1. Introduction

In recent years, question-answering systems (QAS) have experienced rapid growth in multiple sectors, with tourism being one of the most prominent applications. In the tourism industry, QAS are widely used in smart travel guides, online customer service, itinerary planning, and more. The nature of the queries in tourism often involves vague structures, varied phrasing, and heavy reliance on external knowledge, which raises the bar for semantic understanding and response accuracy. The introduction of machine learning (ML), particularly deep learning and pre-trained language models (PLMs), has significantly enhanced the semantic modeling capabilities of QAS. As noted by Núñez et al. in their study, machine learning methods have been systematically reviewed and summarized for their applications in tourism, offering valuable technical frameworks for current question-answering systems [1]. From traditional intent classification models to more advanced BERT-based semantic matching frameworks, numerous technological paths have been explored in various question-answering scenarios. However, in the complex context of tourism, choosing, adapting, and optimizing ML models remains underexplored, and few systematic studies have been conducted on how different techniques can be applied effectively to the unique challenges posed by the tourism domain.

This paper aims to fill this gap by providing a comprehensive review and comparative analysis of ML methods in QAS, with a special focus on the tourism domain. It discusses the performance and domain adaptability of various models, highlights existing challenges, and proposes methods to improve accuracy in tourism QAS. By drawing from user demand perspectives and technological advancements, this study offers insights into optimizing QAS for tourism applications and beyond.

2. Question-answering systems and machine learning methods

2.1. Composition and evolution of question-answering systems

Question-answering systems (QAS) are designed to respond to user queries by providing relevant and accurate answers, often drawn from a vast repository of knowledge. These systems have evolved significantly over the years, from rule-based systems to the more sophisticated, machine-learning-based models that we see today. Traditional QAS were primarily built on keyword matching, where the system would search for exact matches to the user's query in a database of pre-stored answers. This approach, however, was limited by its inability to handle complex sentence structures, paraphrasing, or vague queries.

With the advancement of machine learning, particularly deep learning, QAS has become more capable of understanding the nuances of natural language. Modern QAS utilize algorithms like decision trees, support vector machines (SVM), and neural networks to improve their performance. Neural networks, especially those based on deep learning, have been able to model the intricate relationships between words in a sentence, allowing for more accurate and contextually relevant responses.

2.2. Machine learning techniques in question-answering systems

Machine learning techniques have brought significant improvements to QAS by allowing systems to learn from data rather than relying on static rule-based logic. One of the primary advantages of ML-based QAS is their ability to handle dynamic, varied phrasing and complex sentence structures, which are common in natural language. Algorithms like deep neural networks, convolutional neural networks (CNN), and recurrent neural networks (RNN) have revolutionized QAS by enabling semantic understanding at deeper levels. Pre-trained language models (PLMs), such as BERT, GPT, and T5, have particularly shown impressive results in tasks like semantic matching and context understanding, thereby improving the overall response accuracy in systems that require nuanced understanding.

2.3. Mainstream techniques classification and research framework

In QAS development, the core techniques typically fall into categories like intent recognition, entity extraction, and response generation. Techniques such as decision trees, Naive Bayes, and deep learning architectures like LSTMs, CNNs, and transformers are often classified based on their application. This paper explores the use of BERT-based models in semantic understanding, as well as the application of various optimization methods for better contextual responses. Understanding the evolution and classification of these techniques allows us to assess the best-suited approaches for the tourism domain, considering the unique requirements and challenges of this field.

3. Survey of machine learning methods in various question-answering system modules

3.1. Intent recognition model development and comparison

Intent recognition is a core component of any question-answering system, as it helps the system understand the underlying purpose of the user's query. Traditionally, intent recognition was carried out using rule-based methods or simple machine learning classifiers like decision trees or SVMs. However, with the advent of deep learning, more sophisticated approaches such as Recurrent Neural Networks (RNNs) and Transformer-based models like BERT have emerged as the gold standard.

Recent studies have shown that BERT-based models significantly outperform traditional models in intent classification. These models are capable of understanding the broader context of a sentence and are highly effective in domains where the phrasing of questions can vary widely, such as in tourism, where users may ask about similar information in different ways.

3.2. Semantic matching and question representation methods

Semantic matching in QAS is crucial for understanding the relationship between the user's query and the relevant information in the knowledge base. Traditional methods rely on keyword-based matching, like fine-grained keyword extraction, which allows question-answering systems to more accurately match the semantics of user queries [2], while modern approaches use embedding-based models such as word2vec, GloVe, and BERT, which are capable of capturing the semantic meanings of words in context. The combination of transfer learning and BERT significantly improves the accuracy of question-answering systems, particularly when handling complex semantics [3]. And the integration of spatial and textual information in a dual-encoder model significantly improves semantic understanding in tourism question-answering systems [4]. These models represent words as vectors in a continuous vector space, where semantically similar words are closer to each other. This allows QAS to handle paraphrased questions or those with different wordings but similar meanings, which is particularly important in the tourism domain, where users may phrase their queries in many different ways. Or we can use a semantic category classification method to significantly enhance the accuracy of question-answering systems [5], this is also a good method.

