Presentation describes the method for link discovery that aims to connect debate content (textual documents from the parliament) on a speech level with relevant articles that contain not just the mentions of speakers but also mentions of speakers in a context of topics or events that politicians tackled in their speech in parliament. Method uses semantic and information retrieval techniques to generate automatic queries that contain the context of the parliamentary speeches and to search newspaper, radio and video data sets for the connections between speeches and newspaper articles that are covering them.
2. Background: the PoliMedia project
• The PoliMedia project:
– driven by research questions from historians
– interested in media coverage across several types of
media outlets
– Cross-media comparisons
• conducted over a longer period of time, on different topics
• focus on the coverage of the debates in the Dutch parliament
• insight on the different choices that different media make while
reporting on those debates
– three phases :
• modeling phase: creating a semantic model
• data production phase: creating links between debates and
associated media sources
• application phase: searching and navigating linked datasets
3. Introduction
• Polimediasemantic model needs to represent:
– people
– topics
– time
– media types
• Model has to be expressive enough:
– describing events from the Dutch parliament
4. Data Sets
• Primary data set:
– The Dutch parliamentary debates
(Handelingender Staten-General
or Dutch Hansard)
– transcripts of speeches that
politicians had in the parliament
– this project uses data from the
Political Mashup
– all debates until the year 1995:
• published as XML documents (OCR
with satisfactory quality is being
used).
• data shows a fine-grained structure.
5. Data Sets
• Secondary data set:
– different media types:
• newspaper articles and radio
bulletins
– National Library of the
Netherlands
• newscasts
– evening news and current affairs
programs
6. Semantic model: Goals
• Goal of the project:
– to publish the links on the Web
– to use open Web formats and standards
– Web query language
– unique identifiers (URI’s)
• Model has to be expressive:
– important information regarding parliamentary
debates should be easily accessed
7. Debate: The structure
Metadata
Debate
Metadata NEs={EconomischeZaken, Borssele}
Aan de orde is de behandeling van: - de brief van de minister van
Topic 1 EconomischeZakeninzakeBorssele(16226, nr. 26).
De beraadslagingwordtgeopend.
Speaker 1 Speaker 1 / Content
NEs={Borssele, Partij van de Arbeid, D66}
Speaker 2 Speaker 2 / Content
Mijnheer de Voorzitter! Met de verdragen tot uitbreiding van de EEG met
Denemarken,Engeland, Ierland en Noorwegenwordteen van de doelstellingen
Speaker 3 Speaker 3 / Content van onsbuitenlandsbeleidverwezenlijkt.
Topic 2
Speaker 1 / Content
10. Polimedia linking method
• The challenge: how to create a representation of the speech that contains enough
information, so it can be used as a query to retrieve relevant media articles from
the archive?
• Debate speeches and newspaper articles are generally different types of
documents (so computing document similarity doesn’t work) in the style and
scope
• Speeches can contain large number of NEs and digressions:
– Problem: hard to distinguish the right context for each speech
• Newspaper articles:
– very strict and concise
– words are used sparingly
11. Polimedia linking method
• Our PoliMedia linking method consists of four steps:
1. topics: enriching the existing debate metadata with topics
2. preselection of articles: when the candidate articles were published and who
spoke in the debate (timeframe and speakers)?
