3.1.1 NMT training and translation using sentence alignment

There are two main methods to translate the dataset: Statistical Machine Translation (SMT) and NMT. NMT is preferred due to its lower memory requirements and improved translation accuracy. NMT uses an attention mechanism that considers sentences’ semantic and syntactic contributions, making it more effective in maximizing translations. However, ensuring accurate sentence alignment between the source and target languages is challenging. To address this, a mechanism called sentence alignment (Shi et al. Reference Shi, Huang, Jian and Tang2021) is used, which aligns the sentences during the decoding phase using a beam search algorithm. This approach helps mitigate translation inaccuracies and ensures better overall translation quality. By employing NMT and sentence alignment, the dataset can be effectively translated while maintaining the integrity and accuracy of the data.

A large parallel corpus of en-hi and en-mr parallel text is used to train the NMT model for English-Hindi (en-hi) and English-Marathi (en-mr) translation. For this purpose, the Samanantar dataset (Ramesh et al. Reference Ramesh, Doddapaneni, Bheemaraj, Jobanputra, R., Sharma, Sahoo, Diddee, Kakwani and Kumar2022) is utilized, which is a significant publicly available parallel corpus of English-Indic, Indic-English, and Indic-Indic language pair sentences. This corpus contains 10.126M parallel sentences for en-hi and 3.627M parallel sentences for en-mr languages. From this corpus, 10.12 million en-hi parallel sentences and 3.621 million en-mr parallel sentences are used for training the NMT model. Additionally, 5K parallel sentences are used for validation, and 1K parallel sentences are used for testing. To ensure a balanced dataset, simple random sampling is employed. The NMT model is then trained using this dataset, enabling it to learn the translation patterns and improve its performance.

NMT model is trained using three different NMT architectures: Stacked-LSTM (Wu et al. Reference Wu, Schuster, Chen, Le, Norouzi, Macherey, Krikun, Cao, Gao and Macherey2016; Johnson et al. Reference Johnson, Schuster, Le, Krikun, Wu, Chen, Thorat, Viégas, Wattenberg and Corrado2017), Stacked-Bi-LSTM (Hasan et al. Reference Hasan, Alam, Chowdhury and Khan2019), and Transformer (Vaswani et al. Reference Vaswani, Shazeer, Parmar, Uszkoreit, Jones, Gomez, Kaiser and Polosukhin2017). These architectures are selected to explore different approaches for NMT. The training and implementation of these models are carried out using the OpenNMT-py 2.0 toolkit (Klein et al. Reference Klein, Kim, Deng, Senellart and Rush2017), which provides a reliable framework for building and training NMT models. In the experimental setting, each architecture is configured with specific hyperparameters, such as the number of layers, hidden units, and attention mechanisms, to optimize their performance. The training process involves feeding the parallel training data from the Samanantar dataset into the models and iteratively updating the model parameters using optimization techniques like stochastic gradient descent. During training, the models learn the input–output pairs’ statistical patterns and linguistic features, enabling them to generate accurate translations. The additional experimental setting for all these models is given below:

1. (1) Stacked-LSTM model: The Stacked-LSTM model (see Figure 1(a)) is composed of an encoder and decoder, each consisting of three stacked LSTM layers. During the training process, a pretrained word vector, specifically fastText (Grave et al. Reference Grave, Bojanowski, Gupta, Joulin and Mikolov2018), is used as input to the model. This choice of word vector assists in addressing the Out-Of-Vocabulary (OOV) issue that may arise. The model dimensions for the LSTM, Bi-LSTM, and attention models are 512. The Adam optimizer is employed, with a smoothing label value of 0.1 and gradient clipping set to 1.0. The learning rate is configured to 0.005, while the dropout rate is set to 0.3 for Hindi and 0.1 for Marathi, respectively.

2. (2) Stacked-Bi-LSTM model: The Stacked-Bi-LSTM model (see Figure 1(b)) follows the same experimental setup as the Stacked-LSTM model. Still, it uses bidirectional LSTM instead of unidirectional LSTM. This modification allows the model to consider past and future contexts when making predictions.

3. (3) Transformer model: This model (see Figure 1(c)) follows the experimental setup of (Vaswani et al. Reference Vaswani, Shazeer, Parmar, Uszkoreit, Jones, Gomez, Kaiser and Polosukhin2017). It incorporates 16 attention heads, with input and output vectors of dimension 1024. The feed-forward network in the inner layer has a dimension of 4096. For the Hindi language, a dropout rate of 0.3 is applied, while for the Marathi language, a dropout rate of 0.1 is utilized. The hyperparameter settings include an Adam optimizer with a smoothing label value of 0.1 to improve optimization stability. The learning rate is set to 0.003, ensuring an appropriate adjustment rate during training.

