Application of knowledge graphs to question answering systems

Application of knowledge graphs to question answering systems

A knowledge graph can be defined as “a graph created by describing entities and the relationships among them. Entities” here are things that “exist” physically or non-physically and are not necessarily material entities, but are abstracted to represent things (events in mathematics, law, academic fields, etc.).

Examples of knowledge graphs include simple and concrete things such as “there is a pencil on the table” and “Mt. Fuji is located on the border between Shizuoka and Yamanashi prefectures,” as well as more abstract things such as “if a=b, then a+c = b+c,” “the consumption tax is an indirect tax that focuses on “consumption” of goods and services,” “in electronically controlled fuel injection systems In the case of electronically controlled fuel injection systems, the throttle chamber is an intake throttling device that is attached to the collector of the intake manifold and contains a throttle valve that controls the amount of intake air.

The advantage of using these knowledge graphs, from AI’s perspective, is that machines can access the rules, knowledge, and common sense of the human world through the data in the knowledge graphs. In contrast to the recent black-box approaches, such as deep learning, which require a large amount of teacher data in order to achieve learning accuracy, AI can produce results that are easy for humans to interpret, and AI can generate data based on knowledge data to enable machine learning with small data. Machine learning with small data is possible by generating data based on knowledge data.

By applying this knowledge graph to a question-answer system, it is possible to create a hierarchical structure of key terms, rather than simple FAQ question-answer pairs, and further associate them with context-specific questions and their alternatives, synonyms, and machine-learned response classes, to provide an intelligent FAQ experience. This enables the creation of an intelligent FAQ experience.

The application of this knowledge graph to question and answer systems has been the subject of a number of papers in recent years. The following is a summary of the papers presented at ISWC on the application of the knowledge graph to question answering systems.

• A Survey on Complex Question Answering over Knowledge Base: Recent Advances and Challenges

Question Answering (QA) over Knowledge Base (KB) aims to automatically answer natural language questions via well-structured relation information between entities stored in knowledge bases. In order to make KBQA more applicable in actual scenarios, researchers have shifted their attention from simple questions to complex questions, which require more KB triples and constraint inference. In this paper, we introduce the recent advances in complex QA. Besides traditional methods relying on templates and rules, the research is categorized into a taxonomy that contains two main branches, namely Information Retrieval-based and Neural Semantic Parsing-based. After describing the methods of these branches, we analyze directions for future research and introduce the models proposed by the Alime team.

• EARL:joint entity and relation linking for question answering

Relation linking is an important problem for knowledge graph-based Question Answering. Given a natural language question and a knowledge graph, the task is to identify relevant relations from the given knowledge graph. Since existing techniques for entity extraction and link- ing are more stable compared to relation linking, our idea is to exploit entities extracted from the question to support relation linking. In this paper, we propose a novel approach, based on DBpedia entities, for computing relation candidates. We have empirically evaluated our approach on different standard benchmarks. Our evaluation shows that our approach significantly outperforms existing baseline systems in both recall, precision and runtime.

• Pretrained transformers for simple question answering over knowledge graph

Answering simple questions over knowledge graphs is a well-studied problem in question answering. Previous approaches for this task built on recurrent and convolutional neural network based architectures that use pretrained word embeddings. It was recently shown that finetuning pretrained transformer networks (e.g. BERT) can outperform previous approaches on various natural language processing tasks. In this work, we investigate how well BERT performs on SimpleQuestions and provide an evaluation of both BERT and BiLSTM-based models in datasparse scenarios.

• Difficulty-controllable multi-hop question generation

Knowledge graphs have become ubiquitous data sources and their utility has been amplified by the research on ability to answer carefully crafted questions over knowledge graphs. We investigate the problem of question generation (QG) over knowledge graphs wherein, the level of difficulty of the question can be controlled. We present an end-to-end neural network-based method for automatic generation of complex multi-hop questions over knowledge graphs. Taking a subgraph and an answer as input, our transformer-based model generates a natural language question. Our model incorporates difficulty estimation based on named entity popularity, and makes use of this estimation to generate difficulty-controllable questions. We evaluate our model on two recent multi-hop QA datasets. Our evaluation shows that our model is able to generate high-quality, fluent and relevant questions. We have released our curated QG dataset and code at https://github.com/liyuanfang/mhqg.

