The open domain Q&A QA system based on knowledge graph is essentially an entity alignment and search problem. It needs to nd and extract the information contained and lacked in the question in the comparison with graph data unit. Our team proposed two search models: text similarity and graph computation. Combined them with some basic strategies, the result sets of the two are integrated, and a good effect is obtained. The open domain Q&A QA system based on knowledge graph is essentially an entity alignment and search problem. It needs to nd and extract the information contained and lacked in the question in the comparison with graph data unit. The meta-knowledge form of the ”subject-predicate-object” triple of knowledge graph has a corresponding relationship with the intrinsic expression logic of natural language. The problem of simple sentence can basically be one-to-one correspondence; but for the semantically similar but literally irrelevant form or contains nodes with large information density in the graph, the alignment of the question and the graph data becomes very complicated, and the noise information in the question and the redundant and irregular data in the graph make the process become more difcult and error-prone. Our team proposed two search models: text similarity and graph computation. The former is good at solving the 1-degree problem, and the latter is good at solving the multi-degree problem which has larger number of information nodes. Combined with some basic strategies, the result sets of the two are fused, and a certain effect is obtained.
From the perspective of the knowledge structure of existing graph, sentence is analyzed by information segmentation. The sentence is a path composed of different knowledge nodes, and different nodes correspond to different phrases in the sentence. Text-based search is to nd possible paths in the graph, and then take the search path with the highest similarity.
The evaluation of similarity needs to consider the following aspects: 1) The similar values between graph nodes and phrases in sentence; 2) The similarity matching value of logical order between graph node and node; 3) The question type and the position of the question keyword in sentence correspond to the position of adjacent node in graph. 3.1
Entity discovery Searching for the way that all triples are similarly aligned entities will be affected by larger noise information, resulting in the loss of search entry, and the annotation information of object entity will also affect the ordering of search hit entities. For example: the most similar triple of ”德国著名的汽车品牌” is as follows, { ”id” : 37730758, ”subject” : ” <著名汽车品牌图鉴> ”, ”predict” : ” <软件名称> ”, ”object” : ”著名汽车品牌图鉴” }, causes its entry entity (subject entity - 德国) to be lost.
Therefore, entity discovery needs to rely on mention data, subject entity index data, and object entity index data. All possible subject entities and object entities in the question are the starting point of the question. The rest information of the question belongs to predicate information. Generally, the matching requirements of subject and object entity should be higher than those of predicate, even predicate node information is not included in the question.
1）Use object entity index to nd whole sentence entity Long sentence may appear as object entity as a whole.
For example：首位登上过时代周刊封面的华人女歌手？谁首次实现了中国的首次统一？ 2) Discover entity using mention index 3) Identify the predicate entity or type entity that may appear before main entity For example：书籍《亚利安01》的价格是多少 ”书籍”、”亚利安01”were found to be entities, but the former and the latter have category relation. 1-degree problem solving The search implements the stepwise matching of the information contained in the question by a similar matching cursor mechanism. Each time an information node is matched, the matching cursor will narrow the description information range of the question.
The 1-degree problem solving uses main entity as subject or object. The remaining information after excluding main entity in the question is used as predicate information. The search is performed through triple index, and the search result is veried, return to the triple, if the relationship between predicate and the remaining information in the question is inclusion relationship or the similarity of the largest similar match string is above a certain threshold, the triple is considered to be the answer corresponding to the question description. Then the matching cursor will move to the right. If the remaining information in the question still contains potential predicate entity or object entity after excluding stop words, it is regarded as a 2-degree problem and is executed by other modules; otherwise, it is regarded as 1-degree question and the search results are returned directly. 2-degree problem solving The matching of triple to problem in 2-degree problem requires at least two triples to be satised. Since the alignment sequence between the structure of graph and the phrases in the question is not exactly matched from left to right, in order to avoid the complicated processing of phrase combination sequence, we will spread the nodes related to main entity at 2 degrees. Then, calculate the similarity between each piece and the question, and select the best matching triple as candidate answer. Combined with the structure of triples, the description of 2-degree problem can be summarized into the following six types: 1）S-P-<>-P-? 2）S-<>-O-P-? The main entity starts as a subject and queries its indirect attribute values. Similar to 1), but the subject entity is not directly related to the attribute, but its object entity. for example：澳大利亚的悉尼有什么著名景点？ 3）S-P-?-P-<> Similar to 1), but its problem is another subject of the object entity of main entity. for example：周恩来的妻子曾当过什么主席？ 4）O-P-S-P-? The main entity starts as an object entity and queries its indirect attribute values.
For example：小说《哈利波特》的作者是谁？郦道元的《水经注》编撰于哪个朝代？ 5）O-<>-?-P-O1 6）O-P-?-<>-O1 The main entity starts as an object entity, and with the aid of other entities in the question, queries its subject entity. Similar to 5), but predicate information is the parent attribute of main entity. for example：有哪些位于湖北的公司？
In general, 2-degree problem is divided into three steps: 1) spread 2-degree solution space; 2) pre-screening result set of question similar match; 3) based on the pre-screening result set from 2), similar match is graded according to 6 types of template, and the triple with the highest degree of matching is considered a candidate answer.
The entity in the question can basically use pkuorder as a dictionary for entity connection. The generation method of presolution space is: for possible entities, go to the database to search its child nodes or parent nodes as the rst hop node, and then search for the child node of the rst hop node. It acts as the second hop node. The corresponding graphs of different entity entries are stored separately.