Figure 1.

NMT Architecture (a) LSTM based model, (b) bi-LSTM based model, (c) transformer based model (Vaswani et al. Reference Vaswani, Shazeer, Parmar, Uszkoreit, Jones, Gomez, Kaiser and Polosukhin2017).

The effectiveness of translation is measured using the BLEU metric (Papineni et al. Reference Papineni, Roukos, Ward and Zhu2002), which quantifies the similarity between the translated output and reference translations. BLEU is calculated using equation (1). This metric provides a quantitative measure of translation quality.

(1) \begin{equation} BLEU = \exp \left({\sum _{n=1}^{ng} w_nlogpr_{n}}\right) \cdot BP \end{equation}

where $pr_{n}$ is n-gram precision, and BP is Brevity Penalty, refer equation (2)

(2) \begin{equation} BP= \begin{cases} 1, & \text{if}\ len(\hat{y}) \gt len(y) \\[5pt] \exp \Bigl (1 - \frac{len(y)}{len(\hat{y})}\Bigl ), & \text{if}\ len(\hat{y}) \leq len(y) \end{cases} \end{equation}

where $\hat{y}$ represents the predicted translated sentence while $y$ is the gold standard alignment sentence. Table 2 reports the accomplished BLEU score for all three models. The transformer model outperforms the other models with a BLEU score of 51.74 for en-hi translation and 41.07 for en-mr translation on the test set. This score is considered a good BLEU score and can be used for en-hi and en-mr translation. The transformer model effectively produces accurate and fluent translations for the en-hi and en-mr language pairs.

Table 2.

BLEU score of en-hi and en-mr translation

The trained NMT model is utilized to translate the entire SQUAD dataset, with each tuple being translated into Hindi and Marathi languages. During this translation process, it is observed that for certain tuples, the generated translated answer does not appear in the translated context. This occurrence is particularly noticeable for verbs, abbreviations (such as US, U.S., USA, America for the United States of America), numbers (specifically for the Marathi language), and other similar cases. This situation arises because the NMT model translates the same word differently based on the context. For example, if the translated context contains information about the “Bank of River," indicating that the context is about the river, but the translated answer outputs “Financial Institution Bank," which is unrelated to the context, then the answer is considered inaccurate. In such cases, the tuple consisting of the translated question, context, and answer will not be considered for training the model. To address this problem, retrieving the correct answer is necessary even when its translation is not explicitly present in the context. The process of retrieving the correct answer is discussed in Section 3.1.2.

3.1.2 Retrieval of correct answer

The retrieval of accurate answers is performed to obtain the complete translated version of the SQuAD dataset in both languages for cases where the translated answer ( $a_{hi}$ and $a_{mr}$ ) is not present in the translated context. To ensure the retrieval of correct answers, the best strategy involves utilizing Hindi and Marathi WordNet (Bhattacharyya Reference Bhattacharyya2017), which serves as a thesaurus available for various languages. WordNet is used to acquire different synsets (i.e., sets of synonymous words) for a given word (lemma). It aids in replacing the existing translated answer ( $a_{hi/mr}$ ) with an appropriate alternative found within the translated context ( $c_{hi/mr}$ ).

In this strategy, the word in the translated answer ( $a_{hi/mr}$ ) is searched within WordNet to obtain its various senses. The process provides a pool of searched senses and their matches for a word in the translated context ( $c_{hi/mr}$ ). The translated answer is replaced with the corresponding word from the translated context if a match is found. This ensures that the retrieved answer accurately reflects the information present within the context. By applying this approach, the translated NMT model can effectively retrieve the correct answers, enhancing the overall accuracy and reliability of the translated SQuAD dataset.

The two subtasks described above have resulted in the creation of two versions of the EHMQuAD dataset. The first version consists of translating the original SQuAD dataset with sentence alignment. This version includes 48,769 translated samples in Hindi and 42,192 translated samples in Marathi. On the other hand, the second version incorporates the translation of the dataset with sentence alignment, along with retrieving the correct answer (Section 3.1.2). This version comprises 63,298 translated samples in Hindi and 51,359 translated samples in Marathi. The second version demonstrates an improvement in the number of tuples available in the EHMQuAD dataset. Table 3 presents the statistical details of both versions of the dataset, which were derived from the original 87,599 samples of the SQuADv1.1 dataset.

