KAG: Boosting LLMs in Professional Domains via Knowledge Augmented Generation
KAG：通过知识增强生成提升专业领域的 LLM

Given a dataset, we use fine-tuning-free LLM(such as GPT-3.5, DeepSeek, QWen, etc,.) or our fine-tuned model Hum to extract entities, events, concepts and relations to construct $KG_{fr}$, subsequently, construct the mutual-indexing structure between $KG_{fr}$ and $RC$, enabling cross-document links through entities and relations. This process includes three steps. First, it extracts the entity set $E=\{e_{1},e_{2},e_{3},...\}$ chunk by chunk, second, extracts the event set ${EV}=\{{ev}_{1},{ev}_{2},{ev}_{3},...\}$ associated to all entities and iteratively extracts the relation set $R=\{r_{1},r_{2},r_{3},...\}$ between all entities in $E$, finally, completes all hypernym relations between the instance and its spgClass. To provide more convenience for the subsequent Knowledge Alignment phase, and overcome the problem of low discrimination of knowledge phrases such as Wikidata[22] and ConceptNet[23], in the entity extraction phase, we use LLMs to generate built-in properties description, summary, semanticType, spgClass, descripitonOfSemanticType by default for each instance e at one time, as shown in Figure 2, we store them in the e instance storage according to the structure of e.description,e.summary, <e, belongTo, semanticType> and <e, hasClass, spgClass>.
给定一个数据集，我们使用免微调的 LLM（如 GPT-3.5、DeepSeek、QWen 等）或我们微调的模型 Hum 来提取实体、事件、概念和关系来构建 $KG_{fr}$ ，随后，构建 和 之间的 $KG_{fr}$ $RC$ 互索引结构，通过实体和关系实现跨文档链接。此过程包括 3 个步骤。首先，它逐块提取实体集 $E=\{e_{1},e_{2},e_{3},...\}$ ，其次，提取与所有实体关联的事件集 ${EV}=\{{ev}_{1},{ev}_{2},{ev}_{3},...\}$ ，并迭代提取 $E$ 中所有实体之间的关系集 $R=\{r_{1},r_{2},r_{3},...\}$ ，最后，完成实例与其 spgClass 之间的所有上位词关系。为了给后续的知识对齐阶段提供更多的便利，克服 Wikidata[22] 和 ConceptNet[23] 等知识短语的低辨别度问题，在实体抽取阶段，我们使用 LLM 为每个实例 e 默认生成内置属性 description、summary、semanticType、spgClass、descripitonOfSemanticType，如图 2 所示， 我们根据 e.description、e.summary、<e、belongTo、semanticType> 和 <e、hasClass、spgClass> 的结构将它们存储在 E 实例存储中。