• knowlyBERT- hybrid query answering over language models and knowledge graph

Providing a plethora of entity-centric information, Knowledge Graphs have become a vital building block for a variety of intelligent applications. Indeed, modern knowledge graphs like Wikidata already capture several billions of RDF triples, yet they still lack a good coverage for most relations. On the other hand, recent developments in NLP research show that neural language models can easily be queried for relational knowledge without requiring massive amounts of training data. In this work, we leverage this idea by creating a hybrid query answering system on top of knowledge graphs in combination with the masked language model BERT to complete query results. We thus incorporate valuable structural and semantic information from knowledge graphs with textual knowledge from language models to achieve high precision query results. Standard techniques for dealing with incomplete knowledge graphs are either (1) relation extraction which requires massive amounts of training data or (2) knowledge graph embeddings which have problems to succeed beyond simple baseline datasets. Our hybrid system KnowlyBERT requires only small amounts of training data, while outperforming state-of-the-art techniques by boosting their precision by over 30% in our large Wikidata experiment.

• PNEL pointer Network based end to end entity linking over knowledge graph

Question Answering systems are generally modelled as a pipeline consisting of a sequence of steps. In such a pipeline, Entity Linking (EL) is often the first step. Several EL models first perform span detection and then entity disambiguation. In such models errors from the span detection phase cascade to later steps and result in a drop of overall accuracy. Moreover, lack of gold entity spans in training data is a limiting factor for span detector training. Hence the movement towards end-to-end EL models began where no separate span detection step is involved. In this work we present a novel approach to end-to-end EL by applying the popular Pointer Network model, which achieves competitive performance. We demonstrate this in our evaluation over three datasets on the Wikidata Knowledge Graph.

• Unscripted conversation through knowledge graph

In this paper, we introduce “unscripted conversation” – free form dialog over a domain knowledge graph. We describe a use case around Luggage handling for a commercial airline where we answer users queries regarding various policies such as luggage dimensions, restrictions on carry-on items, travel routes etc. We have encoded the domain entities, relationships, processes and polices in the knowledge graph and created a generic semantic natural language processing engine to process user queries and retrieve the correct results from a knowledge graph.

• Leveraging semantics parsing for relation linking over knowledge graph

Knowledgebase question answering systems are heavily dependent on relation extraction and linking modules. However, the task of extracting and linking relations from text to knowledgebases faces two primary challenges; the ambiguity of natural language and lack of training data. To overcome these challenges, we present SLING, a relation linking framework which leverages semantic parsing using Abstract Meaning Representation (AMR) and distant supervision. SLING integrates multiple relation linking approaches that capture complementary signals such as linguistic cues, rich semantic representation, and information from the knowledgebase. The experiments on relation linking using three KBQA datasets; QALD-7, QALD-9, and LC-QuAD 1.0 demonstrate that the proposed approach achieves state-of-the-art performance on all benchmarks

• Enriching knowledge bases with interesting negative statement

Knowledge bases (KBs), pragmatic collections of knowledge about notable entities, are an important asset in applications such as search, question answering and dialogue. Rooted in a long tradition in knowledge representation, all popular KBs only store positive information, but abstain from taking any stance towards statements not contained in them. In this paper, we make the case for explicitly stating interesting statements which are not true. Negative statements would be important to overcome current limitations of question answering, yet due to their potential abundance, any effort towards compiling them needs a tight coupling with ranking. We introduce two approaches towards automatically compiling negative statements. (i) In peer-based statistical inferences, we compare entities with highly related entities in order to derive potential negative statements, which we then rank using supervised and unsupervised features. (ii) In pattern-based query log extraction, we use a pattern-based approach for harvesting search engine query logs. Experimental results show that both approaches hold promising and complementary potential. Along with this paper, we publish the first datasets on interesting negative information, containing over 1.4M statements for 130K popular Wikidata entities.

• Knowledge based question answering for DIYERS

DIY (Do-It-Yourself) requires extensive knowledge such as the usage of tools, properties of materials, and the procedure of activities. Most DIYers use online search to find information but it is usually time-consuming and challenging for novice DIYers to understand the retrieved results and later apply them to their individual DIY tasks. In the work, we present a Question Answering (QA) system which can address the DIYers’ specific needs. The core component is a knowledge base (KB) which contains a vast amount of domain knowledge encoded in a knowledge graph. The system is able to explain how the answers are derived with reasoning process. Our user study shows that the QA system addresses DIYers’ needs more effectively than the web search.