Table 3.

Statistics of EHMQuAD dataset over SQuADv1.1

The proposed work utilizes the second version of the EHMQuAD dataset to develop an MQA framework, leveraging the enhanced quality and accuracy achieved through answer retrieval. Figure 2 provides a visual representation of a sample from the EHMQuAD dataset, showcasing the structure and format of the translated tuples. In the second version of the dataset, the start position ( $a^{begin}{hi/mr}$ ) and end position ( $a^{end}{hi/mr}$ ) of the translated answer are determined by searching for the answer span within the translated context. These positions provide the necessary information to accurately locate the corresponding answer within the translated context. A comprehensive and enriched multilingual question-answering dataset is constructed through these steps, facilitating the development and evaluation of MQA models for Hindi and Marathi languages.

Figure 2.

An example of the EHMQuAD dataset.

3.3.1 Question & Context encoding layer

Let us consider the English question $q_{en} = \big\{qw^{q_{en}}_1, qw^{q_{en}}_2,\ldots, qw^{q_{en}}_{x_{en}}\big\}$ and context $c_{en} = \big\{cw^{c_{en}}_1, cw^{c_{en}}_2,\ldots, cw^{c_{en}}_{y_{en}}\big\}$ , Hindi question $q_{hi} = \big\{qw^{q_{hi}}_t\big\}^{x_{hi}}_{t=1}$ and context $c_{hi} = \big\{cw^{c_{hi}}_t\big\}^{y_{hi}}_{t=1}$ , and Marathi question $q_{mr} = \big\{qw^{q_{mr}}_t\big\}^{x_{mr}}_{t=1}$ and context $c_{mr} = \big\{cw^{c_{mr}}_t\big\}^{y_{mr}}_{t=1}$ . This layer is responsible for obtaining the encoding of the questions and contexts. From the given question and context, word embedding $\big(\big\{e^{q_{en}}_t\big\}^{x_{en}}_{t=1}$ , $\big\{e^{c_{en}}_t\big\}^{y_{en}}_{t=1}$ , $\big\{e^{q_h}_t\big\}^{x_h}_{t=1}$ , $\big\{e^{c_h}_t\big\}^{y_h}_{t=1}$ , $\big\{e^{q_{mr}}_t\big\}^{x_{mr}}_{t=1}$ , and $\big\{e^{c_{mr}}_t\big\}^{y_{mr}}_{t=1}\big)$ are generated from the pre-trained word embedding model. According to (Lee, Cho, and Hofmann Reference Lee, Cho and Hofmann2017; Xia, Ding, and Liu Reference Xia, Ding and Liu2020), character embedding is an extension of word embedding as it deals with OOV words. Hence, character-level embedding $\big(\big\{c^{q_{en}}_t\big\}^{x_{en}}_{t=1}$ , $\big\{c^{c_{en}}_t\big\}^{y_{en}}_{t=1}$ , $\big\{c^{q_{hi}}_t\big\}^{x_{hi}}_{t=1}$ , $\big\{c^{c_{hi}}_t\big\}^{y_{hi}}_{t=1}$ , $\big\{c^{q_{mr}}_t\big\}^{x_{mr}}_{t=1}$ , and $\big\{c^{c_{mr}}_t\big\}^{y_{mr}}_{t=1}\big)$ is also employed in this work. These character-level embeddings are produced as the final output of a bi-directional gated recurrent unit (Bi-GRU), employed for the token’s character embeddings. The word embeddings and character-level embeddings are concatenated to enhance the word representation.

\begin{equation*} E_i = concat \left(e^{q_k}_t,c^{q_k}_t\right) \end{equation*}

For each question and context, the final word representations $u^{q_k}_t$ and $u^{c_k}_t$ are computed using Bi-GRU as follows:

(3) \begin{equation} \begin{aligned} u^{q_k}_t = Bi\text{-}GRU \left(u^{q_k}_{t-1},[E_i]\right)\\[4pt] u^{c_k}_t = Bi\text{-}GRU \left(u^{c_k}_{t-1},[E_i]\right) \end{aligned} \end{equation}

where $k \in \{en, hi, mr\}$ denotes the English(en), Hindi(hi), and Marathi(mr) languages. Bi-GRU is chosen over LSTM due to its computational efficiency.

