Bring back lost lover in USA, Canada ,Uk ,Australia ,London Lost Love Spell C...
Towards Improving Dialogue Topic Tracking Performances with Wikification of Concept Mentions
1. Towards Improving Dialogue Topic Tracking Performances
with Wikification of Concept Mentions
Seokhwan Kim, Rafael E. Banchs, Haizhou Li
Human Language Technology Department, Institute for Infocomm Research (I2
R), Singapore
Dialogue Topic Tracking
Detecting topic transitions and predicting topic categories
In on-going dialogues which address more than a single topic
Classification problem f(xt) = (yt−1, yt)
Examples of dialogue topic tracking
Speaker Utterance Topic Transition
Guide How can I help you? NONE→NONE
Tourist Can you recommend some good places to visit in Sin-
gapore?
NONE→ATTR
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? ATTR→ATTR
Guide Yes, we use that to symbolise Singapore.
Tourist Okay. ATTR→ATTR
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? ATTR→TRSP
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 destina-
tion?
TRSP→TRSP
Guide Yes, it’ll take only ten minutes on foot.
Tourist Alright. TRSP→FOOD
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?
FOOD→FOOD
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. FOOD→FOOD
Wikipedia-based Dialogue Topic Tracking
To leverage on Wikipedia as an external knowledge source
Obtained without significant effort toward building resources
Kernel methods based on dialogue segment-level correspondences to
Wikipedia articles obtained with vector space models
Kim et al. (ICASSP 2014, ACL 2014)
Limitations
Mismatches between spoken dialogues and written encyclopedia
Multiple relevances from a single dialogue segment to more than one concept
Lack of semantic or discourse aspects in concept retrieval
Wikification of Concept Mentions in Spoken Dialogues
Wikification aims at linking mentions to the relevant entries in Wikipedia
Examples of Wikification results on the previous example
Singapore, Merlion Park, Orchard Road, North South MRT Line, Raffles
Place MRT Station Singapore River, Chilli crab
Dealing with co-references, ambiguities, and variabilities of mentions
Every noun phrase is defined as a single mention
For each mention, candidates are retrieved from a Lucene index
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
The top-ranked item in the list of candidates scored by this model is
considered as the result of Wikification for a given mention
Wikification-based Features for Topic Tracking
Baseline based on vector space model
φ(x) = α1, α2, · · · , α|W| ∈ R|W|
αi =
h
j=0
λj
· tfidf(wi, u(t−j))
Wikipedia-based features (ICASSP 2014, ACL 2014)
φ (x) = α1, α2, · · · , α|W|, β1, β2, · · · , β|D|
βi = sim(x, ci) = cos (θ) =
φ(x) · φ(ci)
|φ(x)||φ(ci)|
Wikification-based features
γi =
h
j=0
λj
· mk ∈ u(t−j)|g(mk) = ci
Experimental Setup
Singapore tour guide dialogues
Human-human mixed initiative dialogues
35 sessions, 21 hours, 31,034 utterances
Manually annotated with nine topic categories
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
Wikification of Concept Mentions
Kim et al. (EMNLP 2015)
Preprocessed by Stanford CoreNLP toolkit
For every noun phrase, top 100 candidates are retrieved from Lucene index
SVMrank
was used to train a ranking function
Wikification Performances
Five-fold cross validation
38.04/31.97/34.74 in Precision/Recall/F-measure
Dialogue Topic Tracking
SVM models were tranined using SVMlight
3 schedules were considered: every turns, only tourist turns, and only guide turns
All the evaluations were done in five-fold cross validation
Evaluation metrics
Accuracy of the predicted topic label for every turn
Precision/Recall/F-measure for each event of topic transition occurred
Experimental Results
Schedule Features
Transition Turn
P R F ACC
all
α 42.08 53.48 47.10 67.97
α + β 42.12 53.38 47.08 67.98
α + γ 47.36 50.19 48.73 72.38
α + β + γ 47.35 50.24 48.75 72.43
α + γ 50.77 49.36 50.06 79.12
α + β + γ 50.82 49.41 50.10 79.15
tourist turns
α 41.88 52.59 46.63 67.15
α + β 41.84 52.75 46.67 67.08
α + γ 46.58 51.09 48.73 71.99
α + β + γ 46.57 51.09 48.72 71.99
α + γ 50.51 49.58 50.04 81.10
α + β + γ 50.43 49.58 50.00 81.10
guide turns
α 41.96 52.11 46.49 67.13
α + β 41.91 52.03 46.42 67.13
α + γ 47.10 48.44 47.76 71.94
α + β + γ 47.02 48.21 47.61 71.93
α + γ 50.94 49.10 50.00 78.92
α + β + γ 50.98 49.02 49.98 78.92
γ is the oracle features with the manual annotations for Wikification
1 Fusionopolis Way, #21-01 Connexis (South Tower), Singapore 138632 Email: kims@i2r.a-star.edu.sg