moqa a multi modal question answering architecture
play

MoQA A Multi-Modal Question Answering Architecture Monica - PowerPoint PPT Presentation

MoQA A Multi-Modal Question Answering Architecture Monica Haurilet, Ziad Al-Halah and Rainer Stiefelhagen Computer Vision for Human Computer Interaction Lab KIT - Computer Vision for Human Computer Interaction Lab www.kit.edu Multi-Modal


  1. MoQA – A Multi-Modal Question Answering Architecture Monica Haurilet, Ziad Al-Halah and Rainer Stiefelhagen Computer Vision for Human Computer Interaction Lab KIT - Computer Vision for Human Computer Interaction Lab www.kit.edu

  2. Multi-Modal Question Anwering Text … a thick layer containing water is called confined aquifer. The earth region that supports the confined aquifer is called confining bed. The hole to obtain water in the unconfined aquifer is.. … Question What layer is underneath the confined aquifer? Answers a) Unconfined Aquifer b) Confined Aquifer c) Water Table d) Confining Bed  MoQA – A Multi-Modal Question Answering Architecture Monica Haurilet, Ziad Al-Halah and Rainer Stiefelhagen

  3. Definitions  ∃ S Q - a set of sentences or nodes that verifies (Q, A i )  We call S Q set of supporting sentences  S Q is used to get probability of correctness of (Q, A i ) Example … Two other types of mass movement are slump and creep. Both may move a lot of soil and rock. However, they usually aren’t as destructive as landslides and mudslides. Slump is the sudden movement of large blocks of rock and soil down a slope. You can see how it happens in Figure 10.32. All the material moves together in big chunks. … Q: Sudden movement of a large block of rock and soil down a slope .... e) Slump MoQA – A Multi-Modal Question Answering Architecture Monica Haurilet, Ziad Al-Halah and Rainer Stiefelhagen

  4. Overview of our Approach Input  Question Q  Set of possible answers {A i }  Set of sentences or nodes {S j } Our Model 1. Select k supporting sentences 2. Verify answer using a deep neural model MoQA – A Multi-Modal Question Answering Architecture Monica Haurilet, Ziad Al-Halah and Rainer Stiefelhagen

  5. 1. Supporting Sentences  Measure similarity of embedded question and sentences  Select top k most similar sentences S 1 Q 1 S 2 Q 2 S 3 Sentences Questions Embedding Space Question: Study of the solid earth Supporting Sentences: 1. Geology is the study of the solid Earth. 2. Scientists who compare geology of other planets to Earth are planetary geologists. Question: Factors that determine how much erosion runoff can cause include Supporting Sentences: 1. Runoff is an important cause of erosion. 2. Runoff is likely to cause more erosion if the land is bare. MoQA – A Multi-Modal Question Answering Architecture Monica Haurilet, Ziad Al-Halah and Rainer Stiefelhagen

  6. 2. Deep Learning Model  Selects answer based on supporting sentences and question K j-1 Low level clouds cause rain. K j Water is also found in the clouds. CNN Visual Information ... MoQA – A Multi-Modal Question Answering Architecture Monica Haurilet, Ziad Al-Halah and Rainer Stiefelhagen

  7. 2. Deep Learning Model  Selects answer based on supporting sentences and question K j-1 Low level clouds cause rain. K j Water is also found in the clouds. CNN Visual Information A i : Low level. ... Q: Which level of clouds cause rain? Bidir. LSTM Bidir. LSTM [A i , Q, K j , Q· K j , Q·A i , A i · K j , Q· K j ·A i ] FC Kj FC Kj-1 ... FC K1 Softmax MoQA – A Multi-Modal Question Answering Architecture Monica Haurilet, Ziad Al-Halah and Rainer Stiefelhagen

  8. 2. Deep Learning Model  Selects answer based on supporting sentences and question K j-1 Low level clouds cause rain. K j Water is also found in the clouds. CNN Visual Information A i : Low level. ... Q: Which level of clouds cause rain? Bidir. LSTM Bidir. LSTM [A i , Q, K j , Q· K j , Q·A i , A i · K j , Q· K j ·A i ] FC Ai FC Kj FC Kj-1 ... FC K1 · Confidence Softmax MoQA – A Multi-Modal Question Answering Architecture Monica Haurilet, Ziad Al-Halah and Rainer Stiefelhagen

  9. Evaluation MoQA won in the TQA challenge • 1st place in the text track • 2nd place in the diagram track Errors for T/F Questions Validation Accuracy Errors for Diag. Questions MoQA – A Multi-Modal Question Answering Architecture Monica Haurilet, Ziad Al-Halah and Rainer Stiefelhagen

  10. Poster S1 MoQA – A Multi-Modal Question Answering Architecture Monica Haurilet, Ziad Al-Halah, Rainer Stiefelhagen Karlsruhe Institute of Technology, Germany haurilet@kit.edu

Recommend


More recommend