RoBERTa Fine-Tuned for Event Expression Extraction (EVENTX)
Overview
This model is a fully fine-tuned version of FacebookAI/roberta-base for Event Expression Extraction (EVENTX).
The model identifies spans in natural language text that correspond to events, including actions, occurrences, processes, and states expressed in a sentence.
The task is formulated as token-level sequence labeling using BIO tags.
Label Map
| ID | Label | Description |
|---|---|---|
| 0 | O | Token is outside an event expression |
| 1 | B-E | Beginning of an event expression |
| 2 | I-E | Inside an event expression |
Task
Event Expression Extraction (EVENTX)
Given an input sentence, the model identifies spans corresponding to event expressions.
For example:
Input:
The company announced the acquisition on Monday.
Event expressions:
announced
acquisition
Each token is assigned one of three BIO labels:
- O: outside an event expression
- B-E: beginning of an event expression
- I-E: inside an event expression
Event expression extraction provides structured information about events mentioned in text and can support downstream temporal reasoning, event ordering, relation extraction, and timeline construction.
Model
RoBERTa is an encoder-only transformer architecture. This model uses RoBERTa with a token-classification head and is fully fine-tuned for sentence-level event expression extraction.
The model is one of the fine-tuned pretrained language model baselines used to study multi-dimensional generalization in temporal extraction tasks.
Dataset
Dataset: mdg-nlp/eventx-recognition-original
Labels
0 -> O
1 -> B-E
2 -> I-E
Training Details
| Configuration | Value |
|---|---|
| Base Model | FacebookAI/roberta-base |
| Task | Event Expression Extraction |
| Formulation | Token Classification |
| Label Scheme | BIO |
| Fine-tuning | Full fine-tuning |
| Epochs | 3 |
| Learning Rate | 2e-5 |
| Optimizer | AdamW |
| Weight Decay | 0.01 |
| Warmup Steps | 500 |
| Train Batch Size | 16 |
| Evaluation Batch Size | 32 |
| Max Sequence Length | 128 |
Evaluation
All reported scores use micro-F1.
| Evaluation Setting | Micro-F1 |
|---|---|
| Base | 0.85 |
| T2C | 0.19 |
| Adversarial (Adv) | 0.82 |
| Length | 0.82 |
| Vocabulary (Voc) | 0.85 |
| Adversarial Perturbation (Adv_P) | 0.50 |
The evaluation settings measure generalization across the original task, cross-domain data, adversarial examples, structural changes, vocabulary variation, and adversarial perturbations.
Usage
from transformers import AutoTokenizer, AutoModelForTokenClassification
model_id = "mdg-nlp/RoBERTa-eventx-recognition-sentence"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForTokenClassification.from_pretrained(model_id)
text = "The company announced the acquisition on Monday."
inputs = tokenizer(text, return_tensors="pt", truncation=True)
outputs = model(**inputs)
predictions = outputs.logits.argmax(dim=-1)[0].tolist()
id2label = {
0: "O",
1: "B-E",
2: "I-E"
}
tokens = tokenizer.convert_ids_to_tokens(inputs["input_ids"][0])
for token, prediction in zip(tokens, predictions):
print(token, id2label[prediction])
Limitations
- Performance may decrease under distribution shifts.
- Cross-domain event expressions may be more difficult to identify.
- Adversarial perturbations can affect extraction performance.
- Tokenization can affect predicted event boundaries.
- Longer or structurally different inputs may produce different behavior.
- Ambiguous or implicit events may not be identified correctly.
- The model extracts event expressions and does not directly perform event relation or temporal relation classification.
Research Context
This model was developed as part of research on multi-dimensional generalization in temporal extraction tasks.
The broader work studies temporal and event expression extraction under different forms of distribution shift, including domain shift, adversarial perturbations, vocabulary variation, compositional changes, and structural changes.
Contact
Author: Arnav Verma Affiliation: New York University
Citation
If you use this model in your research, please cite the associated work when available.