3. automatic query creation: candidate articles are ranked based on similarity to
the query (automatically created from speech text) by comparing vectors of
topics and named entities
4. link creation: links are created between a speech and an article if the
similarity score is above a threshold t
12. Topics
• Topic modeling:
– popular tool for the unsupervised analysis of text,
– used to check models, summarize the corpus, and guide exploration of its contents
– topic models lead to semantically meaningful decompositions of text because they tend to place high probability on
words that represent concepts
• Extracting topics from speech:
– ten words that represent one topic discussed inside the speech are extracted
– all speeches contained inside one debate segment are concatenated into one text
– set of ten words that represent one topic of the debate segment as a whole is extracted from that text
• Input: text /number of iterations/number of topics
• Output: generic names for topics/words that cluster around one topic
• Example:
– Test case: debate nr. 1975/number of iterations: 2000/numbner of topics: 1
13. Automatic query creation
Metadata
NE Speech
Staatssecretaris
Regering
Euro-kapitaalmarkt Debate
Tariefnota
Metadata
TopicSet Topic Financiën
moeten Zwitserland Topic 1
fraude Grave
TopicSet Speech wetgeving Brussel Speaker 1 / Content
inkomstenbelasting sociale EEG
bronheffing Speaker 2 / Content
misbruik Netherlandse
Kombrink ten Contourennota
Speaker 3 / Content
rente fraudebestrijding Kombrink
contourennota vraag Nederland
Nederland gebruik Contou Topic 2
vereenvoudiging kamer OESO-verband
Speaker 1 / Content
tarief Midden-Oosten
word misbruikfraudebest Engwirda
rijdingismo-rapport
tussen Couprie
NE Topic Actor
14. Automatic query creation
Scholten
+(text:wetsontwerptext:latertext:septembertext:prijzentext:lonentext:ontwikkelingtext:zeggentext:staatssecretaristext:gebrachttext:ertoe)
(title:wetsontwerptitle:latertitle:septembertitle:prijzentitle:lonentitle:ontwikkelingtitle:zeggentitle:staatssecretaristitle:gebrachttitle:ertoe)
+(text:staatssecretaristext:huurverhogingtext:jaartext:moetentext:apriltext:uitsteltext:percentagetext:nieuwe)
(title:staatssecretaristitle:huurverhogingtitle:jaartitle:moetentitle:apriltitle:uitsteltitle:percentagetitle:nieuwe)
+(text:regelentext:wet)
(title:regelentitle:wet)
+text:staatssecretaris
title:staatssecretaris
Mijnheerde Voorzitter ! In de memorie van toelichtingbij het voorliggendewetsontwerpzegt de Staatssecretaris , dathij over het
trendmatigehuurstijgingspercentagevoor 1977 nognietskanzeggenomdat de gegevens over de teverwachtenontwikkeling van lonen en prijzenvoor 1977
nognietbekendzijn . Dit is gedateerd 14 september . Impliceertdit , wanneerergeensprakezouzijn van eenwetsontwerp tot verschuiving van de ingangsdatum
, danook ten aanzien van de 8 procent per 1 aprilzougeldendatnogafgewachtmoetworden , of het dat percentage zalworden , omdat men pas later
ietsmeerweet over de ontwikkeling van lonen en prijzen ? De Staatssecretarisvoeltzich door ditwetsontwerpeigenlijkgedwongen op
eenvrijvroegtijdstiptochdaaroverietstezeggen . Immers , een week later namelijkbij brief van 21 septemberkomthijwel met eenbepaaldconcreetvoorstel .
Daarinstelthij : Het overleg met de vastecommissieheeftmijertoegebracht ...
ExpandedQuery =
NERsSpeech TopicSet Speech NER Topic TopicSet Topic
+
Speaker X =
ActorFromSpeech TimeFrame
15. Example of the relevant article
vvd: van dam baseertbeleidteveel op rossige prognoses van planbureau
kamermeerderheidtegenuitstel van huurverhoging
den haag — eenmeerderheid van de tweedekamervoelternietsvoor de huurverhoging van volgendjaaruitte, stellen van 1 april tot 1 juli. de fracties
van kvp, arp, chu, vvd, ds'7o en de kleinechristelijkepairtijenwillen de huurverhoging op 1 aprillatendoorgaan. staatssecretaris van dam van
volkshuisvestingwiluitstelom op 1 julivolgendjaareennieuwhuurbeleidtekunneninvoeren. daarvoorzalhij op kortetetmijndriewetsontwerpenindienen:
de huurprijzenwet, de wet op de huurcommissie en eenwijziging van het burgerlijkwetboek.debewindsmanzeidat met het afwijzen van uitstel in
feiteinvoering van het nieuwehuurbeleid op 1 julivolgendjaaronmogelijkwordtgemaakt. het nieuwestelselzaldan pas in 1978
ingevoerdkunnenworden. „met eenuitstel van driemaandenkomen we preciesuit", aldus de heer van dam. de arp'erscholten, die medenamenskvp en
chusprak, zegde de regeringallemedewerking toe om de nieuwehuurwetnog in dezekabinetsperiodetebehandelen, maarhijtwijfeldeeraan of op 1 juli
1977 het nieuwehuurbeleid al ingevoerdkanworden. de confessionelen en de vvdhouden vast aaneenhuurverhoging van 8 procent op 1 april.
