Intelligent question answering system is an innovative information service system which integrates natural language processing, information retrieval, semantic analysis and artificial intelligence. The system mainly consists of three core parts, which are question analysis, information retrieval and answer extraction. Through these three parts, the system can provide users with accurate, fast and convenient answering services.

The representative systems of the intelligent question answering system include:

(1) Rule-based algorithms (1960s-1980s). The question-answering system based on this pattern mainly relies on writing a lot of rules and logic to implement the dialogue. ELIZA [1], developed by Joseph Weizenbaum in the 1960s, was the first chatbot designed to simulate a conversation between a psychotherapist and a patient. PARRY [2] is a question-and-answer system developed in the 1970s that simulates psychopaths. The emergence of ELIZA and PARRY provided diverse design ideas and application scenarios for subsequent intelligent question answering systems, thereby promoting the diversification and complexity of dialogue systems. However, the main problem of this model is its lack of flexibility and extensibility. It relies too much on rules or templates set by humans, and consumes a lot of time and manpower. When the questions become complicated, it is difficult to get satisfactory answers through simple rules set by the model.

(2) Statistics-based algorithms (1990s-2000s). The question-answering system based on this model adopts the method of statistical learning to learn patterns and rules from a large number of dialogue data. Common algorithms include Vector Space Model [3] and Conditional Random Fields [4]. ALICE (Artificial Linguistic Internet Computer Entity) [5] is an open-source natural language processing project. The system in question is an open-domain question answering platform capable of addressing queries across a multitude of subjects and domains. Jabberwacky [6] is an early intelligent chatbot employing machine learning and conversational models to enhance its responses continually. These systems are designed to train models that can learn the relationships between questions and answers present in the corpus. Therefore, these models can carry out more natural and smooth dialogue. However, the ability of context understanding and generalization ability is weak, so it is difficult to adapt to model sharing and transfer learning in various professional fields. Moreover, considering statistical models are trained on a large corpus, this kind of model may suffer from data bias when dealing with domain-specific problems and fail to provide accurate answers.

(3) Algorithms based on hybrid technology (2010s-early 2020s). The question-answering system, grounded on this model, can amalgamate diverse techniques encompassing rules, statistics, and machine learning. It leverages multiple input modalities, including speech, image, and text, to interoperate seamlessly. The overarching objective is to facilitate users in accomplishing specific tasks or goals within designated domains, such as booking, traveling, shopping, or ordering food. This synergistic integration of multifarious technologies and input modes fosters a more sophisticated and intelligent dialogue system. Typical question answering systems based on hybrid technology model include Apple's Siri [7], Microsoft's Cortana [8], Amazon's Alexa [9], Facebook's M [10] and Google's Google Assistant [11]. These systems are centered around artificial intelligence and natural language processing technology, aiming to furnish users with personalized and convenient information and services to cater to diverse needs.

The system built based on this pattern has stronger context understanding and personalized customization, but there are two shortcomings: first, the quality of dialogue in such a system is not stable; Secondly, the generalization ability of the model is limited. It is difficult to realize model sharing, transfer learning and answer generation in professional fields. The training of this model requires excessive investment in computing and data resources, and its training and deployment speed is slow.

(4) Algorithms based on pre-trained language (2020s). The model is based on pre-trained language models such as BERT [12], GPT (Generative Pre-trained Transformer) [13], etc. These models are pre-trained on large-scale data and they learn rich language representation and context understanding skills to generate more natural, fluid, and accurate responses. In addition, through the supervised training on domain-specific question answering datasets, the question answering system can answer questions in specialized professional fields. [14] proposed a BERTserini algorithm which improves the exact match rate of the question answering system. In comparison to the original BERT algorithm, the proposed method surpasses its processing byte limit and can provide accurate answers for multi-document long texts.

Although systems built on the BERTserini algorithm perform well on public datasets, there are some problems in the application in professional fields such as electrical power engineering. Considering the low exact match rate and poor answer quality, engineering applications of these models are challenging. The problems are mainly caused by the following aspects.

(1) Lack of model expertise: Language models such as BERT or GPT are usually pre-trained from large amounts of generic corpus collected on the Internet. However, the digital realm offers limited professional resources pertaining to industries like electrical power engineering. As a result, the model has insufficient knowledge reserve when dealing with professional question, which affects the quality of the answers; (2) Differences in document format: There are significant differences between the format of documentation in the electrical power engineering field and that of public datasets. The documents in the electrical power engineering field often exhibit unique formatting, characterized by an abundance of hierarchical headings. It is easy to misinterpret the title as the main content and mistakenly use it as the answer to the question, leading to inaccurate results; (3) Different scenario requirements: Traditional answering systems do not need to pay attention to the source of answers in the original document. However, a system designed for professional use must provide specific source information for its answers. If such information is not provided, there may arise doubts regarding the accuracy of the response. This further diminishes the utility of the application in particular domains.