When openIE is applied to professional domains, irrelevant noise will be introduced. Previous researches[3, 5, 24] have shown that noisy and irrelevant corpora can significantly undermine the performance of LLMs. It is a challenge to align the granularity of extracted information and domain knowledge. The domain knowledge alignment capabilities in KAG include: 1) Domain term and concept injection. We use an iterative extraction approach, First, we store domain concepts and terms with description in KG storage. Second, we extract all instances in the document through openIE, then we perform vector retrieval to obtain all possible concept and term sets $E_{d}$. Finally, we add $E_{d}$ to the extraction prompt and perform another extraction to obtain a set $E_{d}^{a}$ that is mostly aligned with the domain knowledge. 2) Schema-constraint Extraction. In the vertical professional domains, the data structure between multiple documents in each data source such as drug instructions, physical examination reports, government affairs, online order data, structured data tables, etc. has strong consistency, and is more suitable for information extraction with schema-constraint, structured Extraction also makes it easier to do knowledge management and quality improvement. For detailed information about knowledge construction based on Schema-constraint, please refer to the SPG¹
当 openIE 应用于专业领域时，将引入不相关的噪音。以前的研究[ 3， 5， 24] 表明，嘈杂和不相关的语料库会显着破坏 LLM 的性能。调整提取的信息和领域知识的粒度是一项挑战。KAG 中的领域知识对齐功能包括：1） 领域术语和概念注入。我们使用迭代提取方法，首先，我们将域概念和术语与描述存储在 KG 存储中。其次，我们通过 openIE 提取文档中的所有实例，然后我们进行向量检索，以获得所有可能的概念和术语集 $E_{d}$ 。最后，我们添加到 $E_{d}$ 提取提示符中并执行另一次提取，以获得与领域知识基本一致的集合 $E_{d}^{a}$ 。2） 模式约束提取。在垂直专业领域中，每个数据源中诸如药物说明书、体检报告、政府事务、在线订单数据、结构化数据表等多个文档之间的数据结构具有很强的一致性，更适合于具有 schema-constraint 的信息抽取，结构化抽取也使得做知识管理和质量提升变得更加容易。关于基于 Schema-constraint 的知识构建的详细信息，请参考 SPG1¹1Official site of SPG: https://spg.openkg.cn/en-US
SPG 俱乐部官方网站：https://spg.openkg.cn/en-US and OneKE[25]. This article will not introduce it in detail. It is worth noting that, as shown in figure 2, for the same entity type, such as Person, we can pre-define properties and relations such as name, gender, placeOfBirth, (Person, hasFather, Person), (Person, hasFriend, Person), and can also extract tripples directly such as (Jay Chou, spgClass, Person), (Jay Chou, constellation, Capricorn), (Jay Chou, record company, Universal Music Group) through openIE. 3) Pre-defined Knowledge Structures By Document Type. Professional documents such as drug instructions, government affairs documents, and legal definitions generally have a relatively standardized document structure. Each type of document can be defined as an entity type, and different paragraphs are different properties of the entity. Taking government affairs as an example, we can pre-define the GovernmentAffair EntityType and properites such as administrative divisions, service procedures, required materials, service locations, and target groups. The divided chunks are the values of different properties. If the user asks "What materials are needed to apply for housing provident fund in Xihu District?", you can directly take out the chunk corresponding to property required materials to answer the question, avoiding the possible hallucinations caused by LLM re-generation.
和 OneKE[25]。 本文不会详细介绍。值得注意的是，如图 2 所示，对于相同的实体类型，比如 Person，我们可以预定义 name， gender， placeOfBirth、（Person， hasFather， Person）、（Person， hasFriend， Person） 等属性和关系，也可以通过 openIE 直接提取 （Jay Chou， spgClass， Person）、（Jay Chou， constellation， Capricorn）、（Jay Chou， 唱片公司， 环球音乐集团） 等三元组。3） 按文档类型预定义的知识结构 。药品说明书、政务文件、法律定义等专业文件一般具有相对标准化的文档结构。每种类型的文档都可以定义为实体类型，不同的段落是实体的不同属性。以政府事务为例，我们可以预先定义 GovernmentAffair EntityType 和行政区划、服务流程、所需材料、服务地点和目标人群等属性。划分的 chunk 是不同 property 的值。如果用户问到“西湖区申请住房公积金需要什么材料”，可以直接拿出物业所需材料对应的 chunk 来回答问题，避免了 LLM 再生可能带来的幻觉。

KAG’s mutual-indexing is a knowledge management and storage mechanism that conforms to the LLMFriSPG semantic representation. As is described in section 2.1, it includes four core data structures: 1) Shared Schemas are coarse-grained-types pre-defined as SPG Classes at project level, it includes EntityTypes, ConceptTypes, and EventTypes, they serve as a high-level categorization such as Person, Organization, GEOLocation, Date, Creature, Work, Event. 2) Instance Graph include all event and entity instances in $KG_{cs}$ and $KG_{fr}$. that is, instances constructed through openIE with schema-free or structured extraction with schema-constraint are both stored as instances in KG storage. 3) Text Chunks are special entity node that conforms to the definition of the Chunk EntityType. 4) Concept Graph is the core component for knowledge alignment. it consists of a series of concepts and concept relations, concept nodes are also fine-grained-types of instances. Through relation prediction, instance nodes can be linked to concept nodes to obtain their fine-grained semantic types. , and two storage structures: 1) KG Store. Store KG data structures in LPG databases, such as TuGraph, Neo4J. 2) Vector Store. Store text and vectors in a vector storage engine, such as ElasticSearch, Milvus, or the vector storage embedded in the LPG engine.
KAG 的互索引是一种符合 LLMFriSPG 语义表示的知识管理和存储机制。如第 2.1 节所述，它包括四个核心数据结构：1） 共享模式是在项目级别预定义为 SPG 类的粗粒度类型，它包括 EntityType、ConceptType 和 EventType，它们用作高级分类，例如人员、组织、GEOLocation、日期、生物、工作、事件 。2） 实例图包括 $KG_{cs}$ 和 中的所有 $KG_{fr}$ 事件和实体实例。也就是说，通过 openIE 构建的具有 schema-free 的实例或具有 schema-constraint 的结构化提取都作为实例存储在 KG 存储中。3） 文本块是符合 Chunk EntityType 定义的特殊实体节点。4） 概念图是知识对齐的核心组件。它由一系列的概念和概念关系组成，概念节点也是细粒度类型的实例。通过关系预测，实例节点可以链接到概念节点，以获得其细粒度的语义类型。和两种存储结构：1） KG Store。将 KG 数据结构存储在 LPG 数据库中，例如 TuGraph、Neo4J。2） 载体存储 。将文本和向量存储在向量存储引擎中，例如 ElasticSearch、Milvus 或 LPG 引擎中嵌入的向量存储。

