During my last year of Master degree at Paul Sabatier university in Toulouse (FR), some colleagues and I, worked on building a personalized information retrieval system for retriving specific tweets.
Personalized information retrieval systems are search engines that are able to build and store a user profile based on user interaction and behaviour in order to rerank the query results to better match the user wishes.
Our project was composed of several blocks :
- Step 1 : Build the Lucene index
- Step 2 : Build the user profile mechanism
- Step 3 : Personalized inforamtion retrieval system with PyLucene, relevent tweets based on a query and the user profile
The working corpus
The corpus we worked on for our experimental project is a 6M+ tweets from 2009 to 2016 about
#iot. The corpus was already labeled with information about tweet’s author using automatic methods and NLP processings :
- sentiment : Opinion / point of view expressed in the tweet towards IoT (
- topicID : topic modeling analysis to group tweets in 6 topics, from
- country :
- gender : gender of the user (
Sample of our corpus
How to build a simple user profile ?
Our approach was pretty simple :
1. We turned all tweets to Word2Vec vectors
Word2Vec is a powerful model that is used to produce word embeddings. It’s trained on huge corpuses (we took the GNews pre-trained model) and produce a vector representation of each individual word. Usually, vectors produced have 50 to 300 dimensions. Word vectors are positioned in a vector space such that words that share common contexts in the corpus are located in close proximity to one another in the space.
We choose to use a very simple vector representation for a tweet : we
avg the words vectors which compose it.
2. We built models to predict user characteristics from a tweet vector
The idea was to get user characteristics (gender, topic, sentiment, country) from a w2v vector. So we trained 4 different models to predict each characteristic of the tweet vector : SVM for gender and topicId, MLPClassifier for sentiment and NaiveBayes for country prediction.
3. User profile representation and update
The user profile we designed was composed of a w2v vector and the 4 characteristics. Each time the user uses the search engine and interact with query results (by clicking or ❤-ing) we update the profile :
- The w2v vector is the
avgof all liked and clicked vector by the user in the current session
- The 4 characteristics are predicted with the models using the avg w2v vector.
Our user profile representation
Rerank the query results
Now that our profile is built, the idea is to use it !
For that, we first used a
OneHotEncoder. The first step is to train the OneHotEncoder model with all different possible values for gender, sentiment, country in order to convert these categorical variables into dummy variables.
The OneHotEncoder process
The produced OneHotEncoder vector from the user’s characteristics is appended to the user’s w2v vector : this new big vector is actually the real profile of the user. The final step or our personalized information retrieval system is to rerank results to match the user’s profile.
- We take the
ntop results retrieved by Lucene and build w2v+OneHotEncoder vectors
- Perform Cosine similairty between the user’s profile vector and all top
Rerank of the results
That’s all ! We have created a very simple personnalized search engine ✌