Hugging Face's logo

zjunlp
/
knowlm-13b-ie

Text Generation
Transformers
PyTorch
llama
text-generation-inference
Inference Endpoints
Model card Files Community
1
knowlm-13b-ie / README_EN.md
ghh001's picture
ghh001
Update README_EN.md
78ee99d verified
preview code
|
raw
history blame
21.1 kB

license: apache-2.0

1.Differences from knowlm-13b-zhixi

Compared to zjunlp/knowlm-13b-zhixi, zjunlp/knowlm-13b-ie exhibits slightly stronger practicality in information extraction but with a decrease in its general applicability.

zjunlp/knowlm-13b-ie samples around 10% of the data from Chinese-English information extraction datasets, which then undergo negative sampling. For instance, if dataset A contains labels [a, b, c, d, e, f], we first sample 10% of the data from A. For a given sample 's', it might only contain labels a and b. We randomly add relationships that it doesn't originally have, such as c and d, from the specified list of relation candidates. When encountering these additional relationships, the model might output text similar to 'NAN'.This method equips the model with the ability to generate 'NAN' outputs to a certain extent, enhancing its information extraction capability while weakening its generalization ability.

2. Information Extraction Template

The template is used to construct an instruction for the input of model, consisting of three parts:

  1. Task Description: Clearly define the model's function and the task it needs to complete, such as entity recognition, relation extraction, event extraction, etc.
  2. Candidate Label List {s_schema} (optional): Define the categories of labels that the model needs to extract, such as entity types, relation types, event types, etc.
  3. Structured Output Format {s_format}: Specify how the model should present the structured information it extracts.

Template with specified list of candidate labels:

Named Entity Recognition(NER): You are an expert specialized in entity extraction. With the candidate entity types list: {s_schema}, please extract possible entities from the input below, outputting NAN if a certain entity does not exist. Respond in the format {s_format}.

Relation Extraction(RE): You are an expert in extracting relation triples. With the candidate relation list: {s_schema}, please extract the possible head entities and tail entities from the input below and provide the corresponding relation triples. If a relation does not exist, output NAN. Please answer in the {s_format} format.

Event Extraction(EE): You are a specialist in event extraction. Given the candidate event dictionary: {s_schema}, please extract any possible events from the input below. If an event does not exist, output NAN. Please answer in the format of {s_format}.

Event Type Extraction(EET): As an event analysis specialist, you need to review the input and determine possible events based on the event type directory: {s_schema}. All answers should be based on the {s_format} format. If the event type does not match, please mark with NAN.

Event Argument Extraction(EEA): You are an expert in event argument extraction. Given the event dictionary: {s_schema1}, and the event type and trigger words: {s_schema2}, please extract possible arguments from the following input. If an event argument does not exist, output NAN. Please respond in the {s_format} format.
Template without specifying a list of candidate labels 
Named Entity Recognition(NER): Analyze the text content and extract the clear entities. Present your findings in the {s_format} format, skipping any ambiguous or uncertain parts.

Relation Extraction(RE): Please extract all the relation triples from the text and present the results in the format of {s_format}. Ignore those entities that do not conform to the standard relation template.

Event Extraction(EE): Please analyze the following text, extract all identifiable events, and present them in the specified format {s_format}. If certain information does not constitute an event, simply skip it.

Event Type Extraction(EET): Examine the following text content and extract any events you deem significant. Provide your findings in the {s_format} format.

Event Argument Extraction(EEA): Please extract possible arguments based on the event type and trigger word {s_schema2} from the input below. Answer in the format of {s_format}.
Candidate Labels {s_schema} 
```json
    NER(Ontonotes): ["date", "organization", "person", "geographical social political", "national religious political", "facility", "cardinal", "location", "work of art", ...]
    RE(NYT): ["ethnicity", "place lived", "geographic distribution", "company industry", "country of administrative divisions", "administrative division of country", ...]
    EE(ACE2005): {"declare bankruptcy": ["organization"], "transfer ownership": ["artifact", "place", "seller", "buyer", "beneficiary"], "marry": ["person", "place"], ...}
    EET(GENIA): ["cell type", "cell line", "protein", "RNA", "DNA"]
    EEA(ACE2005): {"declare bankruptcy": ["organization"], "transfer ownership": ["artifact", "place", "seller", "buyer", "beneficiary"], "marry": ["person", "place"], ...}
```
Structural Output Format {s_format} 
```json
    Named Entity Recognition(NER): (Entity,Entity Type)

    Relation Extraction(RE): (Subject,Relation,Object)

    Event Extraction(EE): (Event Trigger,Event Type,Argument1#Argument Role1;Argument2#Argument Role2)

    Event Type Extraction(EET): (Event Trigger,Event Type)

    Event Argument Extraction(EEA): (Event Trigger,Event Type,Argument1#Argument Role1;Argument2#Argument Role2)
```