Retrieval. According to the the knowledge or information retrieved from SPO, s, p, o should not repeatedly appear multiple times in the same expression. Constraints can be applied to the s, p, o for querying. For multi-hop queries, multiple retrievals are required. When the current variable refers to a previously mentioned variable, the variable name must be consistent with the referenced variable name, and only the variable name needs to be provided. The knowledge type and name are only specified during the first reference.
检索 。根据从 SPO 检索到的知识或信息，s、p、o 不应在同一表达式中多次重复出现。约束可以应用于 s、p、o 以进行查询。对于多跃点查询，需要多次检索。当当前变量引用前面提到的变量时，变量名必须与引用的变量名一致，只需要提供变量名。知识类型和名称仅在第一次引用期间指定。

Sort. Sort the retrieved results. $A$ is the variable name for the retrieved subject-predicate-object(SPO) ($s_{i}$, $o_{i}$, or $s.prop$, $p.prop$, $o.prop$). $direction$ specifies the sorting direction, where $direction=min$ means sorting in ascending order and $direction=max$ means sorting in descending order. $limit=n$ indicates outputting the topN results.
排序 。对检索到的结果进行排序。 $A$ 是检索到的主语谓词对象 （SPO） （ $s_{i}$ 、 $o_{i}$ 或 $s.prop$ 、 、 $p.prop$ ） $o.prop$ 的变量名称。 $direction$ 指定排序方向，其中 $direction=min$ 表示升序排序， $direction=max$ 表示降序排序。 $limit=n$ 表示输出 topN 结果。

Math. Perform mathematical calculations. $expr$ is in LaTeX syntax and can be used to perform calculations on the retrieved results (sets) or constants. $math_{i}$ represents the result of the calculation and can be used as a variable name for reference in subsequent actions.
数学 。执行数学计算。 $expr$ 采用 LaTeX 语法，可用于对检索到的结果（集合）或常量执行计算。 $math_{i}$ 表示计算的结果，并可用作变量名称，以便在后续作中引用。

Deduce. Deduce the retrieval or calculation results to answer the question. $A,B$ can be the variable names from the retrieved SPO or constants. The operator $op=entailment|greater|less|equal$ represents $A$ entails $B$, $A$ is greater than $B$, $A$ is less than $B$, and $A$ is equal to $B$, respectively.
Deduce.推断检索或计算结果以回答问题。 $A,B$ 可以是检索到的 SPO 或常量中的变量名称。运算符 $op=entailment|greater|less|equal$ 分别表示 $A$ 、 $B$ $A$ 大于 $B$ 、 $A$ 小于 $B$ 和 $A$ 等于 $B$ 。

When the query statement represented by natural language is applied to the search, the logic is often fuzzy, such as "find a picture containing vegetables or fruits" and "find a picture containing vegetables and fruits". Whether text search or vector search is used, the similarity between the two queries is very high, but the corresponding answers are quite different. The same is true for problems involving logical reasoning processes such as and or not, and intersection differences. To this end, we use logical form to express the question, so that it can express explicit semantic relations. Similar to IRCOT, we decompose complex original problem and plan out various execution actions such as multi-step retrieval, numerical reasoning, logical reasoning, and semantic deduce. Each sub-problem is expressed using logical form functions, and dependencies between sub-questions are established through variable references. The inference resolution process for each sub-question is illustrated as Algorithm 11. In this process, the GraphRetrieval module performs KG structure retrieval according to the logical form clause to obtain structured graph results. Another key module, HybridRetrieval, combining natural language expressed sub-problems and logical functions for comprehensive retrieval of documents and sub-graph information. To understand how logical functions can be utilized to reason about complex problems, refer to the following examples as Table 14.
当自然语言表示的查询语句应用于搜索时，逻辑往往是模糊的，比如 “find a picture containing vegetables or fruits” 和 “find a picture containing vegetables and fruits” 等。无论使用文本搜索还是向量搜索，这两个查询之间的相似度都很高，但对应的答案却大相径庭。对于涉及逻辑推理过程（如 and or not）和交集差异的问题也是如此。为此，我们用逻辑形式来表达问题，使其能够表达明确的语义关系。与 IRCOT 类似，我们分解复杂的原始问题并规划出多步骤检索、数字推理、逻辑推理和语义演绎等各种执行动作。每个子问题都使用逻辑形式函数表示，子问题之间的依赖关系通过变量引用建立。每个子问题的推理解决过程如算法 11 所示。在此过程中，GraphRetrieval 模块根据逻辑形式子句进行 KG 结构检索，得到结构化图结果。另一个关键模块 HybridRetrieval，结合自然语言表达的子问题和逻辑函数，对文档和子图信息进行综合检索。要了解如何利用逻辑函数来推理复杂问题，请参考 表 14 中的以下示例。

