Keywords

1 Introduction

Knowledge graph (KG) is one of key technologies for implementing an intelligent question and answer service (QA), which can reduce customers’ service costs and improve their self-service capabilities. Although the current methods based on natural language processing can well deal with the problems of context semantics and language fuzziness, the existing QA services based on KG is still far from industrial applications. Firstly, current methods retrieve information with neural network and dialogue model to improve the efficiency of constructing KGs, but they cannot well utilize and merge knowledge in different domains. Furthermore, current methods deal with the shortage knowledge in KG by means of context semantic association [1], web page search [2] and external text [3], but the lack of external knowledge and the automatic integration of KGs lead to a high cost of operating and managing online services and cannot well satisfy customers’ requirements.

To address the above issues, this paper proposes a Knowledge graph-based intelligent question-answering service AIServiceX. Firstly, we propose an entity/assertion recognition method based on a gate recurrent unit (GRU) model combined with the attention model and the dual-phase connection to achieve the accurate recognition of natural language’s expression. Secondly, we propose a semantic enhancement method based on the topic comparison, which considers the loss of two squares and a topic model, and thus the domain knowledge mapped to a local KG supports the online update of a knowledge base. Meanwhile, we employ heuristic query rules to sort the candidate set by combining questions’ features and KG’s status, so that we can easily select the best candidate answers. Finally, we employ the optimization strategy based on the integer linear programming (ILP) to update the local KG according to the users’ satisfaction and services’ cost. The GRU model and semantic enhancement method can realize the construction and expansion of KG and have the flexible online service capability. Heuristic rules and the ILP optimization strategy can support the retrieval and update of KG in AIServiceX and ensure its availability.

2 Methodology

We take questions expressed in Chinese language as input and the answers of matched questions expressed in the KG as output. AIServiceX receiving requests to answer questions includes six phases as follows.

Question input (Phase 1). AIServiceX receives users’ questions from different interfaces, and then transmits these questions to Phase 2 with identity authentication. Different users may have different expressions for similar questions in the input. Each user has his own question intentions or expected answers, so the forwarded questions is uncertain and fuzzy;

GRU identification (Phase 2). AIServiceX receives questions forwarded by Phase 1, and then identifies the entities/assertions from the received texts via word segmentation, which are the query condition of Phase 3. AIServiceX overcomes the fuzziness of Phase 1 through the GRU’s dual-phase connection and attention weight. Thus, AIServiceX outputs more accurate recognition results described as a target vector reflecting users’ main attention;

Tendency analysis (Phase 3). AIServiceX takes the entity/assertion vector identified in Phase 2 as input, queries keywords from different knowledge bases to find all relevant texts of the questions, and then takes the text with the highest matching degree as input of the next phase. AIServiceX takes historical data of various knowledge bases to construct a tendency analysis model based on a topic model, and then outputs the best matched text through combining the maximum likelihood estimation and the least square loss;

KG enhancement (Phase 4). AIServiceX enhances the original KG by adding edges and points to construct an enhanced KG based query database. This phase integrates the KG and external texts to automatically update KG for online service, and thus continuously improves the knowledge quantity and the expression ability of KG;

Heuristic query (Phase 5). AIServiceX queries KG to obtain an answer candidate set ac-cording to the entity/assertion vectors of the updated global KG from Phase 4, sorts the answers by the counting historical accesses, provides the answers with the highest score, and then records the status of this questioning and answering process. The heuristic rules consider the questions and the overall status of KG;

Periodic update (Phase 6). AIServiceX periodically pulls and normalizes the status of all questions in Phase 5, and quantifies the response time and users’ satisfaction with the integer linear programming model to achieve the minimum of the KG update cost.

3 Algorithm

According to the above methodology, Table 1 describes the algorithm of answering questions in detail as follows. AIServiceX trains the GRU model for each entity to minimize the entity identification loss (lines 1–5); trains the GRU model for each assertion to minimize the assertion identification loss (lines 6–10); gets entities and assertions from the trained GRU model’s outputs according to input questions (line 11); queries knowledge bases with the above output to obtain enhanced text of each knowledge base (line 12); employs the topic model and the quadratic loss to find the most similar text, and then constructs the local KG (lines 13–17); uses heuristic rules to sort answer candidate sets (lines 18–23); regularly updates the global KG (lines 24–33); initializes ILP’s input (lines 25–29); performs ILP solution and merges local KGs according to the results (lines 30–33).