• deqa: Deep Web Extraction for Question Answering

From an ISWC paper. “Despite decades of effort, intelligent object search has yet to be realized. Search engines and Semantic Web technologies alone cannot provide usable systems for simple questions such as, “Tell me about a property near a supermarket that has a garden and at least two bedrooms. In this paper, we introduce deqa, a conceptual framework that achieves this challenging goal by combining state-of-the-art semantic technology with effective data extraction. deqa is realized by mapping natural language questions into Sparql patterns. These patterns are then evaluated on an RDF database of current real estate offers. The offers are obtained from major real estate companies in the Oxford area using OXPath, a state-of-the-art data extraction system, and linked to the background via Limes. They provide knowledge of things like the location of supermarkets.”

• Anu Question Answering System

AnuQA is a question answering system built on top of a search index and an enterprise knowledge graph. In this work, we describe five semantic technologies that have helped us address real world challenges in deploying this system. These challenges include bias in knowledge base population, entity re-resolution on streaming data, ontology alignment across data sources, explaining relationships, and providing a single unified query interface for business analytics.

• A Methodology for Hierarchical Classification of Semantic Answer Types of Questions

Question answering systems have recently been integrated with many smart devices and search engines. Answer type prediction plays an important role in question answering systems as it can help filter irrelevant results and improve overall search and retrieval performance. Here, we present our approach for answer type prediction using the datasets provided for the International Semantic Web Conference (ISWC 2020) SMART Task Challenge. Predicting granular answer types for a question from a big knowledge graph is a greater challenge due to the large number of possible classes. Thus, we propose a 3-step approach to tackle the challenge task. We start with building a classifier that predicts the category of the types and build another classifier just for resource types. The latter model will predict the most general (frequent) type for each question, ignoring type hierarchy. We use a multi-class text classification algorithm built-in fastai library for these two models. The models’ accuracies are 0.95 and 0.73 for category and generic type classification respectively in the validation set (20% randomly chosen samples) of the DBPedia dataset. Next, we train a third classifier to find more specific types (sub-classes) for each question based on the previous general predicted types. We achieve 0.62 and 0.61 using NDCG@5 and NDCG@10 metrics respectively for the test set.

• Hierarchical Contextualized Representation Models for Answer Type Prediction

SeMantic AnsweR Type prediction (SMART) challenge proposed a task to determine the types of answers given natural language questions. Understanding answer types play a crucial role in question answering. In this paper, we present Hierarchical Contextualized-based models, namely HiCoRe, for the SAMRT task. HiCoRe builds on top of state of the art contextualized-based models and the hierarchical strategy to deal with the hierarchical answer types. The SMART results show that HiCoRe obtains promising performance for answer type prediction on DBpedia and Wikidata datasets.

• Fine and Ultra-File type Embeddings for Quesiton Answering

We describe our system for the SeMantic AnsweR (SMART) Type prediction task 2020 for both the DBpedia and Wikidata Question Answer Type datasets. The SMART task challenge introduced finegrained and ultra-fine entity typing to question answering by releasing two datasets for question classification using DBpedia and Wikidata classes. We propose a flexible approach for both entity types using paragraph vectors and word embeddings to obtain high quality contextualized question representations. We augment the document similarity provided by paragraph vectors with semantic modeling and sentence alignment using word embeddings. For the answer category prediction, we achieved a maximum accuracy score of 85% for Wikidata and 88.5% for DBpedia. For the answer types prediction, we achieved a maximum MRR of 40% for Wikidata and a maximum nDCG@5 of 54.8% for DBpedia datasets

• Question Embeddings for Semantic Answer Type Prediction

This paper considers an answer type and category prediction challenge for a set of natural language questions, and proposes a question answering classification system based on word and DBpedia knowledge graph embeddings. The questions are parsed for keywords, nouns and noun phrases before word and knowledge graph embeddings are applied to the parts of the question. The vectors produced are used to train multiple multi-layer perceptron models, one for each answer type in a multiclass one-vs-all classification system for both answer category prediction and answer type prediction. Different combinations of vectors and the effect of creating additional positive and negative training samples are evaluated in order to find the best classification system. The classification system that predict the answer category with highest accuracy are the classifiers trained on knowledge graph embedded noun phrases vectors from the original training data, with an accuracy of 0.793. The vector combination that producesthe highest NDCG values for answer category accuracy is the word embeddings from the parsed question keyword and nouns parsed from the original training data, with NDCG@5 and NDCG@10 values of 0.471 and 0.440 respectively for the top five and ten predicted answer types.