In naive RAG, retrieval is achieved by calculating the similarity (e.g. cosine similarity) between the embeddings of the question and document chunks, where the semantic representation capability of embedding models plays a key role. This mainly includes a sparse encoder (BM25) and a dense retriever (BERT architecture pre-training language models). Sparse and dense embedding approaches capture different relevance features and can benefit from each other by leveraging complementary relevance information.
在 naïve RAG 中，检索是通过计算问题块和文档块的嵌入之间的相似性（例如余弦相似度）来实现的，其中嵌入模型的语义表示能力起着关键作用。这主要包括一个稀疏编码器（BM25）和一个密集检索器（BERT 架构预训练语言模型）。稀疏和密集嵌入方法可以捕获不同的相关性特征，并且可以通过利用互补的相关性信息来相互受益。

The existing method of combining the two is generally to combine the scores of the two search methods in an ensemble, but in practice different search methods may be suitable for different questions, especially in questions requiring multi-hop reasoning. When query involves proper nouns, people, places, times, numbers, and coordinates, the representation ability of the pre-trained presentation model is limited, and more accurate text indexes are needed. For queries that are closer to the expression of a paragraph of text, such as scenes, behaviors, and abstract concepts, the two may be coupled in some questions.
现有的将两者结合起来的方法一般是将两种搜索方法的分数组合成一个集合，但在实践中，不同的搜索方法可能适用于不同的问题，尤其是在需要多跳推理的问题中。当查询涉及专有名词、人、地点、时间、数字和坐标时，预训练的表示模型表示能力有限，需要更准确的文本索引。对于更接近于文本段落表达的查询，例如场景、行为和抽象概念，在某些问题中可能会将两者耦合在一起。

In the design of logical form, it is feasible to effectively combine two retrieval methods. When keyword information is needed as explicit filtering criteria, conditions for selection can be specified within the retrieval function to achieve structured retrieval.
在逻辑形式的设计中，将两种检索方法有效结合是可行的。当需要关键字信息作为显式过滤条件时，可以在检索函数中指定选择条件，实现结构化检索。

For example, for the query "What documents are required to apply for a disability certificate at West Lake, Hangzhou?", the retrieval function could be represented as: "Retrieval(s=s1:Event[applying for a disability certificate], p=p1:support_chunks, o=o1:Chunk, s.location=West Lake, Hangzhou)". This approach leverages the establishment of different indices (sparse or dense) to facilitate precise searches or fuzzy searches as needed.
例如，对于查询 “杭州西湖申请残障证需要什么文件？”，检索函数可以表示为 ：“Retrieval（s=s1：Event[申请残障证]， p=p1：support_chunks， o=o1：Chunk， s.location=杭州西湖）”。 此方法利用建立不同的索引（稀疏或密集）来促进所需的精确搜索或模糊搜索。

Furthermore, when structured knowledge in the form of SPO cannot be retrieved using logical functions, alternative approaches can be employed. These include semi-structured retrieval, which involves using logical functions to search through chunks of information, and unstructured retrieval. The latter encompasses methods such as Retrieval-Augmented Generation (RAG), where sub-problems expressed in natural language are used to retrieve relevant chunks of text. This highlights the adaptability of the system to leverage different retrieval strategies based on the availability and nature of the information.
此外，当无法使用逻辑函数检索 SPO 形式的结构化知识时，可以采用替代方法。这些包括半结构化检索（涉及使用逻辑函数搜索信息块）和非结构化检索。后者包括 Retrieval-Augmented Generation （RAG） 等方法，其中用自然语言表达的子问题用于检索相关的文本块。这突出了系统根据信息的可用性和性质利用不同检索策略的适应性。