3.3. Context modeling and multi-turn dialogue systems

Context modeling is essential for QAS that handle complex, multi-turn interactions; deep context modeling greatly improves response selection accuracy in multi-turn dialogue systems [6]. Deep learning techniques and models such as RNNs and BERT have been successfully applied to this task, enabling systems to remember and incorporate previous user queries into the response generation process. The improved BERT model can better understand local context in multi-turn dialogue, enhancing response selection accuracy [7]. This is particularly useful in tourism-related scenarios, where users may ask follow-up questions or need answers that involve multiple steps, like booking a flight and hotel at the same time.

3.4. Pre-trained models in question-answering

Pre-trained models like BERT, GPT, and T5 have revolutionized the field of QAS by providing robust models capable of understanding language context at scale. These models are fine-tuned on domain-specific data to improve performance in specialized areas, such as tourism. By leveraging large-scale language understanding, pre-trained models offer substantial improvements in accuracy and fluency in responses.

4. Current research on question-answering systems in the tourism domain

4.1. Typical tourism question-answering scenarios and user needs

Tourism question-answering systems serve a variety of user needs, which can range from simple informational queries to more complex personalized recommendations. In typical scenarios, users may ask questions about destinations, transportation, local events, or accommodations. The diversity in phrasing and the need for context-based answers make the tourism domain particularly challenging for QAS.

4.2. Current tourism question-answering systems and dataset overview

Tourism question-answering systems (QAS) have seen significant advancements in recent years, with the integration of machine learning (ML) techniques and natural language processing (NLP). NLP techniques are rapidly expanding in tourism research, particularly in question-answering systems [8]. These systems are now crucial components of smart tourism, assisting travelers in accessing a wide variety of information ranging from local attractions to transportation options. Unlike traditional systems that were based on rule-based frameworks or static databases, modern tourism QAS leverage machine learning models to handle more dynamic, context-sensitive interactions. The current landscape of tourism QAS relies heavily on both structured and unstructured data sources. Tourism-specific datasets, such as TourismQA and others, have been developed to aid the training and fine-tuning of models for the domain. These datasets often contain diverse queries related to travel destinations, hotel bookings, and tourist attractions, along with their corresponding responses. However, the limitations in dataset diversity and quality remain significant challenges. While some datasets are comprehensive and still being improved, such as some datasets recently using the integration of geospatial data, the MapQA system is enabled to handle complex open-domain geospatial queries [9]. Many lack the nuances needed for realistic user queries, particularly in multi-turn dialogues.

Recent advancements in using external knowledge sources, such as geographic information systems (GIS) or knowledge graphs, have contributed significantly to improving the depth of answers in tourism QAS. These external sources allow for the integration of real-time information, such as weather conditions or transport schedules, offering users more dynamic and accurate responses.

4.3. Linguistic characteristics of tourism and their impact on models

The tourism domain poses unique linguistic challenges, primarily due to the wide variation in question phrasing and the use of local dialects and multilingual content. Additionally, tourism-related queries often contain implicit contextual information, such as time, location, and user preferences, which may not be explicitly stated. For instance, a query like “What’s the best time to visit Paris for sightseeing?” implicitly refers to the user's interests and time of visit, making it more complex than a typical factual query.

Machine learning models that are trained on generic language data often struggle to fully understand and accurately interpret these subtle contextual cues. As a result, domain-specific models that can capture these nuances are essential. Techniques like transfer learning, where pre-trained models are adapted to the tourism domain using specific datasets, are gaining popularity to tackle these issues effectively.

5. Applicability of existing methods in the tourism domain

5.1. Model adaptability in tourism contexts

While many machine learning methods have been proven effective in general-purpose QAS, their adaptability to the tourism domain varies. For instance, traditional models based on statistical methods often struggle with the dynamic and unstructured nature of tourism queries, such as those involving personal preferences or real-time information like “What are the best restaurants near me right now?” In such cases, a system’s ability to understand context and location-specific data becomes critical for providing relevant answers.

On the other hand, deep learning-based models, particularly pre-trained models like BERT, have shown better performance in handling the complex and varied language used in tourism. These models are capable of understanding subtle nuances and providing answers that are contextually relevant to the specific needs of the user. However, the performance of these models can vary based on the domain-specific data they are trained on. Fine-tuning models with tourism-related data has been shown to improve their accuracy and make them more suitable for handling tourism-specific queries.