• CASQAD – A New Dataset For Context-aware Spatial Question Answering 

The task of factoid question answering (QA) faces new challenges when applied in scenarios with rapidly changing context information, for example on smartphones. Instead of asking who the architect of the “Holocaust Memorial” in Berlin was, the same question could be phrased as “Who was the architect of the many stelae in front of me?” presuming the user is standing in front of it. While traditional QA systems rely on static information from knowledge bases and the analysis of named entities and predicates in the input, question answering for temporal and spatial questions imposes new challenges to the underlying methods. To tackle these challenges, we present the Context-aware Spatial QA Dataset (CASQAD) with over 5,000 annotated questions containing visual and spatial references that require information about the user’s location and moving direction to compose a suitable query. These questions were collected in a large scale user study and annotated semi-automatically, with appropriate measures to ensure the quality.

• RuBQ: A Russian Dataset for Question Answering over Wikidata 

The paper presents RuBQ, the first Russian knowledge base question answering (KBQA) dataset. The high-quality dataset consists of 1,500 Russian questions of varying complexity, their English machine translations, SPARQL queries to Wikidata, reference answers, as well as a Wikidata sample of triples containing entities with Russian labels. The dataset creation started with a large collection of question-answer pairs from online quizzes. The data underwent automatic filtering, crowd-assisted entity linking, automatic generation of SPARQL queries, and their subsequent in-house verification.

• PNEL: Pointer Network based End-To-End Entity Linking over Knowledge Graphs. 

Question Answering systems are generally modelled as a pipeline consisting of a sequence of steps. In such a pipeline, Entity Linking (EL) is often the first step. Several EL models first perform span detection and then entity disambiguation. In such models errors from the span detection phase cascade to later steps and result in a drop of overall accuracy. Moreover, lack of gold entity spans in training data is a limiting factor for span detector training. Hence the movement towards end-to-end EL models began where no separate span detection step is involved. In this work we present a novel approach to end-to-end EL by applying the popular Pointer Network model, which achieves competitive performance. We demonstrate this in our evaluation over three datasets on the Wikidata Knowledge Graph

• Learning to Rank Query Graphs for Complex Question Answering over Knowledge Graphs

In this paper, we conduct an empirical investigation of neural query graph ranking approaches for the task of complex question answering over knowledge graphs. We propose a novel self-attention based slot matching model which exploits the inherent structure of query graphs, our logical form of choice. Our proposed model generally outperforms other ranking models on two QA datasets over the DBpedia knowledge graph, evaluated in different settings. We also show that domain adaption and pre-trained language model based transfer learning yield improvements, effectively offsetting the general lack of training data.

• Summarizing News Articles Using Question-and-Answer Pairs via Learning

The launch of the new Google News in 2018 introduced the Frequently asked questions feature to structurally summarize the news story in its full coverage page. While news summarization has been a research topic for decades, this new feature is poised to usher in a new line of news summarization techniques. There are two fundamental approaches: mining the questions from data associated with the news story and learning the questions from the content of the story directly. This paper provides the first study, to the best of our knowledge, of a learning based approach to generate a structured summary of news articles with question and answer pairs to capture salient and interesting aspects of the news story. Specifically, this learning-based approach reads a news article, predicts its attention map (i.e., important snippets in the article), and generates multiple natural language questions corresponding to each snippet. Furthermore, we describe a mining-based approach as the mechanism to generate weak supervision data for training the learning based approach. We evaluate our approach on the existing SQuAD dataset (https://rajpurkar.github.io/SQuAD-explorer/.) and a large dataset with 91K news articles we constructed. We show that our proposed system can achieve an AUC of 0.734 for document attention map prediction, a BLEU-4 score of 12.46 for natural question generation and a BLEU-4 score of 24.4 for question summarization, beating state-of-art baselines.

