±0.018 0.94±0.0180.95±0.014 0.95±0.014 TF-IDF 0.94±0.016 0.94±0.016 0.95±0.012 0.95±0.012 CBOW 0.92±0.024 0.92±0.024 0.93±0.020 0.93±0.020 SkipGram 0.93±0.020 0.93±0.0200.94±0.016 0.94±0.016 Stratified Shuffle Split BoW0.94±0.014 0.94±0.0140.95±0.011 0.95±0.011 TF-IDF 0.94±0.013 0.94±0.013 0.95±0.010 0.95±0.010 CBOW 0.92±0.020 0.92±0.020 0.93±0.017 0.93±0.017 SkipGram 0.93±0.017 0.93±0.0170.94±0.014 0.94±0.014 A. Inter-Annotator Agreement The Fleiss’ Kappa (κ) value for the developed corpus is 0.82, which is treated as ‘Almost Perfect’ in the standard, indicating a high level of inter-annotator agreement and strong consistency in the labeling process. B. Corpus Statistics As shown in Table I, each of the four functional classes contributes approximately 2,500 sentences, ensuring a near-equal distribution that prevents bias towards any single function. The three sentence length analysis bins which include 0-8 tokens and 9-14 tokens and 15+ tokens show that all functional classes use a combination of brief sentences, intermediate sentences, and extended sentences to create a corpus through which we can train models that can handle different input lengths. C. Experimental Setup and Evaluation Metrics The final curated dataset was partitioned into two subsets using stratified sampling, a training-validation set of 8000 sentences, and an unseen test set of 2000 sentences for the final unbiased evaluation. Model performance is assessed using standard metrics, including accuracy, precision, recall, and F1-Score. D. Result Analysis TableIIpresents the comprehensive performance com- parison across four feature representations and eight classifiers. Linear models, including LSVC, Ridge, and LR demonstrate strong performance across most feature types. RF and GB emerged as the top performers, RF particularly excelled with TF-IDF features reaching 0.94 accuracy and F1-Score. DLE achieved 0.95 on the TF-IDF features, while SLE maintained 0.94 on most metrics. In addition, the superior performance of TF-IDF over Word2Vec can be attributed to the discriminative lexical Table IV: Representative genuine misclassification cases with DLE on TF-IDF features. TextActualPredictedPrimary Error Source বাংলােদেশর এই পৰ্াক ৃ িতক দ ৃ শয্ কত মেনাম ু গ্ধকর!(How mesmerizing this natural scenery of Bangladesh is!) ExclamatoryDeclarativeExclamatory sentence with declara- tive syntax. এই নীিতমালার কাযর্কািরতা িনেয় েকােনা সেন্দহ আেছ?(Is there any doubt about the effectiveness of this policy?) InterrogativeDeclarativeRhetorical question interpreted as statement. সংিশ্লষ্ট কত ৃ র্পক্ষেক দৰ্ুত পৰ্েয়াজনীয় বয্বস্থা গৰ্হণ করেত হেব।(The relevant authorities must take necessary measures promptly.) DeclarativeImperativeObligation modality interpreted as directive. গেবষণা অন ু যায়ী জলবায় ু পিরবতর্েনর পৰ্ভাব কৰ্েমই ব ৃ িদ্ধ পােচ্ছ।(According to research, the impact of climate change is increasing.) DeclarativeExclamatoryStrong evaluative tone misinter- preted as exclamation. markers of sentence functions, such as interrogative words, directive, and exclamatory expressions, etc. TF-IDF emphasizes such class-specific terms, whereas Word2Vec captures broader semantic relationships and may not pre- serve these functional cues as explicitly, particularly when trained on a relatively small corpus. Thus, sparse lexical weighting is well suited to this task. The heterogeneous ensembles are also tested against several different CV methods to confirm their robustness, and the results are shown in Table I. The results confirm the stability of our ensemble approaches. Both SLE and DLE reach 92-95% accuracy with standard deviations between 0.6% and 2.4% for all feature representation. Table IVpresents representative examples of such misclassified sentences which highlight common chal- lenges. Exclamatory sentences are frequently misclassified as declarative when exclamation expressions are under- weighted. The use of directive verbs creates problems that lead to difficulties in determining between declarative and imperative sentence structures. Questions that contain embedded commands are sometimes treated as direct