What people suggests
In a corpus of about 1.300.000 conversations about crypto money, the task was to detect which kind of suggestions in terms of buy/sell the users were exchanging and which were the crypto stocks which were affected by these suggestions. This amounted to configuring ASI to detect specific discourse acts ((dis)suggestion of buying) and identify their specific object. In particular the system was able to detect primitives such as SUGGESTION_TO_BUY (21.500), SUGGESTION_TO_SELL (2.600), SUGGESTION_NOT_TO_BUY (2.650), SUGGESTION_NOT_TO_SELL (900) and associate these primitives with specific crypto symbols.
In order to perform this task, ASI was able to automatically identify:
- Crypto money denominations;
- Suggestions expressions;
- Economic transaction verbs,
- Relations among these actors.
It was then up to the linguist team to put them together in a syntactically and semantically coherent system of rules.
The configuration of the "buy or not" project took about three days. After a phase of testing, the application was integrated with customer processing systems for day by day analysis. In the specific case, the consuming application was a predictive machine learning based system which was using linguistic analysis as one of the possible input features.
It should be noticed that these performances would have been impossible without an extensive amount of research on rule induction and relation extraction as well as a background organization of resources as “cognitive cartridges” (e.g. to detect dialogue acts).