• QaldGen: Towards Microbenchmarking of Question Answering Systems over Knowledge Graphs

Over the last years, a number of Knowledge Graph (KG) based Question Answering (QA) systems have been developed. Consequently, the series of Question Answering Over Linked Data (QALD1–QALD9) challenges and other datasets have been proposed to evaluate these systems. However, the QA datasets contain a fixed number of natural language questions and do not allow users to select micro benchmarking samples of the questions tailored towards specific use-cases. We propose QaldGen, a framework for microbenchmarking of QA systems over KGs which is able to select customised question samples from existing QA datasets. The framework is flexible enough to select question samples of varying sizes and according to the user-defined criteria on the most important features to be considered for QA benchmarking. This is achieved using different clustering algorithms. We compare state-of-the-art QA systems over knowledge graphs by using different QA benchmarking samples. The observed results show that specialised micro-benchmarking is important to pinpoint the limitations of the various QA systems and its components.

• EARL: Joint Entity and Relation Linking for Question Answering over Knowledge Graphs

Many question answering systems over knowledge graphs rely on entity and relation linking components in order to connect the natural language input to the underlying knowledge graph. Traditionally, entity linking and relation linking has been performed either as a dependent, sequential tasks or as independent, parallel tasks. In this paper, we propose a framework called EARL, which performs entity linking and relation linking as a joint task. EARL implements two different solution strategies for which we provide a comparative analysis in this paper: The first strategy is a formalization of the joint entity and relation linking tasks as an instance of the Generalised Travelling Salesman Problem (GTSP). In order to be computationally feasible, we employ approximate GTSP solvers. The second strategy uses machine learning in order to exploit the connection density between nodes in the knowledge graph. It relies on three base features and re-ranking steps in order to predict entities and relations. We compare the strategies and evaluate them on a dataset with 5000 questions. Both strategies significantly outperform the current state-of-the-art approaches for entity and relation linking.

• What is the cube root of 27? Question Answering over CodeOntology

We present an unsupervised approach to process natural language questions that cannot be answered by factual question answering nor advanced data querying, requiring ad-hoc code generation instead.
To address this challenging task, our system, AskCO, performs language-to-code translation by interpreting the natural language question and generating a SPARQL query that is run against CodeOntology, a large RDF repository containing millions of triples representing Java code constructs. The SPARQL query will result in a number of candidate Java source code snippets and methods, ranked by AskCO on both syntactic and semantic features, to find the best candidate, that is then executed to get the correct answer. The evaluation of the system is based on a dataset extracted from StackOverflow and experimental results show that our approach is comparable with other state-of-the-art proprietary systems, such as the closed-source WolframAlpha computational knowledge engine.

• QA4IE: A Question Answering based Framework for Information Extraction

Information Extraction (IE) refers to automatically extracting structured relation tuples from unstructured texts. Common IE solutions, including Relation Extraction (RE) and open IE systems, can hardly handle cross-sentence tuples, and are severely restricted by limited relation types as well as informal relation specifications (e.g., free-text based relation tuples). In order to overcome these weaknesses, we propose a novel IE framework named QA4IE, which leverages the flexible question answering (QA) approaches to produce high quality relation triples across sentences. Based on the framework, we develop a large IE benchmark with high quality human evaluation. This benchmark contains 293K documents, 2M golden relation triples, and 636 relation types. We compare our system with some IE baselines on our benchmark and the results show that our system achieves great improvements.

• Can children teach AI? Towards expressive human-AI dialogs

AI-empowered dialog systems become increasingly widespread, with over 40 million units of Amazon Alexa and Google Home installed alone in US in 2017, according to recent surveys. Whereas this adoption facilitates AI-based systems to engage in end user dialogs on an unprecedented scale, the ability of AI to learn from their human dialog partners is still substantially limited and mostly restricted to pre-defined feedback options and, eventually, basic data labeling. In this paper we discuss opportunities and challenges on the interface of knowledge representation and human computer interaction towards more expressive dialogs between end users and AI dialog systems facilitating AI systems to directly gain knowledge from their end users.

