1. Topic: Semantic Text Mining
Bin Li 1 1

2. Outline Paper 1 : Table Cell Search for Question Answering Background
Table cell search framework
Evaluation 2 2

3. Background complexer question unstructured knowledge resource
limit of query for traditional knowledge bases 3 3

4. Background complexer question unstructured knowledge resource
limit of query for traditional knowledge bases
Precisely retrieve table cells from
web to answer a user question
Measure 4 4

5. Table cell search framework
formulate question chain and relational chain
find the best answer (using deep neural networks)
extract the corresponding answer

6. what languages do people in france speak?
natural language question

7. Question chain what languages do people in france speak? Topic entity
natural language question
Question chain
Topic entity
Question pattern

8. Question chain what languages do people in france speak? Topic entity
natural language question
Question chain
Topic entity
Question pattern

9. Question chain what languages do people in <e> speak?
natural language question
Question chain
Topic entity
Question pattern

10. Question chain
relevant table
what languages do
people in <e> speak?
France ?X

11. Question chain
Relevant table
what languages do
people in <e> speak?
France ?X
Row graph

12. people in <e> speak? France ?X
Question chain
Relevant table
what languages do
people in <e> speak?
France ?X
Row graph
column name
Topic cell
Ending cell
Relation chain

13. people in <e> speak? France ?X
Question chain
Relevant table
what languages do
people in <e> speak?
France ?X
matching
imply inward and outward relations
Row graph
column name
Topic cell
Ending cell
Relation chain

14. Candidate Chains Question chain Relevant table what languages do
people in <e> speak?
France ?X
Candidate Chains
matching
imply inward and out ward relations
Row graph
column name
Topic cell
Ending cell
Relation chain

15. Get a large set of candidate chains via string matching.
Evaluate the relevance of a candidate chain to the input question.
Get more accurate candidate chains.
Explore the information of candidate chains.
Use deep neural networks to evaluate the matching degree.

16. Chain inference Semantic representation
non-linear feed-forward neural network
fixed-lenth global feature vector
extract most salient local feature
llocal contextual feature vector
convolution concatenate
{#-s-p,s-p-e,p-e-a,e-a-k,a-k-#}
what languages do people in <e> speak?

17. Chain inference Semantic representation
non-linear feed-forward neural network
fixed-lenth global feature vector
extract most salient local feature
llocal contextual feature vector
convolution concatenate
{#-s-p,s-p-e,p-e-a,e-a-k,a-k-#}
what languages do people in <e> speak?
As input: question pattern, word sequence, answer type, peseudo-predicate,entity pairs

18. Features Shallow features Deep features
word-level
matching degree
Deep features
answer type
pseudo-predicate
entity pairs
Calculate the cosine similarity between the question and the candidate chain.

19. Evaluation

20. Evaluation Measures Precision Recall measure
(Harmonic mean of P and R)

21. Paper 2 : Dynamic Collective Entity Representations for Entity Ranking
Outline
Background
Dynamic collective entity representations(DCER)
Evaluation 21 21

22. Background Mismatch Manualy query
Entity's description in knowledge base
Manualy query
Context dependency
Time dependency

23. Dynamic collective entity representation
Collective intelligence
Manualy query
Knowledge base entity description
Fielded documents represent entities
continuously update ranking model
incorporate new descriptions
Retrain model
adjust weights associated to entity's fields
Dynamic collective entity representation

24. Dynamic collective entity representations(DCER)
Problem:
Query: q
Knowledge base: KB (consist of entities e E)
Aim: find the best match between e and q
Approach:
Expand entity representations(field document)
Reduce the vocabulary gap between queties and entities
Train classification-based ranker
Combine content from each field

25. Description sources
Knowledge base
External description sources
(static: web achives, web anchors...
dynamic: tweets, query logs...)
Adaptive entity ranking
Supervised entity ranker learns to weight the fields
but two challenges by constructing DCER:
-Heterogeneity(volume, quaity, quantity, type...)
within entities
between entites
-Dynamicness
cannot capture the evolving and continually changing
so “Adaptive entity ranking”(continuously undated)

26. Model Entity representaton:
: Field term vector, represent the content of e
Updating fields:
Estimate e's relevance to a query q:
Supervises single-field weighting model

27. Three features express the importance of field and entity
Field similarity
Query-Field similarity score
Field importance
status of the field at a point in time
Entity importance
resently updated entities

28. Ranker Macine learning Query
(Employ a supervised ranker to learn the optimal feature weights for retrieval)
Weight vector:
( : weights of each of the field features and the entity importance feature)
Top K-retrieval
Query
Candidate entities
user interaction(i.e.,clicks)
feature vectore(x)
Ranker
Optimal
classification's condidence score

29. Evaluation Q1: Does entity ranking effectiveness increase using DCER?
Compare and
Q2: Does entity ranking effectiveness increase when employing field and entity features?
Compare ,KBER(incorperate entity and field importance feature) and DCER
Q3: Does entity ranking effectiveness increase when we continuosly learn the optimal entity representation?
Compare DCER and (non-adaptive)

30. Thank you for your attention!