3.3.2 Shared encoding layer

The shared representation of the question and context is obtained separately for each language from the encoding of the previous layer. This shared representation improves the representation of question–question and context–context relationships. Since the dataset contains the same information for every question and context, this shared representation is achieved through soft alignment between words. A single separate representation of question and context is formed by combining the representation of individual languages. $\big(R^{q_{en}}_t, \text{ } R^{q_{hi}}_t \text{and }R^{q_{mr}}_t\big)$ and $\big(R^{c_{en}}_t, \text{ } R^{c_{hi}}_t \text{and } R^{c_{mr}}_t\big)$ are the English, Hindi, and Marathi question and context representations, with the awareness of other two languages. The English question representation, taking into account the Hindi and Marathi question representations, is computed using a Bi-GRU as shown in equation (4):

(4) \begin{equation} R^{q_{en}}_t = Bi\text{-}GRU\left(R^{q_{en}}_{t-1}, v^{q_{hi}}_t, v^{q_{mr}}_t\right) \end{equation}

(5) \begin{align} l^t_j & = N^T tanh\left(\left[W^{q_{hi}}_u, u^{q_{hi}}_j\right] \left[W^{q_{mr}}_u, u^{q_{mr}}_j\right]\left[W^{q_{en}}_u, u^{q_{en}}_t\right]\left[W^{q_{en}}_r, R^{q_{en}}_{t-1}\right]\right)\nonumber\\[5pt] v^{q_{hi}}_t & = \sum _{i=1}^{x_{hi}}\left(\frac{exp\big(l^t_i\big)}{\sum _{j=1}^{x_{hi}} exp\big(l^t_j\big)}\right)\cdot u^{q_{hi}}_i\nonumber\\[5pt] v^{q_{mr}}_t & = \sum _{i=1}^{x_{mr}}\left(\frac{exp\big(l^t_i\big)}{\sum _{j=1}^{x_{mr}} exp\big(l^t_j\big)}\right)\cdot u^{q_{mr}}_i \end{align}

where $N^T$ is a weight vector, $W^{q_{en}}_u, W^{q_{hi}}_u, W^{q_m}_u, W^{q_{en}}_r$ are the learnable weight matrices, and $\lt v^{q_{hi}}_t$ & $v^{q_{mr}}_t\gt$ are attention-based pooling vectors. The English question representation $\big(R^{q_{en}}_t\big)$ is computed using Bi-GRU by concatenating the pooling vector $v^{q_{hi}}_t$ & $v^{q_{mr}}_t$ with the previous (t-1) representation of $\big(R^{q_{en}}_t\big)$ . Equation (5) computes the pooling vector $v^{q_{hi}}_t$ & $v^{q_{mr}}_t$ of English question representation $u^{q_{hi}}_t$ & $u^{q_{mr}}_t$ at time t. Likewise, the other two question representations, i.e., $R^{q_{hi}}_t$ and $R^{q_{mr}}_t$ are computed in the similar manner. In the end, the single shared question representation is computed by combining the individual question representations from all three languages as shown in equation (6):

(6) \begin{equation} \big\{R^{q}_t\big\}^{(x_{en}+x_{hi}+x_{mr})}_{t=1} = \big\{R^{q_{en}}_t\big\}^{x_{en}}_{t=1} \oplus \big\{R^{q_{hi}}_t\big\}^{x_{hi}}_{t=1} \oplus \big\{R^{q_{mr}}_t\big\}^{x_{mr}}_{t=1} \end{equation}

Similarly, the shared context representation is computed by concatenating the context representations from all three languages, as indicated in equation (7):

(7) \begin{equation} \big\{R^{c}_t\big\}^{(y_{en}+y_{hi}+y_{mr})}_{t=1} = \big\{R^{c_{en}}_t\big\}^{y_{en}}_{t=1} \oplus \big\{R^{c_{hi}}_t\big\}^{y_{hi}}_{t=1} \oplus \big\{R^{c_{mr}}_t\big\}^{y_{mr}}_{t=1} \end{equation}

In equations (6) and (7), $\oplus$ denotes the concatenation operation, resulting in a unified representation encompassing information from all three languages.

