How to automatically Gime a Lever & a Fulcrum for ATA analyze the dynamics SFR Agorantic : LIA UAPV of images on the Web 2.0 Marc El-Bèze Project ImagiWeb funded by the ANR 2014.IX.10 PhD : Jean-Valère COSSU (Last year) Co-direction: E Sanjuan, JM Torres-Moreno 1 2 Thursday 10th September 2014 ( TLSE ) – Lever & Fulcrum for ATA Thursday 10th September 2014 ( TLSE ) – Lever & Fulcrum for ATA ‘Images’ on the Web 2.0 Aims, Goals and Objectives Multiple sides of an ‘Image’: • How a component of an entity is perceived? • What the entity (politician or company) emits • How wide is the gap between • How (s)he/it is perceived by members of OSN – the perception and what was expected? • What can be done to reduce it? In case of a competition (elections for instance) Our goal is not to predict who will be elected … – each entity is not to be seen as if it was alone – his image is exposed to interaction But to Produce dashboards (personal summaries) – and … other actions or events giving an overview of the Images Once spread, is the control of the image lost? How to do that? 3 3 4 4 Thursday 10th September 2014 ( TLSE ) – Lever & Fulcrum for ATA Thursday 10th September 2014 ( TLSE ) – Lever & Fulcrum for ATA
Segmentation Workflow From all the messages: 3 Populations to be segmented: • Probabilistic extraction of discriminant Terms • People sharing the same opinion on one entity ( N grams with N variable : mwe or chunks) • Topics & subTopics = components of an Image Well-Argued Recommendation: Adaptive Models Based on Words in Recommender Systems Gaillard et al. • Messages will be segmented in order to EMNLP’13) – take into account nuanced discourses • For each message, automatically identify – improve the granularity of each partition who is emitting this opinion ? which entity is Sender’s identity and entity taken into account targeted ? which (sub)targets are concerned ? 3W h + 2P (parity & polarity) 5 6 5 6 Thursday 10th September 2014 ( TLSE ) – Lever & Fulcrum for ATA Thursday 10th September 2014 ( TLSE ) – Lever & Fulcrum for ATA Assumptions we have discarded Assumptions for the Polar Task for the 2 Tasks : Pol & Topic In the politics domain: Entities X, Y: opponents supported by B & C Assumption A 1 • During a given short period, the global opinion Assumption D 1 : B has positive opinion on X, of one person on one entity remains quite stable. whatever the topic Assumption A 2 • D 2 : if B has a positive opinion on X about a • Variations of the opinion can be topic s, B opinion on Y about s must be negative – predicted thanks to time series modelization; • D 3 : if B opinion on X about s is positive, C – estimated by comparing 2 successive histograms. opinion on X must be negative 7 7 8 8 Thursday 10th September 2014 ( TLSE ) – Lever & Fulcrum for ATA Thursday 10th September 2014 ( TLSE ) – Lever & Fulcrum for ATA
Assumptions for the Polar Task Assumptions for the Topic Task In the politics domain: • How much these assumptions are true? • Let us call T a set of predefined topic • How can we use it to enrich the annotated • As for the polar task corpus thru an incremental process? – Classical classification approaches can be used In practice: • In this case, assumption A 1 does not apply: � • For a classification method such as the Vector we have to look for another lever model using BOW and Cosine, the user id or its • If the text comes from Twitter, hashtags are membership can be introduced into each vector as good candidates for a fulcrum an additional component 9 10 9 10 Thursday 10th September 2014 ( TLSE ) – Lever & Fulcrum for ATA Thursday 10th September 2014 ( TLSE ) – Lever & Fulcrum for ATA Adaptation of Recommender Systems Proposal P for the Topic Task Whatever the domain: Each new rating (and the comment associated) can be directly used to adapt the model • Collect (aside the training corpus) a huge • Flash reactivity: adaptive models in recommender amount of non annotated tweets or texts systems J Gaillard et al.DMIN‘13 • Keep the S closer tweets (according to time) • for each tag t in T extract the most n t similar This can be used to prevent the system to make tweets ( n t corresponding to what is expected for t ) twice the same error • Classify all the remaining S - Sum n t This is a good way to reduce the difficulties due – which may be related (or not) to new topics to the so-called « cold start » problem 11 11 12 12 Thursday 10th September 2014 ( TLSE ) – Lever & Fulcrum for ATA Thursday 10th September 2014 ( TLSE ) – Lever & Fulcrum for ATA
What cannot be done in our case Some pecularities of our problem 1. Predict Rating R t In the case of topic and opinion detection / blogs TRAINING ADAPT ADAPT TEST No star … no rating … no tag … no hashtag 2. Adapt with Rating R t-k Texts are longer than a tweet When adapting with the prediction instead of the In a same document, authors may give positive, reference, performances not so much improved neutral and negative opinions on several (sub-) topics It becomes a problem of multilabel categorization In case of topic and opinion detection / Twitter More difficult than a monolabel categorization there is no star no rating no thematic tag 13 14 13 14 Thursday 10th September 2014 ( TLSE ) – Lever & Fulcrum for ATA Thursday 10th September 2014 ( TLSE ) – Lever & Fulcrum for ATA A simple way to reduce the In Summary complexity PSEUDOS NON ANNOTATED TRAINING clustering CORPUS CORPUS Divide texts into segments (phrase, sentence,§) FOR A GIVEN Apply Proposal P to this set of segments FLIP FLOP PERIOD t On each unannotated segment apply as principle categorization confidence measure = Intersystem agreement Advantages: no need to align, less complex METADATA SET OF TOPICS EVENTS Drawbacks: SETS OF MWE / TOPIC a picture DATES Context is lost, subject to oscillation …. at time t SETS OF MWE / POLARITY 15 15 16 16 Thursday 10th September 2014 ( TLSE ) – Lever & Fulcrum for ATA Thursday 10th September 2014 ( TLSE ) – Lever & Fulcrum for ATA
Questions of interest Feasibility of AutoGenerating DashBoards � showing for an entity good/bad points • Images evolution – Follow trends for some particular group • List points to focus on – How to Act on these points? – What to do mostly to attract ‘ swing people’ ? 17 17 Thursday 10th September 2014 ( TLSE ) – Lever & Fulcrum for ATA
Recommend
More recommend
