metadata

dataset_info:
  - config_name: ADG
    features:
      - name: text
        dtype: string
      - name: target_entity
        dtype: string
      - name: label
        dtype: int64
    splits:
      - name: train
        num_bytes: 729565
        num_examples: 3201
      - name: validation
        num_bytes: 168501
        num_examples: 759
      - name: test
        num_bytes: 114693
        num_examples: 470
    download_size: 346052
    dataset_size: 1012759
  - config_name: WN
    features:
      - name: text
        dtype: string
      - name: target_entity
        dtype: string
      - name: label
        dtype: int64
    splits:
      - name: train
        num_bytes: 3010007
        num_examples: 14331
      - name: validation
        num_bytes: 665886
        num_examples: 3320
      - name: test
        num_bytes: 862510
        num_examples: 3463
    download_size: 1253281
    dataset_size: 4538403
configs:
  - config_name: ADG
    data_files:
      - split: train
        path: ADG/train-*
      - split: validation
        path: ADG/validation-*
      - split: test
        path: ADG/test-*
  - config_name: WN
    data_files:
      - split: train
        path: WN/train-*
      - split: validation
        path: WN/validation-*
      - split: test
        path: WN/test-*

Named-Entities Recognition on Multi-Domain Documents (NERMUD)

Original paper: https://iris.unitn.it/retrieve/d833b9e4-e997-4ee4-b6aa-f5144a85f708/paper42.pdf

NERMuD is a task presented at EVALITA 2023 consisting in the extraction and classification of named-entities in a document, such as persons, organizations, and locations.

This dataset comes as a word level classification setting, we decided to reframe the task to be prompted to a generative LLMs as a multiclass classification without the extractive part. The prompt will be a text and one of the possible named entities contained in it, then the model is asked to return the correct class of the Named entity (Location, Organization, Person).

To do this, we generated a different sample for each Named-Entity in the original dataset, the data come in an Inside-Outside-Begining (IOB) tagging scheme:

Original format	Output Format
"L'":O, "astronauta":O, "Umberto":B-PER, "Guidoni":I-PER, "dell'":O "Agenzia":B-ORG, "Spaziale":I-ORG	("L'astronauta Umberto Guidoni dell'Agenzia Spaziale", "Umberto Guidoni", PERSONA), ("L'astronauta Umberto Guidoni dell'Agenzia Spaziale", "Agenzia Spaziale", ORGANIZZAZIONE)

Nermud come with three different Domains: ADG (Alcide De Gasperi writings), FIC (Fiction Books), and WN (Wikinews). After our reframing, with additional cleaning strategies (e.g. remove duplicate, ambiguous samples, ...) we decided to mantain only AGD and WN domains, since FIC ended up to be high unbalanced on the test set.

Example

Here you can see the structure of the single sample in the present dataset.

{
  "text": string, # text of the sentence
  "target_entity": string, # text of the entity to classify
  "label": int, # 0: Luogo, 1: Organizzazione, 2: Persona
}

Statitics

NERMUD WN	Luogo	Persona	Organizzazione
Training	4661	5291	4379
Validation	1217	1056	1047
Test	859	1373	1231

NERMUD AGD	Luogo	Persona	Organizzazione
Training	891	839	1471
Validation	220	198	341
Test	97	162	221

Proposed Prompts

Here we will describe the prompt given to the model over which we will compute the perplexity score, as model's answer we will chose the prompt with lower perplexity. Moreover, for each subtask, we define a description that is prepended to the prompts, needed by the model to understand the task.

Description of the task: "Data una frase e un'entità, indica se tale entità rappresenta un luogo, un'organizzazione o una persona.\n\n"

Cloze style:

Label (Luogo): "Data la frase: '{{text}}'\nL'entità {{target_entity}} è un luogo"

Label (Persona): "Data la frase: '{{text}}'\nL'entità {{target_entity}} è una persona"

Label (Organizzazione): "Data la frase: '{{text}}'\nL'entità {{target_entity}} è un'organizzazione"

MCQA style:

Data la frase: \"{{text}}\"\nDomanda: A quale tipologia di entità appartiene \"{{target_entity}}\" nella frase precedente?\nA. Luogo\nB. Organizzazione\nC. Persona\nRisposta:

Results

The following results are given by the Cloze-style prompting over some english and italian-adapted LLMs.

NERMUD (AVG)	ACCURACY (5-shots)
Gemma-2B	55.25
QWEN2-1.5B	65.82
Mistral-7B	83.42
ZEFIRO	83.24
Llama-3-8B	85.64
Llama-3-8B-IT	89.5
ANITA	88.43

Acknowledge

We would like to thank the authors of this resource for publicly releasing such an intriguing benchmark.

Further, We want to thanks the students of MNLP-2024 course, where with their first homework tried different interesting prompting strategies, and different reframing strategies that able us to generate this resource.

The original dataset is freely available for download link

License

All the texts used are publicly available, under a license that permits both research and commercial use. In particular, the texts used for NERMuD are taken from:

Wikinews (WN) as a source of news texts belonging to the last decades;

Writings and speeches from the Italian politician Alcide De Gasperi (ADG).