For example, the question ”董卿主持的正在播出的节目是?”,when we take ”董卿” as entity entry, we can get the following pre-solution graph.
董卿 职业 主持人
起源地 代表人物 主持人 元宵晚会 播出状态 主持人 美国 朱军 播出中 周涛
This example explains why two directions are used in the rst hop. Because ”元宵晚会的主持人是董卿” and ”董卿主持了元宵晚会”, the same fact is expressed, but we can’t determine which mode is used for database to store this knowledge. The pre-solution tree is usually more than one. For this problem, the ”节目” can also generate a pre-solution tree.
A QA search algorithm based on the fusion integration of text similarity and graph computation 4.3
The similarity features of quantitative characteristics For example, ”Double hop problem”: ”张柏芝、谢霆锋合作的动作电影里的服装是谁设计的”.
Let us remember a knowledge triplet as S P O Structural features First, we will get two pre-solution trees with Cecilia Cheung and Nicholas Tse as entries. For the subtree with the form of S1 P1 O P2 S2on each pre-solution tree, it is a subtree of, where S1=”<张柏芝>” or ”<谢霆锋>”. We use the following characteristics to calculate seq, the jaccard distance of seq, S1 + P1 + P2, the number of common characters ofseq, S1 + P1 + P2 + O + S2), hint, seq, the edit distance of S1 + P1 + P2, Les, seq and the most important thing is that we introduce a denition called placeholder. That is, these cue words such as ”是谁,是什么,哪里,何时,哪个,有什么”in the question. It is found that there is often important information of answer reminder near the cue word. For this problem, ”design” is a very important message. If the P2 of the subtree S1 P1 O P2 S2 contains design, then its score should be very high. Specically, we nd a substring Sattach in the sentence for P2, so that the string can coincide with the head or tail of P2, and then take the distance between the string Sattach and the placeholder as a feature.
Node consolidation We got an array with elements (path : S1 P1 O P2 S2, jaccard, hint, les, placeholder) (using S2 as a possible answer node). But it is noticed that what the question asked was the ”张柏芝,谢霆锋共同主演的电影”. As the bridge node O, ”无极” will actually connect with ”张柏芝” and ”谢霆锋” as S1. In the case of bridge node O is the same as answer S2, we will merge the two candidate answers and recalculate relevant features.
Pruning Since the previous entity link used an exhaustive matching method, the size of the candidate answers is very large at this step. If the classier is trained on this, the training will fail due to the category imbalance. Inspired by the decisionmaking tree model, we use heuristic rules for pruning. Using the mean value of one or several indicators as a benchmark, we search for magnication a, so that the number of pre-selected answers which are greater than amean is less than 80. Thus, we control the positive and negative ratios of training.
Similarity features Using pure quantitative characteristics makes us lose a lot of information, so it is necessary to input new features for the nal perceptron. Here CNN is used to measure the similarity between the path of candidate answers S1P1OP2, S1P1 and questions seq. Using single hop, double hop to solve each problem will get answer and relevant path, training is performed by using path, seq as X, the F1 of answer as Y. 5 5.1
Answer type is determined according to the cue words of the question (”是谁, 什么时间, 哪个国家” etc.), and the one accord with the type is prioritized to be selected. For example, in the case of a character, ”董卿” has an attribute of ”类型-> 人物”, thus, we can judge that this entity is a character. In fact, attribute needs to be reconstructed, and the type in pkutype has error information. We can reconstruct the type in an iterative way and clean the noise. Specically, starting from the attribute of ”人物”, we can get a bunch of entities, and then this bunch of entities will generate a bunch of predicates P. We sort the predicates and use some features like tf-idf to screen out common features (such as ”中文名”), so we build the mapping of the ”人物” type to the predicate. We can in turn reconstruct the pkutype through the predicate set so as to eliminate noise. 5.2
Fusion strategy based on graph traversal path When the path that generates the answer accords well with the problem (information is neither too much or too little), then the problem is a priority. 5.3
Voting If there are two models that choose the same answer, then the nal answer will be this answer. 6
The similarity model and the graph computation perceptron model were evaluated on the validation set and the test set, respectively, and the fusion integration test was performed on the test set. The test data is shown in the table below.
The test data shows that the similarity and graph calculation perceptron models have their own strong characteristics, and the two have complementary characteristics. After fusion integration, scores are improved.
This paper proposes a comparison analysis of implementation problems and graph knowledge between two search models: text similarity model and graph calculation model. It is observed in the experiment that text similarity model has certain advantages in the aspect of full enquiry investigation. Full-text indexing works well for entity discovery and alignment in large-scale triple data. In the solving process, full-text indexing combined with synonyms can get the effective primary selection of answer sets and narrow the screen space of the result sets. The graph calculation perceptron model effectively utilizes the association features between nodes, when combined with the entity and problem type information, it can effectively determine the scope of answer entity. In the process of solving complex problems, both models use the link mode between entities in the graph structure to nd the best matching path, and extract answer entities according to problem type and object. The results of the two models are consolidated and integrated according to entity type and path, which can improve the results.
Meanwhile, it is found that some problems can be improved in the optimization process of models: 1) the performance of similar dynamic solution space generation and pre-screening calculation processing is low; 2) a more optimized learning style can be adopted in the weight parameter optimization of similarity values of the entities with different roles in the link mode between entities. In addition, how the relative relationship between entity and predicate in the syntactic parse tree can be further explored as a feature of assisted optimization.