Automated Fact-Checking of Text Summaries for Relational Databases
Classified in Computers
Written at on 
English with a size of 4.06 KB.
Verifying Text Summaries of Relational Data Sets
• Relational data is often summarized by text.
• The focus of this paper is the problem of verifying, in an automated fashion, whether text claims are consistent with the actual database.
• The authors proposed a tool for verifying text summaries of relational data sets, which works similarly to a spell checker and marks up claims that are believed to be erroneous.
• The system converts claims into SQL queries and then evaluates them.
• The main problem is converting natural language claims to SQL queries.
• The tool is called AggChecker.
AggChecker
• AggChecker consists of two parts: a relational data set and a text document.
• The text contains claims about the data.
• The goal is to translate natural language claims into pairs of SQL queries and claimed query results
Keyword Matching
Each claim in the input text can be associated with relevant keywords.
- Extracting Keywords from Text
- Constructing Likely Query Candidates
Probabilistic Model and Query Evaluation
CONCLUSION
• Introduced the problem of fact-checking natural language summaries of relational databases
• Presented a first corresponding approach, encapsulated into a novel tool called AggChecker
• Successfully used it to identify erroneous claims in articles from major newspapers.
AStream: Ad-hoc Shared Stream Processing
Ad-hoc Stream Requirements
- Integration
- Consistency
- Performance
OPTIMIZATION
- Incremental query processing
- Data copy and shuffling
- Memory-efficient dynamic slice data structure
Exactly-Once Semantics
Every input tuple is only processed once, even under failures. Astream is deterministic because all its distributed components are deterministic. Event-time semantics ensure correctness on out-of-order events or during replays of data.
RELATED WORK
- Query-at-a-Time Processing
- Stream Multi-Query Optimization
- Adaptive Query Optimization
- Batch Ad-hoc Query Processing Systems
- Streaming Query Sharing
FUTURE WORK
In future work, we plan to extend AStream with a cost-based optimizer and adaptive query processing techniques. Based on sharing statistics among queries collected at runtime, a more optimal query plan can be generated by grouping similar queries.
An End-to-End Automatic Cloud Database Tuning System Using Deep Reinforcement Learning
PROPOSED APPROACH
• Uses deep RL to learn and recommend configurations for databases
• Uses only a limited number of samples
• Designed to work end-to-end
• It has good adaptability to environment changes
• Significantly outperformed the state-of-the-art tuning tools and DBA experts.
Reinforcement Learning
Reinforcement Learning is a general-purpose framework for decision-making. It basically learns by trial and error.
TRAINING DATA GENERATION
- Cold start
- Incremental Training
METHODOLOGY
• The process starts with offline training
• The training data is a set of training quadruples <q, a, s, r>
-- q: a set of query workloads (i.e., SQL queries)
-- a: a set of knobs as well as their values when processing q
-- s: the database state (which is a set of 63 metrics) when processing q
-- r: the performance when processing q (including throughput and latency).
CONCLUSION
• An End-to-End Automatic Cloud Database Tuning System Using Deep Reinforcement Learning is proposed
• Its superiority is demonstrated with extensive experimental results
• It is much faster than DBAs and other methods
• It has good adaptability against workload and environment changes