With the rapid growth of knowledge bases(KBs), knowledgebase question answering has drawn huge attention in recent years. Most existing KBQA methods translate questions into SPARQLs to help endusers access the knowledge base represented by RDF more naturally. However, a natural language question is always corresponding to multiple candidate SPARQLs due to the gap between the unstructured Question and the structural SPARQL query. To pick the best SPARQL query from the candidate SPARQL query set, in this Poster, we propose a hybrid similarity computing method to rank the SPARQL query. Firstly, we employ two attentive recurrent neural networks to capture the semantic Similarity between the SPARQL query and the Question. Secondly, we compute the string similarity between the SPARQL query and the Question by leveraging the convolutional neural network. Our method can capture the two-level Similarity between the Question and the SPARQL. Experiments show that our method can improve the effectiveness of KBQA.
Generally, we will obtain multiple SPARQL queries after the semantic parsing stage since the ambiguity between the natural language question and the knowledge base. For instance, the entity mention St. Lawrence of the Question \What body of water does St.Lawrence ow into? " will be mapped to a set of semantic instances in the KB, e.g., E = fhSiant Lawrencei, or hSiant Lawrence Riverig. Thus, the main challenge in the semantic parsing stage is how to pick the best SPARQL query in the candidate query set.
Most existing KBQA work maps the Question and the KB facts(triple) to a common embedding space. The Similarity between the question vector and the SPARQL vectors can be conveniently computed. However, these methods tend to lose original word interaction information. To preserve more original information, we propose a hybrid similarity computing method to pick the best SPARQL query from the candidate set. Both consider the semantic Similarity and the strong Similarity between the Question and the SPARQL query.
Copyright 2020 for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).
In this Poster, we present a hybrid ranking model to rank the SPARQL query, which considers both string similarity and Semantic Similarity. Given the natural language question N , for each query qi in the candidate SPARQL set, we compute the similarity score S(N; qi) that represents the semantic Similarity between N and qi. Finally, all candidate queries are ranked via their similarity scores with N .
Semantic-level Similarity. We construct an attentive recurrent neural network for computing semantic-level similarity between question N and the candidate query qi. The model uses an encoder-compare framework which encodes the semantic information of N and qi into high-dimensional embedding space and then estimates their similarity via multilayer perceptron(MLP). { Encoding. Firstly, each elements in question N and query qi is mapped to its corresponding embedding vectors fw1; : : : ; wLg, where L is the length of the question or query. And then all the embeddings will be input into a bidirectional GRUs neural network to learn the hidden representations H1:L = [h1; : : : ; hL], where hi is the concatenation of forward and backward vectors learned at time i. Since each word contribute di erently to the full sentence semantics, the model would pay di erent attention to each word and learn promising vectors to represent the question/query sentences. The self-attention model is used here to learn the weight of each word semantic for the input sentence. The semantic representation Y of sentence can be calculated as follows:
Y =
L X aihi; i=1 a = Attention(Q; K; V ) = sof tmax
QK> pdk
V
(
String-level Similarity. The Similarity over string-level is evaluated via a text-matching model. Some words or phrases with the same meaning may be expressed di erently in the Question and query, i.e., the pair of words (musical; music) have similar semantics. Since the high-level semantic embedding cannot preserve these words interaction information, we construct a similarity matrix whose elements represent the similarities between question words and query words, regarding it as a two-dimensional vector space to utilize the convolutional layer to capture the matching features.
{ Similarity matrix. Firstly, we construct a similarity matrix M , where each
element Mij indicates the basic interaction. The Mij can be calculated as:
Mij = ui
vi
Where ui and vj denotes the i-th and the j-th word was embedding in
Question and query, respectively. The operator stands for a general operator to
compute the Similarity. Here the matrix can obtain the Similarity of words
with di erent expressions.
{ Convolution layer. The di erent levels of matching patterns can be extracted
by a convolutional kernel. The k-th kernel wk scans over the similarity matrix
M to generate a feature map gk:
(
Combination. With three feature (z1; z2; z3) generated from two level similarity, we utilize a linear layer to learn their respective contribution for holistic similarity score:
S(p; qi) = Sigmoid W >[z1; z2; z3] + b : (8) Finally, all candidate queries qi will be sorted with the similarity scores S(p; qi).
Our Method BiCNN(Yih et al.) AMPCNN (Wenpeng et al.) HR-BiLSTM (Yu et al.) Multiple View Matching (Yu et al.)
SimpleQuestions is a single-relation KBQA dataset. This dataset consists of questions annotated with a corresponding fact from Freebase that provides the answer. We report Accuracy as previous studies. We verify our proposed approach on the SimpleQuestion dataset. Table 1 summarizes the experimental results of di erent methods on answer selection and knowledge base question answering. Clearly, our method achieves the state-of-the-art results in SimpleQuestions, which con rms the e ectiveness of our solution.
This work is supported by the National Key Research and Development Program of China (2017YFC0908401) and the National Natural Science Foundation of China (61972455). Xiaowang Zhang is supported by the Peiyang Young Scholars in Tianjin University (2019XRX-0032).