The process of enhancing indexing through semantic reasoning, as shown in Figure 5 , specifically implemented as predicting semantic relations or related knowledge elements among index items using LLM, encompassing four strategies:
通过语义推理增强索引的过程，如图 5 所示，具体实现为使用 LLM 预测索引项之间的语义关系或相关知识元素，包括四种策略：

• •

  Disambiguation and fusion of knowledge instances. Taking entity instance $e_{cur}$ as an example, first, the one-hop relations and description information of $e_{cur}$ are used to predict synonymous relations to obtain the synonym instance set $E_{syn}$ of $e_{cur}$. Then, the fused target entity $e_{tar}$ is determined from $E_{syn}$. Finally, the entity fusion rules are used to copy the properties and relations of the remaining instances in $E_{syn}$ to $e_{tar}$, and the names of these instances are added to the synonyms of $e_{tar}$, the remaining instances will also be deleted immediately.

  Report issue for preceding element
  
  • 知识实例的消歧和融合。以实体实例 $e_{cur}$ 为例，首先，利用 的 $e_{cur}$ 单跳关系和描述信息来预测同义关系，得到 的 $e_{cur}$ 同义词实例集 $E_{syn}$ 。然后，根据 确定融合的目标实体 $e_{tar}$ $E_{syn}$ 。最后，使用实体融合规则将剩余实例的属性和关系复制到 $E_{syn}$ 中 $e_{tar}$ ，并将这些实例的名称添加到 $e_{tar}$ 的同义词中，剩余的实例也会立即删除。
• •

  Predict relations between instances and concepts. For each knowledge instance (such as event, entity), predict its corresponding concept and add the derived triple $<e_{i},\ belongTo,\ c_{j}>$ to the knowledge index. As is shown in Figure 5, <Chamber, belongTo, Legislative Body> means that the Chamber belongs to Legislative Body in classification.

  Report issue for preceding element
  
  • 预测实例和概念之间的关系。对于每个知识实例（如 event、entity），预测其对应的概念，并将派生的三元组 $<e_{i},\ belongTo,\ c_{j}>$ 添加到知识索引中。如图 5 所示， 是指商会在分类上属于立法机构。
• •

  Complete concepts and relations between concepts. During the extraction process, we use concept reasoning to complete all hypernym and isA relations between semanticType and spgClass. As is shown in Figure 5 and Table 2, we can obtain the semanticType of Chamber is Legislative Body, and its spgClass is Organization in the extraction phase. Through semantic completion, we can get <Legislative Body, isA, Government Agency>, <Government Agency, isA, Organization>. Through semantic completion, the triple information of $KG_{fr}$ space is more complete and the connectivity of nodes is stronger.

  Report issue for preceding element
  
  • 完整的概念和概念之间的关系。在提取过程中，我们使用概念推理来完成 semanticType 和 spgClass 之间的所有 hypernym 和 isA 关系。如图 5 和表 2 所示，我们可以得到 Chamber 的 semanticType 是 Legislative Body，它的 spgClass 在提取阶段是 Organization。通过语义完成，我们可以得到 ， 。通过语义完成，空间的 $KG_{fr}$ 三重信息更加完整，节点的连通性更强。

In the retrieval phase, we utilize semantic relation reasoning to search the KG index based on the phrases and types in the logical form. For the types, mentions or relations in the logical form, we employ the method of combining semantic relation reasoning with similarity retrieval to replace the traditional similarity retrieval method. This retrieval method makes the retrieval path professional and logical, so as to obtain the correct answer. First, the hybrid reasoning performs precise type matching and entity linking. If the type matching fails, then, semantic reasoning is performed. As shown in Figure 6, if the type Political Party fails to match, semantic reasoning is used to predict that Political Party contains Political Faction, and reasoning or path calculation is performed starting from Political Faction.
在检索阶段，我们利用语义关系推理，根据逻辑形式的短语和类型搜索 KG 索引。对于逻辑形式的类型、提及或关系，我们采用语义关系推理与相似性检索相结合的方法，以取代传统的相似性检索方法。这种检索方法使检索路径专业且合乎逻辑，从而获得正确答案。首先，混合推理执行精确的类型匹配和实体链接。如果类型匹配失败，则执行语义推理。如图 6 所示，如果 Political Party 类型不匹配，则使用语义推理来预测 Political Party 包含 Political Faction，并从 Political Faction 开始进行推理或路径计算。