5.2. Multilingual expression and cross-language question-answering challenges

Multilingual support remains one of the most significant challenges for tourism QAS, as the tourism industry caters to a global audience. A tourism QAS should ideally be capable of handling queries in multiple languages, which requires sophisticated language modeling techniques. The ability of machine learning models to understand and respond in various languages is crucial for enhancing the user experience in international contexts. One of the prominent challenges lies in training these models to maintain accuracy across different linguistic structures, idioms, and cultural contexts.

Recent advancements in multilingual pre-trained models such as mBERT (multilingual BERT) have shown promise in bridging this gap. These models, when fine-tuned with multilingual tourism data, can enhance the robustness of QAS in handling queries from a global user base.

5.3. Feasibility of model transfer and fine-tuning strategies

Transfer learning and fine-tuning are key strategies for adapting pre-trained models to specific domains like tourism. By fine-tuning a general-purpose pre-trained model on tourism-related data, the model can be adapted to the domain’s unique challenges. This approach leverages the knowledge learned from a broad range of language data and transfers it to a more specialized domain, thus improving both performance and accuracy.

The feasibility of fine-tuning strategies has been widely studied, with evidence suggesting that they significantly enhance the performance of QAS in specific domains. Fine-tuned models tend to outperform their general-purpose counterparts in both semantic matching and context understanding, making them ideal for tourism QAS.

6. Research problems and future optimization paths

6.1. Major limitations of current methods in tourism QAS

Despite significant advancements, several challenges remain in applying machine learning methods to tourism QAS. One major limitation is the ability of current systems to handle the variety and ambiguity of user queries. For example, a tourist might ask, "Is it worth visiting Paris in December?" This type of subjective query, which depends on personal preferences and context, is difficult for many QAS to handle accurately. Many existing models rely on structured data and struggle with open-ended, subjective queries that require contextual understanding.

Another limitation is the lack of sufficient, high-quality domain-specific datasets. While datasets like TourismQA have been created, they remain relatively limited compared to the vast range of possible tourism-related queries. The lack of diverse, high-quality training data affects the accuracy and generalizability of QAS in real-world scenarios.

6.2. Integration of external knowledge and geospatial information

Future advancements in tourism QAS could significantly benefit from the integration of external knowledge sources, such as knowledge graphs and geospatial data. Knowledge graphs can provide additional context about destinations, accommodations, and attractions, enhancing the system’s ability to deliver more accurate and informative responses. Furthermore, geospatial data can help answer location-based queries, such as the nearest restaurants, or guide users based on their current location.

For example, combining QAS with a geospatial system could allow tourists to ask dynamic, real-time questions like "What’s the closest café to the Eiffel Tower right now?" and receive highly personalized responses based on both location and time of day.

6.3. Multimodal, multilingual, and personalized question-answering for the future

As tourism QAS continues to evolve, the development of multimodal systems is likely to play an increasingly important role. Multimodal systems can integrate text, images, and voice, allowing tourists to interact with QAS in more natural and intuitive ways. For instance, a tourist might want to upload a photo of a landmark and ask for information about it, which requires combining image recognition with text-based semantic understanding.

Furthermore, personalized QAS that can adapt to individual user preferences, travel history, and past interactions hold significant potential. Today’s large language models provide more precise analysis for personalized recommendations in tourism [10], but they still need to be continuously improved. Such systems can offer more tailored recommendations and improve user satisfaction by learning from ongoing interactions. The combination of multimodal inputs, multilingual capabilities, and personalization will significantly enhance the tourism QAS experience in the coming years.

7. Conclusion

This paper has provided a comprehensive review of the role of machine learning in enhancing the accuracy of question-answering systems, with a particular focus on the tourism domain. It has discussed the development of QAS, from rule-based systems to modern machine learning-based models, and highlighted the advantages of pre-trained models like BERT in improving semantic understanding and response generation. The research also explored the unique challenges faced by tourism QAS, such as handling ambiguous queries, multilingualism, and domain-specific knowledge.

Despite significant progress, several limitations remain in the application of machine learning methods to tourism QAS, including the need for better domain-specific data and improved handling of subjective queries. Future research could focus on integrating external knowledge sources, such as knowledge graphs and geospatial data, to improve the accuracy and personalization of responses. Additionally, the development of multimodal and multilingual systems will be key to addressing the diverse needs of tourists around the world.

By enhancing the adaptability and accuracy of QAS through machine learning, the tourism industry can offer more personalized, efficient, and dynamic services to travelers, ultimately improving the overall user experience in the age of smart tourism.