MATQA: Microservice-based Architecture for Temporal Question Answering on Knowledge Graph Embeddings

Vestermark, Kristian Simoni ; Otte, Kristian

4. term

Software, Master

2022

2022-06-16

11

Answering questions with knowledge graphs is an important problem in information retrieval. A special challenge is temporal questions—those that ask when something happened or what was true at a particular time. In this work, we explore temporal question answering by using temporal knowledge graph embeddings (TKGEs), compact numerical representations of knowledge graphs that also encode time. We introduce MATQA, a microservice-based architecture for building temporal QA systems on top of knowledge graph embeddings (KGEs). We also evaluate a variation of ensemble learning called Bayesian model averaging (BMA): we combine and re-rank the results of several link prediction tasks produced by different pre-trained TKGE models before selecting the final answers. Experiments on two datasets, ICEWS14 and ICEWS05-15, implemented through this microservice architecture, show that this ensemble approach provides better results than using individual TKGE models on their own.

At besvare spørgsmål ved hjælp af vidensgrafer er vigtigt inden for informationssøgning. En særlig udfordring er tidsafhængige spørgsmål – for eksempel spørgsmål om, hvornår noget skete, eller hvad der var sandt på et bestemt tidspunkt. I dette arbejde undersøger vi tidslig spørgsmålsbesvarelse ved at bruge tidslige vidensgraf-indlejringer (TKGE), som er kompakte, numeriske repræsentationer af vidensgrafer, der også indkoder tid. Vi præsenterer MATQA, en mikroservice-baseret arkitektur til at bygge tidslige QA-systemer oven på vidensgraf-indlejringer (KGE). Derudover afprøver vi en variation af ensemblelæring kaldet Bayesiansk modelgennemsnit (BMA). Her kombinerer og genrangerer vi resultaterne fra flere forbindelsesforudsigelsesopgaver, der er kørt på forskellige fortrænede TKGE-modeller, før vi vælger de endelige svar. Vores eksperimenter på to datasæt, ICEWS14 og ICEWS05-15, implementeret via denne mikroservice-arkitektur, viser, at ensembletilgangen giver bedre resultater end at bruge hver enkelt TKGE-model for sig.

[This apstract has been rewritten with the help of AI based on the project's original abstract]

Question Answering ; Knowledge Graph ; Knowledge Graph Embeddings ; Temporal Knowledge Graphs ; Temporal Knowledge Graph Embeddings ; Ensemble ; Microservice

Download PDF
View record in AAU Student Projects