Abstract
This paper presents an end-to-end neural network model, named Neural Generative Question Answering (GENQA), that can generate answers to simple factoid questions, based on the facts in a knowledge-base. More specifically, the model is built on the encoder-decoder framework for sequence-to-sequence learning, while equipped with the ability to enquire the knowledge-base, and is trained on a corpus of question-answer pairs, with their associated triples in the knowledge-base. Empirical study shows the proposed model can effectively deal with the variations of questions and answers, and generate right and natural answers by referring to the facts in the knowledge-base. The experiment on question answering demonstrates that the proposed model can outperform an embedding-based QA model as well as a neural dialogue model trained on the same data.
Introduction
Traditional Question Answering (QA) systems typically operate within two distinct paradigms. Retrieval-based systems are proficient at extracting facts from structured databases but often lack the linguistic flexibility to produce natural, conversational responses. Conversely, neural generative models, such as those used in modern dialogue systems, excel at mimicking human speech patterns but frequently fail to accurately recall specific, discrete facts—a phenomenon often referred to as "hallucination."
The fundamental challenge lies in the discrepancy between the "smooth" representations of neural networks and the "discrete" nature of world knowledge. To bridge this gap, researchers from Peking University and Huawei Noah’s Ark Lab have developed GENQA, an end-to-end neural network that generates natural language answers by dynamically querying an external knowledge base (KB).
The GENQA Model Architecture
GENQA utilizes an encoder-decoder framework augmented with a specialized "Enquirer" module to facilitate external knowledge retrieval. The architecture consists of three core components:
- The Interpreter (Encoder): Employs a bi-directional Gated Recurrent Unit (GRU) to transform natural language questions into high-dimensional vector representations.
- The Enquirer: Acts as the system's search mechanism. It identifies candidate facts (triples consisting of subject, predicate, and object) from the KB and calculates relevance scores using either a Bilinear Model or a CNN-based Matching Model.
- The Answerer (Decoder): A novel component that uses a mixture model to generate answers word-by-word.
Hybrid Decoding Strategy
The Answerer module is the centerpiece of GENQA's ability to blend natural language with factual data. At each decoding step, the model employs a "soft switch" mechanism to decide between two sources for the next word:
- Common Vocabulary: Words like "is," "born," or "in" are generated from the model's internal language model.
- KB-word: Specific entities or values (e.g., "2.29m," "Germany") are retrieved by the Enquirer from the knowledge base.
This hybrid approach allows the model to handle "unseen" facts during inference, provided it has learned the underlying relationship requested by the question. The task of generating a natural sentence actually helps the model learn better word-embeddings for the questions, improving the accuracy of the fact retrieval itself.
Experimental Setup and Results
The model was trained and evaluated on a massive dataset comprising approximately 11 million KB triples and 235 million QA pairs extracted from Chinese community sites. A specific training set of 696,306 instances was created where questions were explicitly linked to KB triples.
Empirical results demonstrate that GENQA significantly outperforms baseline models. While a standard neural dialogue model (NRM) achieved only 19% accuracy by attempting to "guess" answers from training patterns, GENQA reached 52% accuracy using a CNN-based enquirer. Qualitative analysis further confirms that the model successfully integrates KB facts into fluent sentences, such as "He is 2.29m and is visible from space."
Conclusion
GENQA represents a significant step toward combining the linguistic power of neural networks with the factual precision of structured knowledge. By treating the knowledge base as an external reference rather than a memory task, the model demonstrates robust generalization to facts not seen during training. Future research will focus on extending this framework to multi-turn dialogues and complex queries requiring multi-step reasoning across knowledge graphs.
Summary updated