Take another example. If the user query $q_{1}$ is "Which public places can cataract patients go for leisure?" and the document content $d_{2}$ is "The museum is equipped with facilities to provide barrier-free visiting experience services such as touch, voice interpretation, and fully automatic guided tours for the visually impaired.", It is almost impossible to retrieve $d_{2}$ based on the vector similarity with $q_{1}$. However, it is easier to retrieve $d_{2}$ through the semantic relation of <cataract patient, isA, visually impaired>.
再举一个例子。如果用户查询 $q_{1}$ 为 “哪些公共场所可以白内障患者休闲游”，文档内容 $d_{2}$ 为 “博物馆设有设施，提供触控、语音传译、视障全自动导游等无障碍参观体验服务”，则几乎不可能根据向量相似度检索 $d_{2}$ 。 $q_{1}$ 然而，通过 <cataract 患者 isA、视力受损> 的语义关系更容易检索 $d_{2}$ 。

NLU is one of the most common foundational tasks in natural language processing, including text classification, named entity recognition, relation Extraction, subject and object extraction, trigger detection, event argument extraction, event extraction, and machine reading comprehension. We have collected over 30 public datasets to enhance understanding capabilities. Experiments found that simply transforming the original datasets into instruction datasets can achieve comparable results to specialized models on trained tasks, but this approach does not improve the model’s NLU capabilities on unseen domains. Therefore, we conducted large-scale instruction reconstruction, designing various instruction synthesis strategies to create an NLU instruction dataset with over 20,000 diverse instructions. By utilizing this dataset for supervised fine-tuning on a given base model, the model has demonstrated enhanced NLU capabilities in downstream tasks. The instruction reconstruction strategy mainly consists of the following three types.
NLU 是自然语言处理中最常见的基础任务之一，包括文本分类、命名实体识别、关系提取、主题和对象提取、触发器检测、事件参数提取、事件提取和机器阅读理解 。我们收集了 30 多个公共数据集，以增强理解能力。实验发现，简单地将原始数据集转换为指令数据集可以在训练任务上获得与专业模型相当的结果，但这种方法并不能提高模型在看不见的域上的 NLU 能力。因此，我们进行了大规模的指令重构，设计了各种指令合成策略，以创建一个包含超过 20,000 条不同指令的 NLU 指令数据集。通过利用此数据集对给定的基础模型进行监督微调，该模型在下游任务中展示了增强的 NLU 功能。指令重构策略主要包括以下三种类型。

• •

  Label bucketing: [25]This strategy focuses on label-guided tasks, where the aim is to extract text based on labels or map text to specified labels, including classification, NER, RE, and EE. When labels in a dataset collectively co-occur in the training set, the model may learn this pattern and overfit to the dataset, failing to independently understand the meaning of each label. Therefore, during the instruction synthesis process, we adopt a polling strategy that designates only one label from each training sample as part of a bucket. Additionally, since some labels have similar semantics and can be confused, we group easily confused labels into a single bucket, allowing the model to learn the semantic differences between the two labels more effectively.

  Report issue for preceding element
  
  • 标签分桶：[ 25]此策略侧重于标签导向型任务，其目的是根据标签提取文本或将文本映射到指定的标签，包括分类、NER、RE 和 EE。当数据集中的标签在训练集中共同出现时，模型可能会学习这种模式并过度拟合数据集，从而无法独立理解每个标签的含义。因此，在指令合成过程中，我们采用一种轮询策略，从每个训练样本中仅指定一个标签作为桶的一部分。此外，由于某些标签具有相似的语义并且可能会混淆，因此我们将容易混淆的标签分组到一个桶中，使模型能够更有效地学习两个标签之间的语义差异。
• •

  Flexible and Diverse Input and Output Formats: The LLM employs an instruction-following approach for inference, and a highly consistent input-output format may cause the model to overfit to specific tasks, resulting in a lack of generalization for unseen formats. Therefore, we have flexibly processed the input and output formats. The output is handled as five different formatting instructions, as well as two types of natural language instructions. Additionally, the output format can dynamically be specified as markdown, JSON, natural language, or any format indicated in the examples.