This paper proposes an improved BERTserini algorithm to construct an intelligent question answering system in the field of electrical power engineering. The proposed algorithm is divided into two stages:

 

 

The improved BERTserini algorithm proposed in this paper has three main advantages.

(1) The proposed algorithm implements multi-document long text preprocessing technology tailored for rules and regulations text. Through optimization, the algorithm segments rules and regulations into distinct paragraphs based on its inherent structure and supports answer output with reference to chapters and locations within the document. The effectiveness of this pretreatment technology is reflected in the following three aspects: First, through accurate segmentation, paragraphs that may include questions can be extracted more accurately, thus improving the accuracy of answer generation. Secondly, the original Bert model exhibits a limitation that it outputs the heading of rules and regulations text as the answer frequently. To address this issue, an improved BERTserini algorithm has been proposed. Finally, the algorithm is able to accurately give the location information of answers in the original document chapter. The algorithm enhances the comprehensiveness and accuracy of reading comprehension, generating answers to questions about knowledge and information contained in professional documents related to the field of electric power. Consequently, this leads to a marked improvement in answer quality and user experience for the question answering system.

(2) The proposed algorithm optimizes the training of the corpus in the field of electrical power engineering and fine-tunes the parameters of the large language model. This method eliminates the necessity for manual organization of professional question-answer pairs, knowledge base engineering, and manual template establishment in BERT reading comprehension, thereby effectively reducing labor costs. This enhancement significantly enhances the accuracy and efficiency of the question-answering system.

(3)（1）基于规则的算法（1960年代-1980年代）。基于这种模式的问答系统主要依靠编写大量的规则和逻辑来实现对话。ELIZA The[1] proposed由 algorithm has been developed for the purpose of enhancing question answering systems in engineeringJoseph Weizenbaum 在 1960 年代开发，是第一个旨在模拟心理治疗师和患者之间对话的聊天机器人。PARRY [2] applications.是 This algo1970 年代开发的一种模拟精神病患者的问答系统。ELIZA和PARRY的出现为后续的智能问答系统提供了多样化的设计思路和应用场景，从而促进了对话系统的多样化和复杂性。然而，该模型的主要问题是缺乏灵活性和可扩展性。它过于依赖人类设定的规则或模板，消耗大量的时间和人力。当问题变得复杂时，很难通过模型设定的简单规则获得满意的答案。

（2）基于统计的算法（1990年代至2000年代）。基于该模型的问答系统采用统计学习的方法，从大量对话数据中学习模式和规则。常见的算法包括向量空间模型[3]和条件随机场[4]。ALICE（Artithm exhibits a higher degree of exact match rate of questions and a faster response time for providing answers.

Freqficial Linguistic Internet Compuently Asked Questions (FAQ) er Entity）[5]是一个开源的自然语言处理项目。该系统是一个开放领域的问答平台，能够解决跨多个主题和领域的查询。Jabbere a collection ofwacky [6] common quest是一个早期的智能聊天机器人，它采用机器学习和对话模型来不断增强其响应。这些系统旨在训练可以学习语料库中问题和答案之间关系的模型。因此，这些模型可以进行更自然、更流畅的对话。然而，语境理解能力和泛化能力较弱，难以适应各专业领域的模型共享和迁移学习。此外，考虑到统计模型是在大型语料库上训练的，这种模型在处理特定领域的问题时可能会受到数据偏差的影响，无法提供准确的答案。

（3）基于混合技术的算法（2010年代-2020年代初）。基于此模型的问答系统可以融合各种技术，包括规则、统计和机器学习。它利用多种输入模式（包括语音、图像和文本）进行无缝互操作。总体目标是帮助用户在指定域内完成特定任务或目标，例如预订、旅行、购物或订购食物。这种多种技术和输入模式的协同整合促进了更复杂、更智能的对话系统。基于混合技术模型的典型问答系统包括苹果的Sions and answers designed to help users qui [7]、Mickly find answers to their questions. The key is to establish a rich osoft的Cortana [8]、亚马逊的Alexand [9]、Faccurate database book的M [10]和谷歌的Goof prgledefined questions, which consists of Assistant [11]。这些系统以人工智能和自然语言处理技术为核心，旨在为用户提供个性化和便捷的信息和服务，以满足多样化的需求。

