Povzemanje besedil
V tem repozitoriju se nahaja rezultat aktivnosti A3.4 - R3.4.1 Orodje za povzemanje besedil, ki je nastalo v okviru projekta Razvoj slovenščine v digitalnem okolju.
Available models
At the moment, 6 models are available in this repository (in the src folder): graph-based, metamodel, t5-article, t5-headline, hybrid-long, and sumbasic. Each model can be run separately, or together with the metamodel (see the instructions below). The role of the metamodel is to automatically select the most appropriate summarization model, based on the analysis of the input text.
A brief description of models:
| Model | Description |
|---|---|
| metamodel | neural model based on the Doc2Vec document representation (docs/metamodel.pdf), suggests the best summarizer |
| graph-based [1] | unsupervised extractive approach, graph-based, returns n most important sentences |
| t5-headline | supervised abstractive approach, transformer-based, returns headlines |
| t5-article | supervised abstractive approach, transformer-based, returns short summaries |
| sumbasic | unsupervised, simple summarizer that works with word frequencies, returns n most important sentences |
| hybrid-long | unsupervised hybrid approach, graph-based & transformer-base, returns short summaries of long texts |
Download pre-trained models
Download models from link and extract them into the folders with their source code. For example, the model t5-article should be extracted in the src/t5-article/model folder.
Run locally
Use each model separately
We suggest using virtual environments and python 3.8 for all models. Dependencies for each model can be installed with pip install -r requirements.txt from the root folder of models. After you installed all dependencies, you can run inference in two ways:
- Command-line interface:
python main-cli.py --text 'the content of an article' - Uvicorn server:
uvicorn main-fastapi:app --host 0.0.0.0 --port xxxx
After the models are up and running, you can access Swagger UI documentation at http://localhost:xxxx/docs (xxxx is the port number). You will find the details and examples of both ways in the README.md file of each model.
Use metamodel
To use the metamodel, you have to start each model as a uvicorn server with the following ports:
| Model | Port |
|---|---|
| metamodel | 5003 |
| graph-based | 5004 |
| t5-headline | 5005 |
| t5-article | 5006 |
| sumbasic | 5007 |
| hybrid-long | 5008 |
main.sh is an example of a script that automates the manual creation of uvicorn servers. Set the --root-path as your absolute path to the root of this project and run y | ./main.sh --root-path /absolute/path/to/RSDO-DS3/. This will automatically create conda environments for all models.
After all six models are up and running, you can call only the metamodel, which will automatically select the best summarizer based on the analysis of the input text. You can find example requests in the commands.sh script in each .
Run a docker
Models can be run with Docker as well. The instructions are in the README.md file of each model.
Run a docker-compose
This is the most straightforward method to run the project, but requires the most resources. First, make sure that you have downloaded and extracted pre-trained models to the specified folders. Next, run docker-compose up. This will build the environment defined in the docker-compose.yml file. It will take some time to build the full environment. After that, you can call the metamodel or single models. For example, to call the metamodel, use the following command:
curl -X POST --location "http://localhost:5003/auto-select/" \ -H "Content-Type: application/json" \ -d "{\"text\": \"Marsikdo nam zavida hitrost pri oblikovanju koalicije, a volja ljudi je bila jasna. Če ne bi bila, se nam ne bi uspelo tako hitro dogovoriti, kateri so projekti, smeri in vrednote, ki jih bomo skupaj zagovarjali v prihodnji vladi, je ob podpisu pogodbe dejal verjetni mandatar in predsednik Gibanja Svoboda Robert Golob. Nove organizacije vlade, ki jo pogodba predvideva, še ne morejo takoj udejanjiti, saj jih je ustavil SDS-ov predlog za posvetovalni referendum o zakonu o vladi, a Golob zatrjuje, da bodo to storili v prihodnjih mesecih. Na videz se povečuje kompleksnost vlade, ker se dodajajo nova ministrstva, a v resnici so ta nova ministrstva namenjena ravno tistemu, kar bo našo vlado razlikovalo od prejšnjih. Namenjena so ustvarjanju novih priložnosti, projektov in znanj, je pojasnil. Z ministrstvom za visoko šolstvo, znanost in inovacije, ministrstvom za solidarno prihodnost in ministrstvom zeleni preboj bodo po njegovih besedah omogočili, da bo Slovenija kot država odporna proti spremembam, ki jih prinaša prihodnost. Tudi predsednica SD-ja Tanja Fajon je zatrdila, da so oblikovali vlado sprememb. Naš cilj je, da Sloveniji zagotovimo močno gospodarstvo, socialno varnost za vse, skladen regionalni razvoj in Slovenijo v jedru Evrope. Nova vlada bo usmerjena v dvig dodane vrednosti, v zeleni in digitalni prehod ter v močne javne storitve. Tudi v mednarodni politiki želimo vrniti ugled državi, kjer je bil ta poškodovan. Po besedah koordinatorja Levice Luke Mesca je bilo namreč zadnje desetletje desetletje izgubljenih priložnosti, ko je Slovenija prehajala iz krize v krizo. Ta koalicijska pogodba je za dva mandata, da do leta 2030 ljudem organiziramo državo, kakršno si zaslužijo, je dodal.\" }"