• AMUSE: Multilingual Semantic Parsing for Question Answering over Linked Data

The task of answering natural language questions over RDF data has received wIde interest in recent years, in particular in the context of the series of QALD benchmarks. The task consists of mapping a natural language question to an executable form, e.g. SPARQL, so that answers from a given KB can be extracted. So far, most systems proposed are i) monolingual and ii) rely on a set of hard-coded rules to interpret questions and map them into a SPARQL query. We present the first multilingual QALD pipeline that induces a model from training data for mapping a natural language question into logical form as probabilistic inference. In particular, our approach learns to map universal syntactic dependency representations to a language-independent logical form based on DUDES (Dependency-based Underspecified Discourse Representation Structures) that are then mapped to a SPARQL query as a deterministic second step. Our model builds on factor graphs that rely on features extracted from the dependency graph and corresponding semantic representations. We rely on approximate inference techniques, Markov Chain Monte Carlo methods in particular, as well as Sample Rank to update parameters using a ranking objective. Our focus lies on developing methods that overcome the lexical gap and present a novel combination of machine translation and word embedding approaches for this purpose. As a proof of concept for our approach, we evaluate our approach on the QALD-6 datasets for English, German & Spanish.

• A Corpus for Complex Question Answering over Knowledge Graphs

Being able to access knowledge bases in an intuitive way has been an active area of research over the past years. In particular, several question answering (QA) approaches which allow to query RDF datasets in natural language have been developed as they allow end users to access knowledge without needing to learn the schema of a knowledge base and learn a formal query language. To foster this research area, several training datasets have been created, e.g.~in the QALD (Question Answering over Linked Data) initiative. However, existing datasets are insufficient in terms of size, variety or complexity to apply and evaluate a range of machine learning based QA approaches for learning complex SPARQL queries. With the provision of the Large-Scale Complex Question Answering Dataset (LC-QuAD), we close this gap by providing a dataset with 5000 questions and their corresponding SPARQL queries over the DBpedia dataset. In this article, we describe the dataset creation process and how we ensure a high variety of questions, which should enable to assess the robustness and accuracy of the next generation of QA systems for knowledge graphs.

• Tractable Query Answering for Expressive Ontologies and Existential Rules

The disjunctive skolem chase is a sound and complete (albeit non-terminating) algorithm that can be used to solve conjunctive query answering over DL ontologies and programs with disjunctive existential rules. Even though acyclicity notions can be used to ensure chase termination for a large subset of real-world knowledge bases, the complexity of reasoning over acyclic theories still remains high. Hence, we study several restrictions which not only guarantee chase termination but also ensure polynomiality. We include an evaluation that shows that almost all acyclic DL ontologies do indeed satisfy these general restrictions.

• Temporal Query Answering in DL-Lite over Inconsistent Data

In ontology-based systems that process data stemming from different sources and that is received over time, as in context-aware systems, reasoning needs to cope with the temporal dimension and should be resilient against inconsistencies in the data. Motivated by such settings, this paper addresses the problem of handling inconsistent data in a temporal version of ontology-based query answering. We consider a recently proposed temporal query language that combines conjunctive queries with operators of propositional linear temporal logic and extend to this setting three inconsistency-tolerant semantics that have been introduced for querying inconsistent description logic knowledge bases. We investigate their complexity for DL-Lite R temporal knowledge bases, and furthermore complete the picture for the consistent case.

• A Practical Acyclicity Notion for Query Answering Over Horn-SRIO-Ontologies

Conjunctive query answering over expressive Horn Description Logic ontologies is a relevant and challenging problem which, in some cases, can be addressed by application of the chase algorithm. In this paper, we define a novel acyclicity notion which provides a sufficient condition for termination of the restricted chase over Horn-

SRIQ TBoxes. We show that this notion generalizes most of the existing acyclicity conditions (both theoretically and empirically). Furthermore, this new acyclicity notion gives rise to a very efficient reasoning procedure. We provide evidence for this by providing a materialization based reasoner for acyclic ontologies which outperforms other state-of-the-art systems.

• Zero-shot Visual Question Answering using Knowledge Graph

Incorporating external knowledge to Visual Question Answering (VQA) has become a vital practical need. Existing methods mostly adopt pipeline approaches with different components for knowledge matching and extraction, feature learning, etc.However, such pipeline approaches suffer when some component does not perform well, which leads to error propagation and poor overall performance. Furthermore, the majority of existing approaches ignore the answer bias issue — many answers may have never appeared during training (i.e., unseen answers) in real-word application. To bridge these gaps, in this paper, we propose a Zero-shot VQA algorithm using knowledge graphs and a mask-based learning mechanism for better incorporating external knowledge, and present new answer-based Zero-shot VQA splits for the F-VQA dataset. Experiments show that our method can achieve state-of-the-art performance in Zero-shot VQA with unseen answers, meanwhile dramatically augment existing end-to-end models on the normal F-VQA task.

