So this long keyword query will be split into 3 separate queries. Each called an aspect query.These aspect queries are scored separately and the results are then combined.
-for each entity, average the numerical ratings of each aspect-assumption: this would be a good approximation to human judgment
Otherwise, this tells you that the system is not really doing well in ranking.
-could not obtain natural queries, so we used semi synthetic queries.-what we did was-and then we randomly combined queries…to form a set of queries.
Then finally we conducted a user study where users were asked to manually determine the relevance of the the sysGen results to query. This is to validate that the results made sense to real usersAnd also to validate the effectiveness of the gold standard rankings which is based on the…Based on this we found that…Which means that this evaluation method can be safely used for similar ranking tasks…
Ganesan & Zhai 2012, Information Retrieval, Vol 15, Number 2Kavita Ganesan (www.kavita-ganesan.com)University of Illinois @ Urbana ChampaignJournalProject Page
Currently: No easy or direct way of finding entities (e.g. products, people, businesses) based on online opinions You need to read opinions about different entities to find entities that fulfill personal criteria e.g. finding mp3 players with ‘good sound quality’
Currently: No easy or direct way of finding entities (e.g. products, people, businesses) based on online opinions You need to read opinions about different entities to find entities that fulfill personal criteria (e.g. finding mp3 players with ‘good sound quality’ Time consuming process & impairs user productivity!
Use existing opinions to rank entities based on a set of unstructured user preferences Example of user preferences: Finding a hotel: “clean rooms, heated pools” Finding a restaurant: “authentic food, good ambience”
Most obvious way: use results of existing opinion mining methods Find sentiment ratings on various aspects ▪ For example, for an mp3 player: find ratings for screen, sound, battery life aspects ▪ Then, rank entities based on these discovered aspect ratings Problem is that this is Not practical! ▪ Costly – It is costly to mine large amounts of textual content ▪ Prior knowledge – You need to know the set of queriable aspects in advance. So, you may have to define aspects for each domain either manually or through text mining ▪ Supervision – Most of the existing methods rely on some form of supervision like the presence of overall user ratings. Such information may not always be available.
Leverage Existing Text Retrieval Models Why? Retrieval models can scale up to large amounts of textual content The models themselves can be tweaked or redefined This does not require costly information extraction or text mining
Leveraging robust text retrieval models Indexed rank Entity 1 Entity 1 Reviews rank retrieval User Preferences Entity 2 models (query) Entity 2 Reviews (BM25, LM, PL2) rank Entity 3 Entity 3 Reviews Keyword match between user prefs & textual reviews
Leveraging robust text retrieval models Indexed rank Entity 3 Entity 3 Reviews rank retrieval User Preferences Entity 2 models (query) Entity 2 Reviews (BM25, LM, PL2) rank Entity 1 Entity 1 Reviews Keyword match between user prefs & textual reviews
Based on the basic setup, this ranking problem seems similar to regular document retrieval problem However, there are important differences:1. The query is meant to express a users preferences in keywords Query is expected to be longer than regular keyword queries Query may contain sub-queries expressing preferences for different aspects It may actually be beneficial to model these semantic aspects2. Ranking is to capture how well an entity satisfies a users preferences Not the relevance of a document to a query (as in regular retrieval) The matching of opinion/sentiment words would be important in this case
Investigate use of text retrieval models for the task of Opinion-Based Entity Ranking Explore some extensions over IR models Propose evaluation method for the ranking task User Study To determine if results make sense to users Validate effectiveness of evaluation method
In standard text retrieval we cannot distinguish the multiple preferences in a query. For example: “clean rooms, cheap, good service” Would be treated as a long keyword query even though there are 3 preferences in the query Problem with this is that an entity may score highly because of matching one aspect extremely well To improve this: We try to score each preference separately and then combine the results
Aspect Queries“clean rooms, cheap, “good “clean rooms” “cheap” service” good service” scored retrieval model separately retrieval model result set 1 result set 2 result set 3 Results results Results combined
In standard retrieval models the matching of an opinion word & a standard topic word is not distinguished However, with Opinion-Based Entity Ranking: It is important to match opinion words in the query, but opinion words tend to have more variation than topic words Solution: Expand a query with similar opinion words to help emphasize the matching of opinions
Similar Meaning toFantastic battery life “Fantastic battery life” Query Good battery life Great battery life Excellent battery life Review documents
Similar Meaning toFantastic battery life “Fantastic battery life” Query Add synonyms of Good battery life word “fantastic” Fantastic, good, Great battery life great,excellent… battery life Excellent battery life Expanded Query Review documents
Document Collection: Reviews of Hotels – Tripadvisor Reviews of Cars – Edmunds Numerical aspect ratings Gold standard Free text reviews
Gold Standard: Needed to asses performance of ranking task For each entity & for each aspect (in dataset): Average numerical ratings across reviews. This will give the judgment score for each aspect Assumption: Since the numerical ratings were given by users, this would be a good approximation to actual human judgment
Gold Standard: Ex. User looking for cars with “good performance” Ideally, the system should return cars with ▪ High numerical ratings on performance aspect ▪ Otherwise, we can say that the system is not doing well in ranking Should have high ratings on performance
User Queries Semi synthethic queries Not able to obtain natural sample of queries Ask users to specify preferences on different aspects of car & hotel based on aspects available in dataset ▪ Seed queries ▪ Ex. Fuel: “good gas mileage”, “great mpg” Randomly combine seed queries from different aspects forms synthetic queries ▪ Ex. Query 1: “great mpg, reliable car” ▪ Ex. Query 2: “comfortable, good performance”
Evaluation Measure: nDCG This measure is ideal because it is based on multiple levels of ranking The numerical ratings used as judgment scores has a range of values and nDCG will actually support this.
Users were asked to manually determine the relevance of system generated rankings to a set of queriesTwo reasons for user study: Validate that results made sense to real users On average, users thought that the entities retrieved by the system were a reasonable match to the queries Validate effectiveness of gold standard rankings Gold standard ranking has relatively strong agreement with user rankings. This means the gold standard based on numerical ratings is a good approximation to human judgment
Most effective Most effective on BM25 (p23) on BM25 (p23)8.0% Hotels 2.5% Cars6.0% 2.0% 1.5%4.0% 1.0%2.0% 0.5%0.0% 0.0% PL2 LM BM25 PL2 LM BM25 QAM QAM + OpinExp QAM QAM + OpinExpImprovement in ranking using QAMImprovement in ranking using QAM + OpinExp
Lightweight approach to ranking entities based on opinions Use existing text retrieval models Explored some enhancements over retrieval models Namely opinion expansion & query aspect modeling Both showed some improvement in ranking Proposed evaluation method using user ratings User study shows that the evaluation method is sound This method can be used for future evaluation tasks