（4）基于预训练语言的算法（2020年代）。该模型基于预训练语言模型，如BERT qu[12]、GPT（Gestions and their corresponding answers. They are manually curated from target documents. FAQs provide anerative Pre-trained Transformer）[13]等。这些模型在大规模数据上进行了预训练，它们学习了丰富的语言表示和上下文理解技能，以生成更自然、流畅和准确的响应。此外，通过对特定领域问答数据集的监督训练，问答系统可以回答专业专业领域的问题。文献[14]主要研究了农业问答系统。该系统利用人工智能技术和相关数据集，为农民提供有关天气、市场价格、植物保护和政府计划等主题的信息。文献[15]提出了一种基于BERT的TD-BERT模型。该模型利用BERT强大的语义表示能力，整合目标信息，提高情感分类的准确性。文献[16]提出了一种BERTswers corresponding to uini算法，该算法提高了问答系统的精确匹配率。与原始的BERT算法相比，所提方法超越了其处理字节限制，能够为多文档长文本提供准确的答案。

虽然基于BERTser questions by matchingni算法构建的系统在公共数据集上表现良好，但在电力工程等专业领域的应用中存在一些问题。考虑到精确匹配率低、答案质量差，这些模型的工程应用具有挑战性。这些问题主要由以下几个方面引起。

常见问题（FAQ）是常见问题和答案的集合，旨在帮助用户快速找到问题的答案[17]。关键是要建立一个丰富而准确的预设问题数据库，该数据库由问题和相应的答案组成。它们是从目标文档中手动整理的。FAQ them with the most similar questions.通过将用户的问题与最相似的问题进行匹配，提供与用户问题相对应的答案。

The Best Match 25 (BM25) algorithm, initially proposed by （BM25）算法[18,19]最初由Stephen Robertson and his team in 1994, is applied in the field of information retrieval. It is commonly used to calculate the relevance score between documents and queries. The main logic of 及其团队于1994年提出，并应用于信息检索领域。它通常用于计算文档和查询之间的相关性分数。BM25 is as follows: first, the query statement involves tokenization to generate morphemes的主要逻辑如下：首先，查询语句涉及分词生成语素; then, it calculates the relevance score between each morpheme and the search results.然后，计算每个语素与搜索结果之间的相关性分数。

Finally, by weighting and summing the relevance scores of morphemes with the search results, the relevance score between the retrieval query and the search result documents is obtained. The calculation formula of the 最后，通过将语素的相关性得分与搜索结果进行加权求和，得到检索查询与搜索结果文档之间的相关性得分。BM25 algorithm is as follows:

算法的计算公式如下：

In在此上下文中，Q 表示查询语句，𝑞_我表示从 thisQ context,获得的语素。对于中文来说，通过标记化查询 representsQ a query statement, represents a morpheme obtained from 得到的分割结果可以看作是语素𝑞_我.D For表示搜索结果文档。𝑊_我表示语素的权重𝑞_我和𝑅(𝑞_我，D)表示语素之间的相关性分数𝑞_我和文件D。重量参数有多种计算方法𝑊_我，逆文档频率 Chinese, the segmented results obtained from tokenizing query can be considered as morpheme . represents a search result document. represents the weight of morpheme , and represents the relevance score between morpheme and document . There are multiple calculation methods for weight parameter , with （Inverse Document Frequency (IDF) being one of the commonly used approaches. The calculation process for IDF is as follows:

DF） 是常用的方法之一。IDF 的计算过程如下：

In在等式中，N the equation, represents the total number of documents in the index, and represents the number of documents that contain 表示索引中的文档总数，并且𝑛(𝑞_我)表示包含以下内容的文档数𝑞_我.

最后，BM25算法的相关性评分公式可以总结如下：

Finally, the relevance scoring formula for the BM25 algorithm can be summarized as follows:

where and are adjustment factors, represents the frequency of morpheme appearing in document , denotes the length of document , and represents the average length of all documents.