staatssecretaris van dam wil pas op 1 julizonverhoging. zou de verhogingtoch op 1 aprilmoeteningaan, danwilhijeenverhoging van 7 procent. de
bewindsmankomtvolgens de confessionelentevroeg met eenverlaging van de jaarlijksehuurverhoging.het d'66-kamerlid nypelsdiendeeenmotie in
waarinhij de regeringverzoektbijverwerping van het uitsteltekomen met eenwetsontwerpvoor 7 procent op 1 april. ook de
pvda'etkombrinksuggereerdedezeoplossing. de heerkombrink deed eendringendberoep op de confessionelenom het uitstelteaanvaarden. de vvder de
beer vonddatelrnietvoldoenderedenenzijnvooruitstel van de huurverhoging. de staatssecretarisbaseertzijnbeleidteveel op „de rossige prognoses van
het centraal plan bureau", vindt de vvd. ook de christendemocratenvindendat van dam teveel van prognoses uitgaat die vaaktelaagzijn.depvda is het
met de regeringeensdat de huren op 1 juli met 8 procentomhoogmoeten. wijst de kamerdataf, danmoeten de huren op 1 april met 7
procentwordenverhoogd. men moetnietalleenkijkennaar de ontwikkeling van lonen en prijzen, men moetookkijkennaar het vrijbesteedbareinkomen.
de stijgingdaarvanzal in de komendejarenuiterstgeringzijn", zeikombrink. cpn-woordvoerderdraagstrazeidat de hurenbevrorenmoetenworden op het
huidigepeil.
17. Evaluation
•We tried three different approaches:
• Experiment 1: NEs in speech
• Experiment 2: NEs + topics in speech
• Experiment 3: NEs + topics in speech and debate
• Conclusion:
• best approach:
• named entities (speech + debate descriptions) and topics (speech + debate)
18. Results discussion
• structural elements of transcript:
• used to create complex and rich query from the speech
• treating particular speech as a part of the bigger context (conversation) and creating a query that is a
mixture of those elements:
• higher number or related articles retrieved
• What we learned?
• definition of link can be vague
• simple document similarity methods doesn't work
• journalist use their own “compression” methods when writing about debates
• long speeches with dozens of NEs and topics are sometimes represented with few concise sentences
19. End
• Thank you for listening
• more information on polimedia.nl
20. similarity measures
• similarity measures: metric that measures similarity or dissimilarity (distance) between two text strings for
approximate string matching or comparison and in fuzzy string searching
• Given two segments, the expanded query Q and the document from media archive D, the term frequency (TF) is
associated to a term t from the query Q and the document D, the similarity between Q and D is computed
according to the cosine similarity formula, where the generated value varies between 0 and 1:
• CosineSimilarity(Q,D) =
• BM25 is a bag-of-words retrieval function that ranks a set of documents based on the query terms appearing in
each document, regardless of the inter-relationship between the query terms within a document (e.g., their
relative proximity).
Given a query Q, containing keywords t1, ..., tn, the BM25 score of a document D is:
• BM25Score(Q,D) =
• where function represents term frequency of the term qtfrom the document D, is the length of the document D in
words, and avgdl is the average document length in the text collection from which documents are drawn.
Parameters k1 and bare free parameters. Function is the inverse document frequency weight of the query term qt.
21. similarity measures
• The overlap coefficient is a similarity measure related to the Jaccard index that
computes the overlap between two sets which is defined as follows:
overlap(Q,D) =
• If set X is a subset of Y or the converse then the overlap coefficient is equal to one.