  Report issue for preceding element
  
  • 灵活多样的输入和输出格式：LLM 采用指令跟随方法进行推理，高度一致的输入输出格式可能会导致模型过度拟合特定任务，从而导致对看不见的格式缺乏泛化。因此，我们灵活地处理了输入和输出格式。输出作为 5 种不同的格式指令以及两种类型的自然语言指令进行处理。此外，输出格式可以动态指定为 markdown、JSON、自然语言或示例中指示的任何格式。
• •

  Instructoin with Task Guildline: Traditional NLP training often employs a "sea of questions" approach, incorporating a wide variety of data in the training set. This allows the model to understand task requirements during the learning process, such as whether to include job titles when extracting personal names. For the training of LLMs, we aim for the model to perform tasks like a professional annotator by comprehending the task description. Therefore, for the collected NLU tasks, we summarize the task descriptions using a process of self-reflection within the LLM. This creates training data that includes task descriptions within the instructions. Additionally, to enhance task diversity, we implement heuristic strategies to rephrase the task descriptions and answers. This enables the model to understand the differences between task descriptions more accurately and to complete tasks according to the instructions.

  Report issue for preceding element
  
  • 使用 Task Guildline 进行指导：传统的 NLP 训练通常采用“问题之海”方法，在训练集中包含各种数据。这允许模型了解学习过程中的任务要求，例如在提取个人姓名时是否包含职称。对于 LLM 的训练，我们的目标是通过理解任务描述，让模型像专业注释者一样执行任务。因此，对于收集的 NLU 任务，我们使用 LLM 中的自我反思过程来总结任务描述。这将创建包含说明中任务描述的训练数据。此外，为了增强任务多样性，我们实施了启发式策略来重新表述任务描述和答案。这使模型能够更准确地理解任务描述之间的差异，并根据说明完成任务。

We fine-tuned six foundational models: qwen2, llama2, baichuan2, llama3, mistral, phi3, and used six understanding benchmarks recorded on OpenCompass for performance validation. The table 3 shows that the KAG-Model has a significant improvement in NLU tasks.
我们微调了六个基础模型：qwen2、llama2、baichuan2、llama3、mistral、phi3，并使用了 OpenCompass 上记录的六个理解基准测试进行性能验证。表 3 显示 KAG 模型在 NLU 任务方面有显着改进。

The NLI task is used to infer the semantic relations between given phrases. Typical NLI tasks include entity linking, entity disambiguation, taxonomy expansion, hypernym discovery, and text entailment. In the context of knowledge base Q&A, due to the diversity and ambiguity of natural language expressions, as well as the subtle and different types of semantic connections between phrases, it often requires further alignment or retrieval of related information through NLI tasks based on NLU. As described in section 2.4, we categorize the key semantic relation in knowledge base applications into six types. Among these, relations such as isA, isPartOf and contains exhibit directional and distance-based partial order relations. During the reasoning process, it is crucial to accurately determine these semantic relations to advance towards the target answer. In traditional approaches, separate training of representation pre-training models and KG completion(KGC) models is often employed to reason about semantic relations. However, these KGC models tend to focus on learning graph structures and do not fully utilize the essential textual semantic information for semantic graph reasoning. LLMs possess richer intrinsic knowledge, and can leverage both semantic and structural information to achieve more precise reasoning outcomes. To this end, we have collected a high-quality conceptual knowledge base and ontologies from various domains, creating a conceptual knowledge set that includes 8,000 concepts and their semantic relations. Based on this knowledge set, we constructed a training dataset that includes six different types of conceptual reasoning instructions to enhance the semantic reasoning capabilities of a given base model, thereby providing semantic reasoning support for KAG.
NLI 任务用于推断给定短语之间的语义关系。典型的 NLI 任务包括实体链接、实体消除歧义、分类扩展、上位词发现和文本蕴涵 。在知识库问答的语境中，由于自然语言表达的多样性和歧义性，以及短语之间微妙而不同类型的语义联系，往往需要通过基于 NLU 的 NLI 任务对相关信息进行进一步的对齐或检索。如 2.4 节所述，我们将知识库应用程序中的关键语义关系分为六种类型。其中，isA、isPartOf 和 contains 等关系表现出方向和基于距离的偏序关系。在推理过程中，准确确定这些语义关系以推进目标答案至关重要。在传统方法中，通常采用表征预训练模型和 KG 完成 （KGC） 模型的单独训练来推理语义关系。然而，这些 KGC 模型往往专注于学习图结构，并没有充分利用基本的文本语义信息进行语义图推理。LLM 拥有更丰富的内在知识，可以利用语义和结构信息来实现更精确的推理结果。为此，我们从各个领域收集了高质量的概念知识库和本体论，创建了一个包含 8000 个概念及其语义关系的概念知识集。 基于这个知识集，我们构建了一个训练数据集，其中包括六种不同类型的概念推理指令，以增强给定基础模型的语义推理能力，从而为 KAG 提供语义推理支持。