For a more comprehensive understanding of the templates, please refer to the files ner_converter.py、re_converter.py、ee_converter.py、eet_converter.py、eea_converter.py and configs.

3.Common relationship types 

wiki_cate_schema_en =  {
    'Person': ['place of birth', 'date of birth', 'country of citizenship', 'occupation', 'work', 'achievement', 'ancestral home', 'position', 'spouse', 'parent', 'alternative name', 'affiliated organization', 'date of death', 'sibling', 'place of death'], 
    'Geographic_Location': ['located in', 'alternative name', 'population', 'capital', 'area', 'achievement', 'length', 'width', 'elevation'], 
    'Building': ['located in', 'alternative name', 'achievement', 'event', 'creation time', 'width', 'length', 'creator', 'height', 'area', 'named after'], 
    'Works': ['author', 'publication date', 'alternative name', 'country of origin', 'based on', 'cast member', 'publisher', 'achievement', 'performer', 'director', 'producer', 'screenwriter', 'tracklist', 'composer', 'lyricist', 'production company', 'box office', 'publishing platform'], 
    'Creature': ['distribution', 'parent taxon', 'length', 'main food source', 'alternative name', 'taxon name', 'weight', 'width', 'height'], 
    'Artificial_Object': ['alternative name', 'brand', 'production date', 'made from material', 'country of origin', 'has use', 'manufacturer', 'discoverer or inventor'], 
    'Natural_Science': ['alternative name', 'properties', 'composition', 'product', 'has use', 'country of origin', 'discoverer or inventor', 'causes'], 
    'Organization': ['located in', 'alternative name', 'has subsidiary', 'date of incorporation', 'product', 'achievement', 'member', 'founded by', 'dissolution time', 'event'], 
    'Transport': ['located in', 'inception', 'connecting line', 'date of official opening', 'pass', 'area', 'alternative name', 'length', 'width', 'achievement', 'class of station'], 
    'Event': ['participant', 'scene', 'occurrence time', 'alternative name', 'sponsor', 'casualties', 'has cause', 'has effect', 'organizer', 'award received', 'winner'], 
    'Astronomy': ['alternative name', 'of', 'time of discovery or invention', 'discoverer or inventor', 'name after', 'absolute magnitude', 'diameter', 'mass'], 
    'Medicine': ['symptoms', 'alternative name', 'affected body part', 'possible consequences', 'etiology']
}

Here schema provides 12 text topics and common relationship types under the topic.

4.Datasets

Name Download Quantity Description
InstructIE Google drive
Baidu Netdisk		 20w+ InstrumentIE dataset (bilingual in Chinese and English)

The InstructIE dataset contains two core files: InstructIE-zh.json and InstructIE-en.json. Both files cover a range of fields that provide detailed descriptions of different aspects of the dataset:

  • 'id': A unique identifier for each data entry, ensuring the independence and traceability of the data items.
  • 'cate': The text's subject category, which provides a high-level categorical label for the content (there are 12 categories in total). -'text ': The text to be extracted.
  • 'relation': Represent relationship triples, respectively. These fields allow users to freely construct instructions and expected outputs for information extraction.
Explanation of each field
Field Description
id The unique identifier for each data point.
cate The category of the text's subject, with a total of 12 different thematic categories.
input The input text for the model, with the goal of extracting all the involved relationship triples.
instruction Instructions guiding the model to perform information extraction tasks.
output The expected output result of the model.
relation Describes the relationship triples contained in the text, i.e., the connections between entities (head, relation, tail).
Example of data 
```json
{
    "id": "6e4f87f7f92b1b9bd5cb3d2c3f2cbbc364caaed30940a1f8b7b48b04e64ec403", 
    "cate": "Person", 
    "input": "Dionisio Pérez Gutiérrez  (born 1872 in Grazalema (Cádiz) - died 23 February 1935 in Madrid) was a Spanish writer, journalist, and gastronome. He has been called \"one of Spain's most authoritative food writers\" and was an early adopter of the term Hispanidad.\nHis pen name, \"Post-Thebussem\", was chosen as a show of support for Mariano Pardo de Figueroa, who went by the handle \"Dr. Thebussem\".", 
    "entity": [
        {"entity": "Dionisio Pérez Gutiérrez", "entity_type": "human"}, 
        {"entity": "Post-Thebussem", "entity_type": "human"}, 
        {"entity": "Grazalema", "entity_type": "geographic_region"}, 
        {"entity": "Cádiz", "entity_type": "geographic_region"}, 
        {"entity": "Madrid", "entity_type": "geographic_region"}, 
        {"entity": "gastronome", "entity_type": "event"}, 
        {"entity": "Spain", "entity_type": "geographic_region"}, 
        {"entity": "Hispanidad", "entity_type": "architectural_structure"}, 
        {"entity": "Mariano Pardo de Figueroa", "entity_type": "human"}, 
        {"entity": "23 February 1935", "entity_type": "time"}
    ], 
    "relation": [
        {"head": "Dionisio Pérez Gutiérrez", "relation": "country of citizenship", "tail": "Spain"}, 
        {"head": "Dionisio Pérez Gutiérrez", "relation": "place of birth", "tail":"Grazalema"}, 
        {"head": "Dionisio Pérez Gutiérrez", "relation": "place of death", "tail": "Madrid"}, 
        {"head": "Mariano Pardo de Figueroa", "relation": "country of citizenship", "tail": "Spain"}, 
        {"head": "Dionisio Pérez Gutiérrez", "relation": "alternative name", "tail": "Post-Thebussem"}, 
        {"head": "Dionisio Pérez Gutiérrez", "relation": "date of death", "tail": "23 February 1935"}
    ]
}
```

5.Convert script

Training Data Transformation

Before inputting data into the model, it needs to be formatted to include instruction and input fields. To assist with this, we offer a script kg2instruction/convert.py, which can batch convert data into a format directly usable by the model.

Before using the kg2instruction/convert.py script, please ensure you have referred to the data directory. Please consult sample.json to understand the format of the data before conversion, schema.json illustrates the organization of the schema, and processed.json describes the format of the data after conversion.

python kg2instruction/convert.py \
  --src_path data/NER/sample.json \
  --tgt_path data/NER/processed.json \
  --schema_path data/NER/schema.json \
  --language zh \      # Specifies the language for the conversion script and template, options are ['zh', 'en']
  --task NER \         # Specifies the task type: one of ['RE', 'NER', 'EE', 'EET', 'EEA']
  --sample -1 \        # If -1, randomly samples one of 20 instruction and 4 output formats; if a specific number, uses the corresponding instruction format, range is -1<=sample<20
  --neg_ratio 1 \      # Sets the negative sampling ratio for all samples
  --neg_schema 1 \     # Sets the negative sampling ratio from the schema
  --random_sort        # Determines whether to randomly sort the list of schemas in the instruction

Negative Sampling: Assuming dataset A contains labels [a, b, c, d, e, f], for a given sample s, it might involve only labels a and b. Our objective is to randomly introduce some relationships from the candidate relationship list that were originally unrelated to s, such as c and d. However, it's worth noting that in the output, the labels for c and d either won't be included, or they will be output as NAN.

