linto-ai / linto-platform-nlp-keyphrase-extraction

linto-platform-nlp-keyphrase-extraction

Description

This repository is for building a Docker image for LinTO's NLP service: Keyphrase Extraction on the basis of linto-platform-nlp-core, can be deployed along with LinTO stack or in a standalone way (see Develop section in below).

LinTo's NLP services adopt the basic design concept of spaCy: component and pipeline, components (located under the folder components/) are decoupled from the service and can be easily re-used in other spaCy projects, components are organised into pipelines for realising specific NLP tasks.

This service can be launched in two ways: REST API and Celery task, with and without GPU support.

Usage

See documentation : https://doc.linto.ai

Deploy

With our proposed stack https://github.com/linto-ai/linto-platform-stack

Develop

Build and run

1 Download models into ./assets on the host machine (can be stored in other places), make sure that git-lfs: Git Large File Storage is installed and availble at /usr/local/bin/git-lfs.

cd linto-platform-nlp-keyphrase-extraction/
bash scripts/download_models.sh

2 configure running environment variables

cp .envdefault .env

Environment Variable	Description	Default Value
APP_LANG	A space-separated list of supported languages for the application	fr en
ASSETS_PATH_ON_HOST	The path to the assets folder on the host machine	./assets
ASSETS_PATH_IN_CONTAINER	The volume mount point of models in container	/app/assets
LM_MAP	A JSON string that maps each supported language to its corresponding language model	{"fr":"sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2","en":"sentence-transformers/all-MiniLM-L6-v2"}
SERVICE_MODE	The mode in which the service is served, either "http" (REST API) or "task" (Celery task)	"http"
CONCURRENCY	The maximum number of requests that can be handled concurrently	1
USE_GPU	A flag indicating whether to use GPU for computation or not, either "True" or "False"	True
SERVICE_NAME	The name of the micro-service	kpe
SERVICES_BROKER	The URL of the broker server used for communication between micro-services	"redis://localhost:6379"
BROKER_PASS	The password for accessing the broker server	None

4 Build image

sudo docker build --tag lintoai/linto-platform-nlp-keyphrase-extraction:latest .

or

sudo docker-compose build

5 Run container with GPU support, make sure that NVIDIA Container Toolkit and GPU driver are installed.

sudo docker run --gpus all \
--rm -p 80:80 \
-v $PWD/assets:/app/assets:ro \
--env-file .env \
lintoai/linto-platform-nlp-keyphrase-extraction:latest

Check running with CPU only setting

• remove --gpus all from the first command.
• set USE_GPU=False in the .env.

or

sudo docker-compose up

Check running with CPU only setting

• remove runtime: nvidia from the docker-compose.yml file.
• set USE_GPU=False in the .env.

6 If running under SERVICE_MODE=http, navigate to http://localhost/docs or http://localhost/redoc in your browser, to explore the REST API interactively. See the examples for how to query the API. If running under SERVICE_MODE=task, plese refers to the individual section in the end of this README.

Specification for http://localhost/kpe/{lang}

Supported languages

{lang}	Model	Size
en	sentence-transformers/all-MiniLM-L6-v2	80 MB
fr	sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2	418 MB

Request

{
  "articles": [
    {
      "text": "Apple Inc. is an American multinational technology company that specializes in consumer electronics, computer software and online services."
    },
    {
      "text": "Unsupervised learning is a type of machine learning in which the algorithm is not provided with any pre-assigned labels or scores for the training data. As a result, unsupervised learning algorithms must first self-discover any naturally occurring patterns in that training data set."
    }
  ]
}