orders. Figure 5: Interpreting model decisions through LIME, visualizing feature importance for predicted sentence class. Fig.5provides an example of clear and intuitive illustration of the model’s decision-making process using LIME, highlighting the contribution of individual words toward the predicted Imperative class. Key token ‘খাও’ is assigned higher importance scores, which aligns well with their semantic association with actions and directives. This indicates that the model is able to capture meaningful lexical cues relevant to imperative sentence construction. Similar patterns were observed for the other classes. V. Conclusion The research presents a high-quality Bangla sentence function corpus together with a study that compares traditional learning methods and heterogeneous ensemble learning techniques. The experimental findings demon- strate that TF-IDF which represents data in sparse form continues to perform effectively for the current task. The ensemble methods demonstrate superior performance sta- bility among all tested models whereas DLE produces the highest overall performance results. The analysis of errors shows difficulties that arise from linguistic uncertainty which exists between different types of sentences. The newly developed corpus together with its benchmark tests offers valuable resources which enable future researchers to advance Bangla NLP studies. The future direction would be using transformer-based language models and larger datasets to enhance both classification accuracy and system resilience. References [1]M. H. I. Bijoy, M. Hasan, A. N. Tusher, M. M. Rahman, M. J. Mia, and M. Rabbani, “An automated approach for bangla sentence classification using supervised algorithms,” in2021 12th International Conference on Computing Communication and Networking Technologies (ICCCNT). IEEE, 2021, p. 1– 6. [2]A. Al Shafi, S. K. Argha, M. Moyeen, A. Muntakim, and S. B. Polok, “Bist: A gold standard bangla-english bilingual corpus for sentence structure and tense classification with inter- annotator agreement,” inSIGUL 2026 Joint Workshop with ELE, EURALI, and DCLRL: Towards Inclusivity and Equality: Language Resources and Technologies for Under-Resourced and Endangered Languages@ LREC 2026. European Language Resources Association (ELRA), 2026, p. 143–152. [3]Z. Yi, J. Ouyang, Z. Xu, Y. Liu, T. Liao, H. Luo, and Y. Shen, “A survey on recent advances in llm-based multi-turn dialogue systems,”ACM Computing Surveys, vol. 58, no. 6, p. 1–38, 2025. [4]S. Castilho and R. Knowles, “A survey of context in neural machine translation and its evaluation,”Natural Language Pro- cessing, vol. 31, no. 4, p. 986–1016, 2025. [5]A. Al Shafi, M. M. Islam, S. I. Hossain, and K. M. A. Hasan, “Stancemoe: Mixture-of-experts architecture for stance detec- tion,” inProceedings of the 2nd International Workshop on Nakba Narratives as Language Resources @ LREC 2026. Euro- pean Language Resources Association (ELRA), 2026, p. 244– 251. [6]X. Tan, J. Chen, H. Liu, J. Cong, C. Zhang, Y. Liu, X. Wang, Y. Leng, Y. Yi, L. Heet al., “Naturalspeech: End-to-end text- to-speech synthesis with human-level quality,”IEEE Transac- tions on Pattern Analysis and Machine Intelligence, vol. 46, no. 6, p. 4234–4245, 2024. [7]R. K. Das, S. S. Sammi, K. Kobra, M. R. Ajmain, S. A. khushbu, and S. R. H. Noori, “Analysis of bangla transformation of sen- tences using machine learning,” inInternational Conference on Deep Learning, Artificial Intelligence and Robotics. Springer, 2022, p. 36–52. [8]M. M. Islam and M. T. Alam, “Bard: Bangla article classifica- tion using a new comprehensive dataset,” in2018 International Conference on Bangla Speech and Language Processing (ICB- SLP), 2018. [9]J. Sikder, P. Chakraborty, U. K. Das, and K. Dhar, “A hybrid approach for bengali sentence validation,”Artificial Intelligence Review, vol. 57, no. 11, p. 316, 2024. [10]A. M. Salih, Z. Raisi-Estabragh, I. B. Galazzo, P. Radeva, S. E. Petersen, K. Lekadir, and G. Menegaz, “A perspective on explainable artificial intelligence methods: Shap and lime,” Advanced Intelligent Systems, vol. 7, no. 1, p. 2400304, 2025.