3.3.3 Question-context attention layer

The question-aware-context representation is computed in this layer using an attention-based RNN mechanism. This approach captures the relevance and significance of context information given a specific question. The question-aware-context representation is computed separately for all three languages, considering the shared question representation and shared context representation. Since the shared question representation only includes a limited number of words, it provides a condensed representation. On the other hand, the shared context representation incorporates a larger number of words, as the context vocabulary is generally more extensive than that of the corresponding question. This facilitates a more comprehensive and enriched question-aware-context representation. The English question-aware-context representation is denoted by equation (8):

(8) \begin{equation} S^{c_{en}}_t = Bi\text{-}GRU\left(S^{c_{en}}_{t-1}, z^{c_{en}}_t\right) \end{equation}

(9) \begin{align} l^t_j & = N^T tanh \left( \left[W^{q}_r, R^{q}_j \right] \left[W^{c}_r, R^{c}_j \right] \left[W^{c_{en}}_u, u^{c_{en}}_t \right] \left[W^{c_{en}}_s, S^{q_{en}}_{t-1} \right] \right)\nonumber\\[5pt] z^{c_{en}}_t & = \sum _{i=1}^{x_{en}+x_{hi}+x_{mr}} \left(\frac{exp\big(l^t_i\big)}{\sum _{j=1}^{x_{en}+x_{hi}+x_{mr}} exp\big(l^t_j\big)} \right)\cdot R^{q}_i + \sum _{i=1}^{x_{en}+x_{hi}+x_{mr}} \left(\frac{exp\big(l^t_i\big)}{\sum _{j=1}^{x_{en}+x_{hi}+x_{mr}} exp\big(l^t_j\big)} \right)\cdot R^{c}_i \end{align}

where $W^{q}_r, W^{c}_r, W^{c_{en}}_u, W^{c_{en}}_s$ are the learnable weight matrices and $z^{c_{en}}_t$ is an attention-based pooling vector (equation (9)). Likewise, the Hindi question-aware-context representation $\big(S^{c_{hi}}_t\big)$ and Marathi question-aware-context representation $\big(S^{c_{mr}}_t\big)$ are derived by applying the same principles as the English representation.

3.3.4 Context-context attention layer

In the previous layer, the question-aware-context representation highlights the important features of the context, but the problem with this representation is that it handles very limited information. Hence, one potential answer fails to overlook the significant information in the context beyond its window. This approach is designed to address lexical or syntactical differences between the question–answer pairs in the SQuAD and EHMQuAD datasets. By dynamically matching the question-aware-context representation with the original context, the model can capture the relevant information from the context and encode it appropriately with the corresponding question, resulting in a more comprehensive and accurate context representation. For the English context, the final context representation is computed using the following equation (10):

(10) \begin{equation} B^{c_{en}}_t = Bi\text{-}GRU\left(B^{c_{en}}_{t-1}, S^{c_{en}}_{t}, z^{c_{en}}_t\right) \end{equation}

(11) \begin{align} l^t_j & = N^T tanh\left(\left[W^{c_{en}}_{b'},W^{c_{en}}_{b''}\right]\left[S^{c_{en}}_j, S^{c_{en}}_{t}\right]^T\right)\nonumber\\[5pt] z^{c_{en}}_t & = \sum _{i=1}^{y_{en}}\left(\frac{exp\big(l^t_i\big)}{\sum _{j=1}^{y_{en}} exp\big(l^t_j\big)}\right)\cdot S^{c_{en}}_i \end{align}

where $W^{c_{en}}_{b'}$ and $W^{c_{en}}_{b''}$ are the learnable weight matrices. Similarly, the final context representation of Hindi $\big(\big\{B^{c_{hi}}_t\big\}^{y_{hi}}_{t=1}\big)$ and Marathi $\big(\big\{B^{c_{mr}}_t\big\}^{y_{mr}}_{t=1}\big)$ is computed in the same fashion by equations (10) and (11).

3.3.5 Answer layer

A pointer network is employed to handle the multilingual nature of the MQA framework and generate answers in multiple languages. The pointer network is based on the seq-to-seq pointer network (Chen et al. Reference Chen, Zhang, Hu and Huang2021), which is responsible for extracting the start position $\big(a^{begin}\big)$ and end position $\big(a^{end}\big)$ of the answer from either the English, Hindi, or Marathi context. Given the context in any language as input, the answer layer’s objective is to determine the appropriate start and end positions within the context that correspond to the answer. This is achieved by computing the following equation (12):

(12) \begin{align} l^t_j & = N^T tanh \left( \left[W^{c_{en}}_{p},B^{c_{en}}_j \right] \left[W^{a_{en}}_{h},h^{a_{en}}_{t-1} \right] \right)\nonumber\\[4pt] a^t_i & = \left(\frac{exp\big(l^t_i\big)}{\sum _{j=1}^{y_{en}} exp\big(l^t_j\big)} \right) \end{align}

where $h^{a_{en}}_{t-1}$ is the last hidden state of the answer layer. The attention pooling vector $a^t_i$ depends on the recently predicted probability $a^t$ , which serves as an input to this layer. The hidden state of the answer layer is computed by equation (13):