To call a single model, check the specified ports defined above and use the following command, e.g., t5-headline:
curl -X POST --location "http://localhost:5005/summarize/" \ -H "Content-Type: application/json" \ -d "{\"text\": \"Marsikdo nam zavida hitrost pri oblikovanju koalicije, a volja ljudi je bila jasna. Če ne bi bila, se nam ne bi uspelo tako hitro dogovoriti, kateri so projekti, smeri in vrednote, ki jih bomo skupaj zagovarjali v prihodnji vladi, je ob podpisu pogodbe dejal verjetni mandatar in predsednik Gibanja Svoboda Robert Golob. Nove organizacije vlade, ki jo pogodba predvideva, še ne morejo takoj udejanjiti, saj jih je ustavil SDS-ov predlog za posvetovalni referendum o zakonu o vladi, a Golob zatrjuje, da bodo to storili v prihodnjih mesecih. Na videz se povečuje kompleksnost vlade, ker se dodajajo nova ministrstva, a v resnici so ta nova ministrstva namenjena ravno tistemu, kar bo našo vlado razlikovalo od prejšnjih. Namenjena so ustvarjanju novih priložnosti, projektov in znanj, je pojasnil. Z ministrstvom za visoko šolstvo, znanost in inovacije, ministrstvom za solidarno prihodnost in ministrstvom zeleni preboj bodo po njegovih besedah omogočili, da bo Slovenija kot država odporna proti spremembam, ki jih prinaša prihodnost. Tudi predsednica SD-ja Tanja Fajon je zatrdila, da so oblikovali vlado sprememb. Naš cilj je, da Sloveniji zagotovimo močno gospodarstvo, socialno varnost za vse, skladen regionalni razvoj in Slovenijo v jedru Evrope. Nova vlada bo usmerjena v dvig dodane vrednosti, v zeleni in digitalni prehod ter v močne javne storitve. Tudi v mednarodni politiki želimo vrniti ugled državi, kjer je bil ta poškodovan. Po besedah koordinatorja Levice Luke Mesca je bilo namreč zadnje desetletje desetletje izgubljenih priložnosti, ko je Slovenija prehajala iz krize v krizo. Ta koalicijska pogodba je za dva mandata, da do leta 2030 ljudem organiziramo državo, kakršno si zaslužijo, je dodal.\" }"
Train your models
In this section, you can find the complete instructions for training and building your models from scratch.
Datasets and pre-trained models
To train our models, we used SURS (Statistical Office of the Republic of Slovenia) and STA (Slovene Press Agency) datasets which can not be publicly released. We also used KAS corpus which has limited availability and AutoSentiNews which is in the public domain and can be freely used. For the pre-trained models, we used t5-sl-small as our pre-trained Slovene text generator and LaBSE as a multilingual sentence encoder.
Metamodel
metamodel is trained on the KAS corpus, AutoSentiNews, STA, and SURS and consists of two independent models: Doc2vec and the main model. Before training, please make sure that you have installed DVC on your system. The dvc.yaml file contains all stages that need to be executed to get the final model. The params.yaml file contains parameter values for each of the corresponding stages. The stages get executed with the dvc run command within your terminal in the root directory. Before running this command, please make sure that you have defined all the paths to the datasets described in the previous section in the get_data.py file. Alternatively, you can also run commands independently of DVC by executing the python command defined within each stage.
Article summarization model
t5-article was trained on a Slovene machine-translated CNN / DailyMail dataset which can be downloaded from link. The original English dataset can be accessed through huggingface. The CNN / DailyMail dataset is an English-language dataset containing approximately 300k unique news articles written by journalists at CNN and the Daily Mail. The dataset consists of texts and highlights that are used as short summaries of an article. You can find the script train.py for training and evaluating the model in this repository in the folder src. If you want to train your model, download the pre-trained t5-sl-small, the Slovene machine-translated dataset, and set the correct paths in the script. After that, run python src/train.py. The trained summarization model can also be used as an abstractor model in the hybrid-long summarization model.
Graph-based model
graph-based is an unsupervised model that uses only pre-trained sentence encoder LaBSE and a script to calculate sentence centrality scores.
Hybrid-long summarization model
hybrid-long is a combination of abstractive and extractive approaches. It combines t5-article and graph-based.
SumBasic model
SumBasic is an unsupervised basic summarization model.
Headline generator
t5-headline was trained on STA articles as a source and their headlines as target data in a similar way to t5-article except for a few different parameter values. In the script src/train.py, change the paths to the pre-trained t5-sl-small model and your datasets. After that, run python src/train.py.
References
[1] Aleš Žagar and Marko Robnik-Šikonja. 2021. Unsupervised Approach to Multilingual User Comments Summarization. In Proceedings of the EACL Hackashop on News Media Content Analysis and Automated Report Generation, pages 89–98, Online. Association for Computational Linguistics.
Operacijo Razvoj slovenščine v digitalnem okolju sofinancirata Republika Slovenija in Evropska unija iz Evropskega sklada za regionalni razvoj. Operacija se izvaja v okviru Operativnega programa za izvajanje evropske kohezijske politike v obdobju 2014-2020.
