If you love #government #data but hate faceted search and csv tables? How to make your database talk? How to squeeze a table into a sequence of characters?
If you find this work useful, feel free to cite us:
@article{DBLP:journals/corr/abs-2012-03704,
author = {Svitlana Vakulenko and
Vadim Savenkov and
Maarten de Rijke},
title = {Conversational Browsing},
journal = {CoRR},
volume = {abs/2012.03704},
year = {2020},
url = {https://arxiv.org/abs/2012.03704},
}
source myvenv/bin/activate
pip install -r requirements.txt
elasticsearch
To start the chatbot web UI on localhost port 8008:
python app_browse.py
python app_search.py
gunicorn --bind 0.0.0.0:8008 wsgi_browse:app &
gunicorn --bind 0.0.0.0:5008 wsgi_search:app &
Ctrl+C
otherwise
kill -9 $(sudo lsof -t -i:5008)
sudo pkill python
sudo pkill gunicorn
Datasets metadata of the Austrian Government Open Data portal (data.gv.at)
2,028 items 5 attributes: title, license, organization, categorization, tags (array 2,028+886)
2,028 items * 5 attributes (representation layers) = 10,140 information units
https://www.data.gv.at/katalog/dataset/b6385a14-b39c-4a00-a262-a8c2308759a6
-
Index table (crawl the knowledge graph)
-
Entity ranking
2.1 Frequency count (estimate percentage of the information space occupied by the entity) -
Storytelling
The agent is designed to talk as long as the user is listening.
a) Storytelling approaches
3.0. Baseline: pick entities at random.
3.1. Greedy approach to storytelling is based on a pessimistic assumption that the conversation can be interupted any time. This way the agent will always communicate the most important entities first.
3.2. Clustering approach to storytelling is based on an optimistic assumption that the conversation will last at least k number of turns. This way the agent can design an optimal combination of entities assuming k number of turns is available.
b) Optimization criteria (performance metrics of the story quality)
The agent is selecting a policy to maximize the expected reward.
Cold start + incorporating user feedback (reinforcement learning)
Information space coverage is the sum of the weights of all the entities communicated up to the ith point of time. The weights are derived from the entity ranking, e.g. counts. This measure can be converted into a proportion (%) with respect to the whole information space (normalised by the size of the input table).
We also design an information theory-inspired measure of information units per symbol to measure communication efficiency.
- Top 10 entities by count
(-1852, (u'license', u'cc-by-at-30')) (-351, (u'organization', u'stadt-linz')) (-335, (u'organization', u'stadt-wien')) (-242, (u'organization', u'land-oberoesterreich')) (-218, (u'tags', u'linz')) (-198, (u'categorization', u'finanzen-und-rechnungswesen')) (-191, (u'tags', u'haushalt')) (-188, (u'tags', u'au\xdferordentlicher')) (-188, (u'tags', u'ordentlicher')) (-186, (u'categorization', u'{umwelt}'))
- greedy approach: max counts first priority queue
license cc-by-at-30 1852 information units per message 102 information units per symbol 0.18 of the information space communicated
organization stadt-linz stadt-wien land-oberoesterreich 1390 information units per message 39 information units per symbol 0.27 of the information space communicated
tags linz 999 information units per message 38 information units per symbol 0.30 of the information space communicated
categorization finanzen-und-rechnungswesen 799 information units per message 26 information units per symbol 0.32 of the information space communicated
tags haushalt außerordentlicher ordentlicher 752 information units per message 23 information units per symbol 0.37 of the information space communicated
categorization umwelt 658 information units per message 21 information units per symbol 0.39 of the information space communicated
Total: communicated 3949 information units via 180 symbols in 6 messages 0.39 of the information space covered
- clustering approach: group the top entities by attribute
** story_size = 10
license cc-by-at-30 1852 information units per message 102 information units per symbol 0.18 of the information space communicated
organization stadt-linz stadt-wien land-oberoesterreich 1390 information units per message 39 information units per symbol 0.27 of the information space communicated
tags linz haushalt außerordentlicher ordentlicher 1188 information units per message 31 information units per symbol 0.35 of the information space communicated
categorization finanzen-und-rechnungswesen umwelt 987 information units per message 24 information units per symbol 0.39 of the information space communicated
Total: communicated 3949 information units via 162 symbols in 4 messages 0.39 of the information space covered
** story_size = 20
license cc-by-at-30 1852 information units per message 102 information units per symbol 0.18 of the information space communicated
organization stadt-linz stadt-wien land-oberoesterreich land-salzburg stadt-innsbruck 1535 information units per message 31 information units per symbol 0.30 of the information space communicated
tags linz haushalt außerordentlicher ordentlicher finanzen 1347 information units per message 26 information units per symbol 0.40 of the information space communicated
categorization finanzen-und-rechnungswesen umwelt bevoelkerung geographie-und-planung 1181 information units per message 20 information units per symbol 0.47 of the information space communicated
Total: communicated 4724 information units via 232 symbols in 4 messages 0.47 of the information space covered
- facets ranked by Gini index
(-0.75768675768675775, u'categorization') (-0.69880119880119884, u'license') (-0.67403120441801634, u'tags') (-0.64241991430320344, u'organization') (-0.019027233611292091, u'title')
categorization finanzen-und-rechnungswesen umwelt bevoelkerung geographie-und-planung verkehr-und-technik 794 information units per message 7 information units per symbol 0.08 of the information space communicated
license cc-by-at-30 cc-by other-pd cc-by-sa 1398 information units per message 20 information units per symbol 0.28 of the information space communicated
tags linz haushalt außerordentlicher ordentlicher finanzen 1255 information units per message 19 information units per symbol 0.37 of the information space communicated
organization stadt-linz stadt-wien land-oberoesterreich land-salzburg stadt-innsbruck 1246 information units per message 18 information units per symbol 0.49 of the information space communicated
title Denkmalliste Steiermark Denkmalliste Wien Ehescheidungen Polytechnische Schule Sonderschulen 999 information units per message 13 information units per symbol 0.49 of the information space communicated
Total: communicated 4995 information units via 365 symbols in 5 messages 0.49 of the information space covered
- search: results irrelevant to the user query
U: stadt wien S: There are many datasets with tags: linz, haushalt, außerordentlicher, finanzen S: For example: Rechnungsabschluss der Stadt Wien: Ausgabenart Zeitreihe Wien Offener Haushalt Budgetdaten Wien Erläuterungen Subventionen Amtsbericht Voranschlag der Landeshauptstadt Salzburg 2014 Voranschlag der Stadt Graz
Beliebteste Vornamen in Linz: bevoelkerung -> stadt-linz -> geschlecht
- Online JSON viewer for ES mapping
- Web UI is based on flask-chatterbot
- CSS Button Generator