(13) \begin{align} h^{a_{en}}_{t} & = Bi\text{-}GRU \left(h^{a_{en}}_{t-1}, z^{c_{en}}_t\right)\nonumber\\[4pt] z^{c_{en}}_t & = \sum _{i=1}^{y_{en}}\left( a^t_i B^{c_{en}}_i\right) \end{align}

The answer for the English language is obtained from the answer network by predicting the start position at time step $t=1$ , and subsequently predicting the end position at the next time step. This prediction is based on the probabilities assigned to each position in English. The prediction of the answer is determined by selecting the maximum probability among the predicted positions.

(14) \begin{equation} a^t_{en} = argmax\left(a^t_{1}, a^t_{2},\ldots, a^t_{y_{en}}\right) \end{equation}

For Hindi and Marathi languages, the answer’s start and end positions are also predicted from their corresponding context using equation (14). This approach allows the answer network to identify the most likely start and end positions of the answer for each language, facilitating the generation of language-specific answers tailored to the corresponding context.

4.1.1 MMQA

MMQAFootnote c dataset (Gupta et al. Reference Gupta, Kumari, Ekbal and Bhattacharyya2018) is an MQA dataset for English and Hindi languages. This dataset has comparable corpora of 500 documents for both English and Hindi languages but not the parallel corpora along with the sample of $\lt q, c, a\gt$ . Gupta et al. (Reference Gupta, Ekbal and Bhattacharyya2019) proposed a novel algorithm that uses the MMQA dataset of 5,495 QA pairs to evaluate and create a snippet for each pair. These selected QA pairs and their corresponding created snippets are used to evaluate the EHMMQA model. This dataset is converted into SQuAD-like sample instances.

4.1.2 Translated SQuAD

The Translated SQuADFootnote d dataset (Gupta et al. Reference Gupta, Ekbal and Bhattacharyya2019) is derived from the original SQuAD dataset and involves the translation (using Google Translate API) of 18,454 samples from English to the Hindi language. For each context available in the original SQuAD dataset, a sample consisting of the question ( $q$ ), context ( $c$ ), and answer ( $a$ ) is selected and transformed into the format of $\lt q,c,a,a_{start}, a_{end}\gt$ for both English and Hindi languages. This ensures that the translated dataset covers all the contexts present in the SQuAD dataset while providing bilingual information, which undergoes a preprocessing step to address any discrepancies that may arise from the direct translation between languages. These discrepancies are likely due to the inherent syntax, grammar, and language structure differences between English and Hindi. The preprocessing step helps mitigate these issues, ensuring the dataset’s quality and reliability for training and evaluation purposes.

4.1.3 XQuAD

DeepMind has developed a Cross-lingual Question Answering Dataset (XQuAD)Footnote e (Artetxe et al. Reference Artetxe, Ruder and Yogatama2020) for MQA tasks. XQuAD covers 12 languages, including Hindi, and is a reliable benchmark for evaluating MQA models. The dataset consists of 1,190 question–answer pairs for each of the 12 languages. These pairs are derived from 240 contexts originally found in the SQuAD dataset. XQuAD is particularly valuable for evaluating English-Hindi Machine Reading Comprehension (MRC) tasks since it has been translated by linguistic experts, making this dataset more reliable for evaluating MQA models. XQuAD is a monolingual variant of the SQuAD dataset allowing, it to be utilized for evaluating the monolingual settings of the EHMMQA model in monolingual settings, such as English-English ( $Q_{en}-C_{en}$ ) and Hindi-Hindi ( $Q_{hi}-C_{hi}$ ) scenarios.

4.1.4 MLQA

MLQAFootnote f (Lewis et al. Reference Lewis, Oguz, Rinott, Riedel and Schwenk2020) is an MQA dataset containing QA pairs along with the context in seven languages, including Hindi. MLQA is a comprehensive and widely used benchmark for evaluating multilingual and cross-lingual QA models. This dataset contains 5,425 QA pairs and follows a structure similar to SQuAD, ensuring familiarity and compatibility with existing MQA models. The inclusion of Hindi in MLQA enables the specific evaluation of MQA models within the context of this language. By leveraging MLQA, researchers can conduct rigorous and reliable evaluations of multilingual and cross-lingual QA models, facilitating advancements in the field.