Semantic reasoning is one of the core ability required in KAG process, we use NLI tasks and general reasoning Q&A tasks to evaluate the ability of our model, the results are as shown in Table 4 and Table 5. The evaluation results indicates that our KAG-Model demonstrates a significant improvement in tasks related with semantic reasoning: First, Table 5 shows that on the Hypernym Discovery task(which is consistent in form with the reasoning required in semantic enhanced indexing and retrieval.), our fine-tuned KAG-llama model outperforms Llama3 and ChatGPT-3.5 significantly. In addition, the better performance of our model on CMNLI, OCNLI and SIQA compared with Llama3 in Table 4 shows that our model has good capabilities in general logical reasoning.
语义推理是 KAG 过程中需要的核心能力之一，我们使用 NLI 任务和一般推理 Q&A 任务来评估我们模型的能力，结果如表 4 和表 5 所示。评估结果表明，我们的 KAG 模型在与语义推理相关的任务方面表现出显着改进：首先，表 5 显示，在上位词发现任务（在形式上与语义增强索引和检索所需的推理一致）上，我们微调的 KAG-llama 模型明显优于 Llama3 和 ChatGPT-3.5。此外，与表 4 中的 Llama3 相比，我们的模型在 CMNLI、OCNLI 和 SIQA 上的性能更好，这表明我们的模型具有良好的一般逻辑推理能力。

Models that have not undergone domain adaptation training often exhibit significant differences from the target text in domain logic and writing style. Moreover, acquiring sufficient amounts of annotated data in specialized domains frequently poses a challenge. Therefore, we have established two efficient fine-tuning methods for specific domain scenarios, allowing the generation process to better align with scene expectations: namely, K-Lora and AKGF.
未经过领域适应训练的模型通常在领域逻辑和写作风格上表现出与目标文本的显著差异。此外，在专业领域中获取足够数量的注释数据经常会带来挑战。因此，我们为特定领域场景建立了两种高效的微调方法，使生成过程更好地与场景预期保持一致：K-Lora 和 AKGF。

Pre-learning with K-LoRA. First of all, we think that using knowledge to generate answers is the reverse process of extracting knowledge from text. Therefore, by inverting the previously described extraction process, we can create a ’triples-to-text’ generation task. With extensive fine-tuning on a multitude of instances, the model can be trained to recognize the information format infused by the KG. Additionally, as the target text is domain-specific, the model can acquire the unique linguistic style of that domain. Furthermore, considering efficiency, we continue to utilize LoRA-based SFT. We refer to the LoRA obtained in this step as K-LoRA.
使用 K-LoRA 进行预学习 。首先，我们认为使用知识生成答案是从文本中提取知识的相反过程。因此，通过反转前面描述的提取过程，我们可以创建一个 'triples-to-text' 生成任务。通过对大量实例进行广泛的微调，可以训练模型识别 KG 注入的信息格式。此外，由于目标文本是特定于域的，因此模型可以获取该域的独特语言风格。此外，考虑到效率，我们继续使用基于 LoRA 的 SFT。我们将此步骤中获得的 LoRA 称为 K-LoRA。

Alignment with KG Feedback. The model may still exhibit hallucinations in its responses due to issues such as overfitting. Inspired by the RLHF(Reinforcement Learning with Human Feedback) approach[27, 28], we hope that the KG can serve as an automated evaluator, providing feedback on knowledge correctness of the current response, thereby guiding the model towards further optimization. First, we generate a variety of responses for each query by employing diverse input formats or random seeds. Subsequently, we incorporate the KG to score and rank these responses. The scoring process compare generated answer with knowledge in KG to ascertain their correctness. The reward is determined by the number of correctly matched knowledge triples. The formula for calculating the reward is represented by Formula 1.
与 KG 反馈保持一致 。由于过度拟合等问题，模型可能仍会在其响应中表现出幻觉。受 RLHF（Reinforcement Learning with Human Feedback）方法 [27， 28] 的启发，我们希望 KG 可以作为一个自动评估器，提供当前响应的知识正确性反馈，从而引导模型进一步优化。首先，我们通过采用不同的输入格式或随机种子为每个查询生成各种响应。随后，我们纳入 KG 来对这些回答进行评分和排名。评分过程将生成的答案与 KG 中的知识进行比较，以确定它们的正确性。奖励由正确匹配的知识三元组的数量决定。计算奖励的公式由公式 1 表示。

where $\alpha$ is a hyperparameter, rspo represents the number of SPO matches, and re represents the number of entity matches.
其中 $\alpha$ 是超参数，rspo 表示 SPO 匹配项的数量，re 表示实体匹配项的数量。