哪里𝑘₁b为调整因子，𝑓(𝑞_我，D)表示语素的频率𝑞_我出现在文档 D 中， |D|表示文档 D 的长度，vgdl 表示所有文档的平均长度。

Anserini [1820] is an open-source information retrieval toolkit that supports various text-based information retrieval research and applications. The goal of 是一个开源的信息检索工具包，支持各种基于文本的信息检索研究和应用。Anserini is to provide an easy-to-use and high-performance toolkit that supports tasks such as full-text search, approximate search, ranking, and evaluation on large-scale text datasets. It enables the conversion of text datasets into searchable index files for efficient retrieval and querying. 的目标是提供一个易于使用且高性能的工具包，支持对大规模文本数据集进行全文搜索、近似搜索、排名和评估等任务。它支持将文本数据集转换为可搜索的索引文件，以便进行有效的检索和查询。Anserini incorporates a variety of commonly used text retrieval algorithms, including the BM25 algorithm. With 结合了多种常用的文本检索算法，包括 BM25 算法。使用 Anserini, it becomes effortless to construct a BM25-based text，可以毫不费力地构建基于 BM25 的文本检索系统，并对大规模文本集合进行有效的搜索和排名。该算法的流程图如图 retrieval1 system and perform efficient search and ranking on large-scale text collections. The flowchart of the algorithm is illustrated in Figure 1.所示。

Bidirectional Encoder Representations from Transformers (BERT) [12] is a pre-trained language model proposed by Google in 2018. The model structure is shown in Figure 2. In the model, 𝐸_𝑖 represents the encoding of words in the input sentence, which is composed of the sum of three word embedding features. The three word embedding features are Token Embedding, Position Embedding, and Segment Embedding. The integration of these three words embedding features allows the model to have a more comprehensive understanding of the text’s semantics, contextual relationships, and sequence information, thus enhancing the BERT model’s representational power. The transformer structure in the figure is represented as Trm. The 𝑇_𝑖 represents the word vector that corresponds to the trained word 𝐸_𝑖.

BERT exclusively employs the encoder component of the Transformer architecture. The encoder is primarily comprised of three key modules: Positional Encoding, Multi-Head Attention, and Feed-Forward Network. Input embeddings are utilized to represent the input data. Addition and normalization operations are denoted by “Add&norm”. The fundamental principle of the encoder is illustrated in Figure 3.

In recent years, several Chinese 近年来，在中文领域提出了几种中文BERT models have been proposed in the Chinese language domain. Among these, the chinese-模型。其中，哈大讯飞语言认知计算实验室（HFL）发布的中文BERT-wwm-ext model 模型[1921] released by the HIT·iFLYTEK Language C备受关注，具有代表性。该模型基于原始的Gognitive Computing Lab (HFL) has gained significant attention and serves as a representative example. This model was pre-trained on a corpus of Chinese encyclopedias, news articles, and question-and-answer texts, which contains a total of 5.4 billion words. The model uses the whole-word masking (wwm) strategy. ogle BERT模型，使用54亿个单词的总词汇进行了进一步的预训练，包括中文百科全书、新闻和问答数据集。该模型采用全词掩码（wwm）策略，这是针对汉语语言特点量身定制的改进。在中文处理中，由于单词是由字符组成的，而一个单词可能由一个或多个字符组成，因此有必要掩盖整个单词而不仅仅是单个字符。wwm 策略旨在更好地理解和捕捉汉语词汇的语义。综上所述，该模型是BERT的改进版中文版，通过全词屏蔽，在中文理解方面表现出更高的性能。

The architecture of BERTserini algorithm算法 [1416] is的架构如图 depicted4 in所示。该算法将 Figure 4. The algorithm employs the Anserini information extraction algorithm in conjunction with a pre-trained BERT model. In this algorithm, the 信息提取算法与预训练 BERT 模型结合使用。在该算法中，Anserini retriever is responsible for selecting text paragraphs containing the answer, which are then passed to the BERT reader检索器负责选择包含答案的文本段落，然后将其传递给 BERT 阅读器以确定答案范围。从图 4 to determine the answer scope. 可以看出，BERThis algorithm exhibits significant advantages over traditional algorithms. It demonstrates fast execution speed similar to traditional algorithms while also possessing the characteristics of end-to-end matching, resulting in more precise answer results. Furthermore, it supports extracting answers to questions from multiple documents.serini 是一个智能问答系统，它将 BERT 语言模型与 Anserini 信息检索系统相结合。它协同利用了BERT强大的语言理解能力和Anserini的高效检索功能。与传统算法相比，该算法具有显著优势。它表现出类似于传统算法的快速执行速度，同时还具有端到端匹配的特点，从而获得更精确的答案结果。此外，它还支持从多个文档中提取问题的答案。该算法主要应用于开放域问答任务，其中系统需要从大量非结构化文本中查找问题的答案。