Presentation

[mc-cat-tty]

• [ ] EDA - Exploratory Dataset Analysis. Eg. unbalanced reviews sentiment
• [ ] Objectives and functional requirements. Assumptions and user's degrees of freedom (TUI).
• [ ] System architecture
• [ ] Basic retrieval models
• [ ] Weighting models
• [ ] Query expansion: global with BERT vs WordNet, local with Whoosh (?)
• [ ] Sentiment analysis
• [ ] Show results
• [ ] Future improvements: Clustering tree 4 w2v (embeddings from BIG LLM, not BERT)

[mc-cat-tty]

Developing presentation structure with more details: [@giuliocorradini]

• [x] AirBnB flaws / our needs: AirBnB doesn't provide textual search or filtering by reviews.
• [x] Dataset: Inside AirBnb; number of listings, number of reviews, ...
• [ ] Benchmarking: tailor-made benchmarking dataset with lower number of listings and more control, why precision and recall and not DCG - reusable benchmark dataset over different search engines
• [ ] EDA forward pointer: why w2v is less suitable than S.A.
• [ ] Assumptions + TUI/user interaction: short queries and not very expressive; the user has to use an additional field as a proxy for expressing emotions
• [ ] Schema: Whoosh inverted index and its schema (stored, scored, keyword/textual, ...); sentiment forward index
• [ ] Query lang: natural language + keyword-based filtering (room type) + sentiment keywords

[@mc-cat-tty]

• System architecture: Whoosh, customization, forward index, packages, ...
• Models overview: monotonic improvement curve
• Basic model: Whoosh sucks; UnionIR work around
• Increasing recall with global query expansion: BERT vs WordNet based; benchmarks feedback proves it
• Sentiment ranking: host redemption model (exp decay), but few reviews
• Sentiment ranking with number of reviews weighting

[@ent0n29]

• Exploratory Dataset Analysis: common context (no w2v) and flat sentiment in listings; unbalanced towards host's profit. Colorful sentiment in reviews (enabling us to do effective S.A.), unbalanced towards positive reviews (hence the decision of using fine-grained classes)
• Sentiment analysis internals: previous researches on binary classification, model (BERT multi-class classifier), training dataset forward pointer, examples
• GoEmotions - sentiment dataset: 27 classes, fine-grained emotions, etc. (see GoEmotions paper)
• Benchmarks results and considerations
• Future improvements: w2v clustering tree, ...
• Demo