• KnowlyBERT – Hybrid Query Answering over Language Models and Knowledge Graphs

Providing a plethora of entity-centric information Knowledge Graphs have become a vital building block for a variety of intelligent applications. Indeed, modern knowledge graphs like Wikidata already capture several billions of RDF triples, yet they still lack a good coverage for most relations. On the other hand, recent developments in NLP research show that neural language models can easily be queried for relational knowledge without requiring massive amounts of training data. In this work, we leverage this idea by creating a hybrid query answering system on top of knowledge graphs in combination with the masked language model BERT to complete query results. We thus incorporate valuable structural and semantic information from knowledge graphs with textual knowledge from language models to achieve high precision query results. Standard techniques for dealing with incomplete knowledge graphs are either (1) relation extraction which requires massive amounts of training data or (2) knowledge graph embeddings which have problems to succeed beyond simple baseline datasets. Our hybrid system KnowlyBERT requires only small amounts of training data, while outperforming stateof-the-art techniques by boosting their precision by over 30% in our large Wikidata experiment.

• Augmentation-based Answer Type Classification of the SMART dataset 

Recent progress in deep-learning-enabled AI researchers and developers to invest minimal efforts to achieve state-of-the-art results. Specifically, in such a task as text classification – text preprocessing and feature generation does not play a significant role anymore thanks to such a landmark model as BERT and other related models. In this paper, we present our solution for the Semantic Answer Type prediction task (SMART task). The solution is based on the application of several data augmentation techniques: machine translation to popular languages, round-trip translation, named entities annotation with linked data. The final submission was generated as a weighted result from several successful system outputs.

• Semantic Answer Type Prediction Using BERT 

This paper summarizes our participation in the SMART Task of the ISWC 2020 Challenge. A particular question we are interested in answering is how well neural methods, and specifically transformer models, such as BERT, perform on the answer type prediction task compared to traditional approaches. Our main finding is that coarse-grained answer types can be identified effectively with standard text classification methods, with over 95% accuracy, and BERT can bring only marginal improvements. For fine-grained type detection, on the other hand, BERT clearly outperforms previous retrieval-based approaches.

• Two-stage Semantic Answer Type Prediction for QA using BERT and Class-Specificity Rewarding

Answer type prediction is a key task in Question Answering (QA) that aims at predicting the type of the expected answer for a user query expressed in natural language. In this paper we focus on semantic answer type prediction where the candidate types come from a class hierarchy of a general-purpose ontology. We model the problem as a two-stage pipeline of sequence classification tasks (answer category prediction, answer literal/resource type prediction), each one making use of a fine-tuned BERT classifier. To cope with the harder problem of answer resource type prediction, we enrich the BERT classifier with a rewarding mechanism that favors the more specific ontology classes that are low in the class hierarchy. The results of an experimental evaluation using the DBpedia class hierarchy (∼760 classes) demonstrate a superior performance of answer category prediction (∼96% accuracy) and literal type prediction (∼99% accuracy), and a satisfactory performance of resource type prediction (∼78% lenient NDCG@5).

• COALA – A Rule-Based Approach to Answer Type Prediction 

For answering a question correctly, the previous detection of the answer type is essential. Especially in the field of Question Answering (QA) over knowledge bases, answers might be of many different types as natural language is ambiguous and a question might lead to different relevant queries. For semantic knowledge bases data types (such as date, string, or number) as well as all ontology classes (such as athlete, championship, or television show) have to be taken into account. Therefore, the previous detection of the answer type is a helpful sub-task for QA systems, but also a complex classification problem. We present our rulebased approach COntext Aware anaLysis of Answer types (COALA). Our approach is based on the extraction of several question features and the context aware disambiguation to retrieve the correct answer type. COALA has been developed in the course of the SMART task challenge and we evaluated our approach based on over 21,000 questions.

• Unscripted Conversation driven through Knowledge Graph

In this paper, we introduce “unscripted conversation” – free form dialog over a domain knowledge graph. We describe a use case around Luggage handling for a commercial airline where we answer users queries regarding various policies such as luggage dimensions, restrictions on carry-on items, travel routes etc. We have encoded the domain entities, relationships, processes and polices in the knowledge graph and created a generic semantic natural language processing engine to process user queries and retrieve the correct results from a knowledge graph.