The schema_path specifies the path to a schema file (a JSON file). The schema file consists of three lines of JSON strings, organized in a fixed format. Taking Named Entity Recognition (NER) as an example, the meanings of each line are as follows:

["BookTitle", "Address", "Movie", ...]  # List of entity types
[]  # Empty list
{}  # Empty dictionary
More 
For Relation Extraction (RE) tasks:
[]                                                 # Empty list
["Founder", "Number", "RegisteredCapital", ...]    # List of relation types
{}                                                 # Empty dictionary


For Event Extraction (EE) tasks:
["Social Interaction-Thanks", "Organizational Action-OpeningCeremony", "Competition Action-Withdrawal", ...]        # List of event types
["DismissingParty", "TerminatingParty", "Reporter", "ArrestedPerson"]       # List of argument roles
{"OrganizationalRelation-Layoff": ["LayoffParty", "NumberLaidOff", "Time"], "LegalAction-Sue": ["Plaintiff", "Defendant", "Time"], ...}         # Dictionary of event types


For Event Type Extraction(EET) tasks:
["Social Interaction-Thanks", "Organizational Action-OpeningCeremony", "Competition Action-Withdrawal", ...]         # List of event types
[]                               # Empty list
{}                               # Empty dictionary


For Event Argument Extraction(EEA) tasks:
["Social Interaction-Thanks", "Organizational Action-OpeningCeremony", "Competition Action-Withdrawal", ...]                  # List of event types
["DismissingParty", "TerminatingParty", "Reporter", "ArrestedPerson"]           # List of argument roles
{"OrganizationalRelation-Layoff": ["LayoffParty", "NumberLaidOff", "Time"], "LegalAction-Sue": ["Plaintiff", "Defendant", "Time"], ...}             # Dictionary of event types

For more detailed information on the schema file, you can refer to the schema.json file in the respective task directories under the data directory.

Testing Data Transformation

For test data, you can use the kg2instruction/convert_test.py script, which does not require the data to contain label fields (entity, relation, event), just the input field and the corresponding schema_path.

python kg2instruction/convert_test.py \
    --src_path data/NER/sample.json \
    --tgt_path data/NER/processed.json \
    --schema_path data/NER/schema.json \
    --language zh \      
    --task NER \          
    --sample 0

Data Transformation Examples

Here is an example of data conversion for Named Entity Recognition (NER) task:

Before Transformation:
{
    "input": "In contrast, the rain-soaked battle between Qingdao Sea Bulls and Guangzhou Songri Team, although also ended in a 0:0 draw, was uneventful.",
    "entity": [{"entity": "Guangzhou Songri Team", "entity_type": "Organizational Structure"}, {"entity": "Qingdao Sea Bulls", "entity_type": "Organizational Structure"}]
}

After Transformation:
{
    "id": "e88d2b42f8ca14af1b77474fcb18671ed3cacc0c75cf91f63375e966574bd187",
    "instruction": "Please identify and list the entity types mentioned in the given text ['Organizational Structure', 'Person', 'Geographical Location']. If a type doesn't exist, please indicate it as NAN. Provide your answer in the format (entity, entity type).",
    "input": "In contrast, the rain-soaked battle between Qingdao Sea Bulls and Guangzhou Songri Team, although also ended in a 0:0 draw, was uneventful.",
    "output": "(Qingdao Sea Bulls,Organizational Structure)\n(Guangzhou Songri Team,Organizational Structure)\nNAN\nNAN"
}

Before conversion, the data format needs to adhere to the structure specified in the DeepKE/example/llm/InstructKGC/data directory for each task (such as NER, RE, EE). Taking NER task as an example, the input text should be labeled as the input field, and the annotated data should be labeled as the entity field, which is a list of dictionaries containing multiple key-value pairs for entity and entity_type.

After data conversion, you will obtain structured data containing the input text, instruction (providing detailed instructions about candidate entity types ['Organization', 'Person', 'Location'] and the expected output format), and output (listing all entity information recognized in the input in the form of (entity, entity type)).

6.Usage

We provide a script, inference.py, for direct inference using the zjunlp/knowlm-13b-ie model. Please refer to the README.md for environment configuration and other details.

CUDA_VISIBLE_DEVICES="0" python src/inference.py \
    --model_name_or_path 'models/knowlm-13b-ie' \
    --model_name 'llama' \
    --input_file 'data/NER/processed.json' \
    --output_file 'results/ner_test.json' \
    --fp16

If GPU memory is not enough, you can use --bits 8 or --bits 4.

7.Evaluate

We provide a script at evaluate.py to convert the string output of the model into a list and calculate F1

python kg2instruction/evaluate.py \
  --standard_path data/NER/processed.json \
  --submit_path data/NER/processed.json \
  --task NER \
  --language zh