1. Building Dynamic Knowledge Graphs From Text Using Machine Reading Comprehension
Rajarshi Das, Tsendsuren Munkhdalai, Xingdi Yuan, Adam Trischler, Andrew McCallum (ICLR’19)
Presented by: Shen Yan

2. Automatically Building Knowledge Graphs
Raw information ⇒ Structured form:
Nodes (entities)
Edges (relationships)
Track the changing relations among entities
Make implicit information more explicit

3. Example:

4. KG-MRC Knowledge Graph - Machine Reading Comprehension
A neural machine-reading model that constructs dynamic knowledge graphs from text
Focus on tracking the Locations of participant entities
Bipartite graph:
Two sets of nodes: entities ( 𝑒 𝑖,𝑡 ) and locations ( 𝜆 𝑖,𝑡 )
Explicitly constructs dynamic knowledge graphs to track state changes in procedural text
Conditions on its own constructed knowledge graphs to improve downstream question answering on the text

5. KG-MRC Pipeline
At each time step 𝑡, reading prefixes of the paragraph up to and including sentence 𝑠 𝑡
Engage Machine reading comprehension (MRC) model to query for the state of each participant entity (e.g., “Where is E located?”). MRC model returns an answer span describing the entities’ current location at 𝑡, encoding the text span as the vector Ψ 𝑖,𝑡
Conditioning on the span vector Ψ 𝑖,𝑡 , the model constructs the graph 𝐺 𝑡 by updating 𝐺 𝑡−1 from the previous time step

6. MRC architecture: DrQA (Chen et al. ACL’17）
Two components: 1. Document retriever (non-machine learning) TF-IDF weighted bag-of-word vectors, wikipeida
2. Document reader (RNN model)

7. Soft Co-reference Across time steps: Within each time step:
Ψ 𝑖,𝑡 : incoming location vector
𝜆′ 𝑖,𝑡 : intermediate node representation
Λ 𝑡−1 = [𝜆 𝑖,𝑡 ] 𝑖=1 𝑁 : matrix of location node representations
Λ′ 𝑡 = [𝜆′ 𝑖,𝑡 ] 𝑖=1 𝑁 : matrix of intermediate node representations
𝑈 𝑡 = [𝑢 𝑖,𝑡 ] 𝑖=1 𝑁 : co-reference adjacency matrix, to track the related nodes within t

8. Graph Update
Compose all connected nodes with their history summary using an LSTM unit
Update node information
Perform a co-reference pooling operation for location node representations
Recurrent graph: Stack L such layers to propagate node information along the graph’s edges.

9. Experiments & Evaluation Procedural text comprehension tasks
Task 1 Sentence-level evaluation (Dalvi et al. 2018)
Answer 3 categories of questions
Cat 1: Is E created, (destroyed, ,moved) in the process?
Cat 2: When (step #) is E created, (destroyed, moved)?
Cat 3: Where is E created, (destroyed, moved from/to)?
Task 2 Document-level evaluation (Tandon et al. 2018)
Answer 4 categories of questions
Cat 1: What are the inputs to the process?
Cat 2: What are the outputs of the process?
Cat 3: What conversions occur, when and where?
Cat 4: What movements occur, when and where?

10. Experiments & Evaluation Procedural text comprehension tasks
PROPARA dataset: procedural text about scientific processes.

11. Experiments & Evaluation Procedural text comprehension tasks
PROPARA dataset

12. Experiments & Evaluation 2. Ablation study
Removing the soft-coreference disambiguation within the steps → 1% performance drop
Removing the soft-coreference across time steps → more significant performance drop
Replace the recurrent graph module with LSTM → lack the information propagation across graph nodes

13. Experiments & Evaluation 3. Commonsense constraints
Commonsense constraints: (Tandon et al. 2018)
An entity must exist before it can be moved or destroyed
An entity cannot be created if it already exists
An entity cannot change until it is mentioned in the paragraph

14. Experiments & Evaluation 4. Qualitative analysis
Tracking the state of entity blood across 6 sentences
Blue: true location
Orange: predicted results from Pro-Local (Dalvi et al. 2018)
Red: predicted results from KG-MRC

15. Conclusions
Proposed a model that constructs dynamic knowledge graphs from text to track locations of participants entities in procedural text.
KG-MRC improves the downstream comprehension of text and achieves state-of-the art results on two question-answering tasks.

16. Questions?