Table 1. The algorithm of answering questions.
Full size table

4 Design

Figure 1 shows the service architecture of AIServiceX including an entrance layer, a resource access layer, a KG service layer and a knowledge base layer.

Fig. 1.
figure 1

System architecture of AIServiceX

Full size image

Entrance layer. AIServiceX and other applications share the same traffic entry with a unified LDAP authentication for permission. Thus, AIServiceX can implement the integrated management for multiple services to avoid confusion in resource utilization and security issues;

KG service layer. AIServiceX implements key technologies in a microservice architecture (i.e., SpringCloud), where the GRU component is connected to a TensorFlow cluster that scales independently; the enhanced semantic querier uses the interface of existing applications for data exchange to avoid complicated logics and adaptation; the local KG builder accesses a lexical processing tool Jieba [4] to implement semantic operations such as word segmentation; the periodic update process interacts with graph database Neo4j [5], and decouples the IIL solver through database operations;

Resource access layer. AIServiceX adopts a typical distributed service architecture including analyzer, database and resource schedulers, and interacts with the KG model service layer through standard Restful interfaces;

Knowledge base layer. AIServiceX interacts with multiple knowledge bases through the unique adaptation of reflection services. All adaptation codes are not diffused to avoid abnormal propagation caused by linking interruption and inconsistent events;

5 Evaluation

5.1 Experiment Setting

We adopt a real business dataset of an e-commerce company. We pre-process and label the data, construct the KG, and then train the GRU model. Existing works often employ the precision, recall and F1 measure to evaluate methods, but they only count the number of binary problems. We extend the evaluation metrics by employing the matching degree of entities [11]. By counting queried answers, labeled correct answers, and the entities in each answer, we can accurately get the values of precision, recall and F1, respectively.

5.2 Heuristic Rules-Based Answer Ranking

We describe the similarities as “count” and “cos”, and the random sort that returns the first result meeting the retrieval rules as “random”. We sort the query results of entity assertions and KG on the test dataset to validate the effects. The experiment compares the results of QA-KG [11], CAN for QA [12] and CKB [7] We validate the performance improvement by comparing the output answers and the labeled correct answers with the distribution of entity assertions. Figure 2 shows that combining two heuristic rules can achieve better results than a single neural network or a dialogue model. Compared with single state-of-the-art methods, our heuristic rules can significantly improve the precision, recall and F1 indexes.

Fig. 2.
figure 2

Comparison with heuristic rules-based answer ranking

Full size image
Fig. 3.
figure 3

Comparison with ILP-based answer ranking

Full size image

5.3 Integer Linear Programming-Based Answer Ranking

The experiment compares the precision, recall and the average response time (seconds) of the full update strategy and the ILP strategy. Figure 3 shows the comparison of TAQA [1], CQA [2] and Text2KB [8]. After adopting different updating strategies, the ILP method brings a slight decrease of precision by 17%, and the decrease of average response time by 9%, which indicates that the space of answer retrieval is relatively concentrated, causing a low processing cost.

6 Related Work

Answering complex questions often involves the analysis of multiple entities and assertions, and requires as a long information retrieval chain. QUINT [6] automatically generates a question template to describe the mapping between questions and answers. CKB [7] defines a dialogue model collecting key elements to obtain accurate answers by guiding users to supplement semantic information. TAQA [1] proposes an N-tuple assertion model to deal with complex semantic constraints introduced by adjectives, verbs, and prepositions in complex questions. Seq2seq [9] employs the overlapping method to connect, duplicate and segment the questions and answers in large datasets. TEQUILA [10] proposes the detection and transformation method of a KG model to eliminate the time constraints of complex time-sensitive questions. Complex question-answering services have high cost and requires in-depth customization of applications, so these methods ought to spend much manual operations on customizing and training KG in the initialized period. AIServiceX accurately recognizes entities, automatically maps domain knowledge to the KG, and online updates the KG.

7 Conclusion

This paper proposes a KG based intelligent QA service integrating domain knowledge. We employ the GRU-based attention model to overcome the fuzziness of domain questions’ expression, topic comparison-based enhanced semantic to construct the local KG and expand the knowledge of the global KG, and ILP update strategy to deal with the dynamic update of KG. AIServiceX can achieve a high precise answering results with a low response delay by accurately recognizing entities, automatically mapping domain knowledge to the KG, and online updating the KG. The experimental results show that AIServiceX improves the precision and decreases the response time.