Response

{
  "kpe": [
    {
      "text": "Apple Inc. is an American multinational technology company that specializes in consumer electronics, computer software and online services.",
      "keyphrases": [
        {
          "text": "apple",
          "score": 0.6539
        },
        {
          "text": "inc",
          "score": 0.3941
        },
        {
          "text": "company",
          "score": 0.2985
        },
        {
          "text": "multinational",
          "score": 0.2635
        },
        {
          "text": "electronics",
          "score": 0.2143
        }
      ]
    },
    {
      "text": "Unsupervised learning is a type of machine learning in which the algorithm is not provided with any pre-assigned labels or scores for the training data. As a result, unsupervised learning algorithms must first self-discover any naturally occurring patterns in that training data set.",
      "keyphrases": [
        {
          "text": "unsupervised",
          "score": 0.6663
        },
        {
          "text": "learning",
          "score": 0.3155
        },
        {
          "text": "algorithms",
          "score": 0.3128
        },
        {
          "text": "algorithm",
          "score": 0.2494
        },
        {
          "text": "patterns",
          "score": 0.2476
        }
      ]
    }
  ]
}

Component configuration

This is a component wrapped on the basis of KeyBERT.

Parameter	Type	Default value	Description
candidates	List[str]	null	Candidate keywords/keyphrases to use instead of extracting them from the document(s)
diversity	Float	0.5	The diversity of results between 0 and 1 if use_mmr is True
keyphrase_ngram_range	Tuple[int, int]	[1,1]	Length, in words, of the extracted keywords/keyphrases
min_df	int	1	Minimum document frequency of a word across all documents if keywords for multiple documents need to be extracted
nr_candidates	int	20	The number of candidates to consider if use_maxsum is set to True
seed_keywords	List[str]	null	Seed keywords that may guide the extraction of keywords by steering the similarities towards the seeded keywords
stop_words	Union[str, List[str]]	null	Stopwords to remove from the document
top_n	int	5	Return the top n keywords/keyphrases
use_maxsum	bool	false	Whether to use Max Sum Similarity for the selection of keywords/keyphrases
use_mmr	bool	false	Whether to use Maximal Marginal Relevance (MMR) for the selection of keywords/keyphrases

Component's config can be modified in components/config.cfg for default values, or on the per API request basis at runtime:

{
  "articles": [
    {
      "text": "Unsupervised learning is a type of machine learning in which the algorithm is not provided with any pre-assigned labels or scores for the training data. As a result, unsupervised learning algorithms must first self-discover any naturally occurring patterns in that training data set."
    }
  ],
  "component_cfg": {
    "kpe": {"keyphrase_ngram_range": [2,2], "top_n": 1}
  }
}

{
  "kpe": [
    {
      "text": "Unsupervised learning is a type of machine learning in which the algorithm is not provided with any pre-assigned labels or scores for the training data. As a result, unsupervised learning algorithms must first self-discover any naturally occurring patterns in that training data set.",
      "keyphrases": [
        {
          "text": "unsupervised learning",
          "score": 0.7252
        }
      ]
    }
  ]
}

Advanced usage

For advanced usage, such as Max Sum Similarity and Maximal Marginal Relevance for diversifying extraction results, please refer to the documentation of KeyBERT and medium post to know how it works.

Testing Celery mode locally

1 Install Redis on your local machine, and run it with:

redis-server --protected-mode no --bind 0.0.0.0 --loglevel debug

2 Make sure in your .env, these two variables are set correctly as SERVICE_MODE=task and SERVICES_BROKER=redis://172.17.0.1:6379

Then start your docker container with either docker run or docker-compose up as shown in the previous section.

3 On your local computer, run this python script:

from celery import Celery
celery = Celery(broker='redis://localhost:6379/0', backend='redis://localhost:6379/1')
r = celery.send_task(
    'kpe_task', 
    (
        'en', 
        [
            "Apple Inc. is an American multinational technology company that specializes in consumer electronics, computer software and online services.",
            "Unsupervised learning is a type of machine learning in which the algorithm is not provided with any pre-assigned labels or scores for the training data. As a result, unsupervised learning algorithms must first self-discover any naturally occurring patterns in that training data set."
        ],
        {"kpe": {"top_n": 3}}
    ),
    queue='kpe')
r.get()