Wikification of Concept Mentions within Spoken Dialogues Using Domain Constraints from Wikipedia
1. Wikification of Concept Mentions within Spoken Dialogues
Using Domain Constraints from Wikipedia
Seokhwan Kim, Rafael E. Banchs, Haizhou Li
Human Language Technology Department, Institute for Infocomm Research (I2
R), Singapore
Wikification on Spoken Dialogues
Linking mentions to the relevant concepts in Wikipedia
Differences between spoken dialogues and written texts
Number of speakers
Dependencies to background knowledge
Degree of informal and noisy expressions
Examples of Wikification on Singapore tour guide dialogues
Guide How can I help you?
Tourist Can you recommend some good places to visit in Singapore?
Guide Well if you like to visit an icon of Singapore, Merlion park will be a nice
place to visit.
Tourist That is a symbol for your country, right?
Guide Yes, we use that to symbolise Singapore.
Tourist Okay.
Guide The lion head symbolised the founding of the island and the fish body
just symbolised the humble fishing village.
Tourist How can I get there from Orchard Road?
Guide You can take the red line train from Orchard and stop at Raffles Place.
Tourist Is this walking distance from the station to the destination?
Guide Yes, it’ll take only ten minutes on foot.
Tourist Alright.
Guide Well, you can also enjoy some seafoods at the riverside near the
place.
Tourist What food do you have any recommendations to try there?
Guide If you like spicy foods, you must try chilli crab which is one of our
favourite dishes here.
Tourist Great! I’ll try that.
Singapore, Merlion Park, Orchard Road, North South MRT Line, Raffles
Place MRT Station Singapore River, Chilli crab
Three-step Approach for Wikification on Dialogues
Input Mention
mi
Linking
Validity
Analysis
In-dialogue
Reference
Analysis
Domain
Relevance
Analysis
Speaker
Relatedness
Analysis
Candidate
Generation
Wikipedia
Concepts
History
<mj, f(mj)>j=0..(i-1)
Candidate
Ranking
Output Concept
f(mi)
Step 1
Step 2
Step 3
Step 1: Mention Analysis
Analyzing four binary properties of a given mention
Linking validity, In-dialogue reference, Domain relevance, Speaker relatedness
Guide: In the morning I suggest to you to go to Botanical Garden.
LV ID DR SRG SRT
- - - - -
LV ID DR SRG SRT
+ - + + -
Tourist: Oh, we also have Botanical Garden.
LV ID DR SRG SRT
+ - - - +
Tourist: That is actually one of my favourite places here.
LV ID DR SRG SRT
+ + - - +
LV ID DR SRG SRT
+ - - - +
Guide: If so, you might like this place also.
LV ID DR SRG SRT
+ + + + -
Step 2: Candidate Generation
Candidates retrieval from a Lucene index on the Wikipedia collection
With filtering constraints based on the analyzed properties in step 1
Combination of multiple constraints: Intersection or Union
Step 3: Candidate Ranking
Ranking SVM: Supervised learning to rank algorithm
s(m, c) =
4 if c is the exactly same as g(m),
3 if c is the parent article of g(m),
2 if c belongs to the same article
but different section of g(m),
1 otherwise.
m: a mention
c: a candidate concept
g(m): the manual annotation for the most relevant concept of m
Datasets
Singapore tour guide dialogues
Human-human mixed initiative dialogues
35 sessions, 21 hours, 31,034 utterances
Manually annotated with relevant Wikipedia concepts
Preprocessed by Stanford CoreNLP toolkit
Wikipedia collection
4,797,927 articles and 25,577,464 sections in total
Collected from Wikipedia database dump as of January 2015
Indexed into a Lucene index
Evaluation: Mention Analysis
SVMlight
was used for training four mention analyzers
With four sets of features: mention (M), utterance (U), dialogue (D),
and Wikipedia-based (W) features
Five-fold cross validation with F-measure
Features LV ID SRG SRT
M 86.29 69.15 71.10 72.94
M+U 86.90 70.43 70.43 68.85
M+D 86.17 71.09 70.56 71.52
M+W 86.21 68.96 70.66 71.86
M+U+D 86.82 72.37 70.12 68.30
M+U+W 86.84 70.13 70.19 68.78
M+U+D+W 86.77 72.20 69.94 68.10
Evaluation: Candidate Generation
Four sets of candidates were prepared for each mention
Baseline: Retrieved with no filtering
Intersection: Filtered with intersection of analyzed properties
Union: Filtered with union of analyzed properties
Oracle: Filtered with manually annotated properties
Top 100 candidates were retrieved from a Lucene index for each set
Evaluation: Candidate Ranking
SVMrank
was used for training ranking functions
The top-ranked item in the list is considered as the result of Wikification
Five-fold cross validation with Precision/Recall/F-measure
Method P R F
Baseline 26.85 22.52 21.24
Intersection 44.37 27.35 33.84
Union 38.04 31.97 34.74
Manual Filtering 39.90 34.72 37.13
1 Fusionopolis Way, #21-01 Connexis (South Tower), Singapore 138632 Email: kims@i2r.a-star.edu.sg