qasem-he-dictalm2-full/README.md

---
base_model: dicta-il/dictalm2.0-instruct
library_name: transformers
license: mit
language:
- he
tags:
- qasem
- hebrew
- causal-lm
- semantic-parsing
datasets:
- biu-nlp/Multilingual_QASem_Datasets
pipeline_tag: text-generation
---

# QASem Hebrew Full Model (DictaLM 2.0)

This model performs **QA-based semantic parsing (QASem) in Hebrew**.

## Overview

This repository provides a **fully fine-tuned model** for performing **QA-based semantic parsing (QASem) in Hebrew**.

QASem represents predicate–argument structure using **natural-language question–answer pairs**, rather than predefined semantic role labels. This makes the representation more interpretable and flexible across languages.

The model is based on:

**Base model:** `dicta-il/dictalm2.0-instruct`

and was fully fine-tuned for QA-based semantic parsing.


## ✨ Why this model matters

Traditional semantic role labeling methods rely on fixed label schemas and costly expert annotation.

This model takes a different approach by:

- Representing semantics using **natural-language question–answer pairs**
- Enabling **automatic dataset construction** via cross-lingual projection
- Supporting **scalable semantic parsing across languages**
- Achieving strong performance with **efficient fine-tuned models**

This makes it possible to build semantic parsers for new languages with minimal cost.

## Use Cases

This model can be used for:
- Research in **QA-based semantic parsing (QASem)** and semantic representation learning
- Extraction of **predicate–argument structures** from Hebrew text  
- Automatic **dataset creation** for training semantic models in new languages  
- Downstream NLP applications such as:
  - Information extraction  
  - Text understanding  
  - Factuality and attribution evaluation  

## Language

- Hebrew 🇮🇱


## Training Data

The model was trained on the **Multilingual QASem Dataset**:

👉 https://huggingface.co/datasets/biu-nlp/Multilingual_QASem_Datasets  

The dataset includes:

- Automatically generated QASem annotations  
- Train / Development / Test splits  
- Multiple languages: **French, Hebrew, Russian**  
- Tens of thousands of QA pairs per language  

The data was constructed using a cross-lingual projection approach, ensuring scalability across languages.


## 📄 Associated Work

This model and the underlying dataset are introduced in:
[Effective QA-Driven Annotation of Predicate-Argument Relations Across Languages](https://aclanthology.org/2026.eacl-long.112/).

The paper presents the full methodology, dataset construction process, and evaluation across multiple languages.


## 🚀 Quick Start (Recommended)

### Using the XQASem Parser

For a simple and structured interface, you can use the XQASem parser.

### Installation

```bash
pip install xqasem
```

### Basic Example

```python
from xqasem import XQasemParser

parser = XQasemParser.from_language("he")

sentences = [
    "המומחים הדגישו שהאלגוריתם החדש מאיץ משמעותית את עיבוד הבקשות המורכבות."
]

df = parser(sentences)

print(df)
```


## Output Format

The model produces structured predicate–argument representations in the form of:

- A predicate (verb or nominal)
- A natural-language question
- A corresponding answer span from the sentence

This structure can be easily converted into tabular or JSON format for downstream use.


### Example Output

| sentence | predicate | predicate_type | question | answer |
| --- | --- | --- | --- | --- |
| המומחים הדגישו שהאלגוריתם החדש מאיץ משמעותית את עיבוד הבקשות המורכבות. | הדגישו | verb | מי הדגיש משהו? | המומחים |
| המומחים הדגישו שהאלגוריתם החדש מאיץ משמעותית את עיבוד הבקשות המורכבות. | הדגישו | verb | מה מישהו הדגיש? | שהאלגוריתם החדש מאיץ משמעותית את עיבוד הבקשות המורכבות |
| המומחים הדגישו שהאלגוריתם החדש מאיץ משמעותית את עיבוד הבקשות המורכבות. | מאיץ | verb | מה מאיץ משהו? | האלגוריתם החדש |
| המומחים הדגישו שהאלגוריתם החדש מאיץ משמעותית את עיבוד הבקשות המורכבות. | מאיץ | verb | מה משהו מאיץ? | את עיבוד הבקשות המורכבות |


👉 For more details and advanced usage, see the project repository:  
https://github.com/JohnnieDavidov/xqasem

## Manual Model Loading (Advanced)

```python
from transformers import AutoTokenizer, AutoModelForCausalLM

model_id = "YonatanDavidov/qasem-he-dictalm2-full"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id)
```


## Limitations

- Performance may degrade on out-of-domain text
- Complex or ambiguous predicates may lead to inconsistent outputs
- The model is optimized for QASem-style generation and not for general-purpose text generation


## 📄 Citation

If you use this model, please cite our work:
```
@inproceedings{davidov-etal-2026-effective,
    title = "Effective {QA}-Driven Annotation of Predicate{--}Argument Relations Across Languages",
    author = "Davidov, Jonathan  and
      Slobodkin, Aviv  and
      Klein, Shmuel Tomi  and
      Tsarfaty, Reut  and
      Dagan, Ido  and
      Klein, Ayal",
    editor = "Demberg, Vera  and
      Inui, Kentaro  and
      Marquez, Llu{\'i}s",
    booktitle = "Proceedings of the 19th Conference of the {E}uropean Chapter of the {A}ssociation for {C}omputational {L}inguistics (Volume 1: Long Papers)",
    month = mar,
    year = "2026",
    address = "Rabat, Morocco",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2026.eacl-long.112/",
    doi = "10.18653/v1/2026.eacl-long.112",
    pages = "2484--2502",
    ISBN = "979-8-89176-380-7",
}
```