Llama-3-8B-multilingual-temporal-expression-normalization

Model Card for Model ID

This model provides a ready-to-go temporal expression normalization tool.

Model Details

Model Description

We present a series of multilingual temporal expression normalization models. Proposed in the {paper}.

This model has been trained over Timebank and E3C multilingual corpora with the sole objevtive of normalizing temporal expressions following the ISO TimeML schema.

A temporal expression is classified into four types: Date, Time, Duration and Set. This model predicts the value of a temporal expression. For example for the expression of type duration "8 hours" the model would predict the value PT8H.

• Developed by: Alejandro Sánchez de Castro, Lourdes Araujo and Juan Martínez-Romo
• Language(s) (NLP): Multilingual

Uses

Downstream Use [optional]

For using the model just follow the general PEFT guide. When using the model please use the following prompt:

Given the temporal expression type the reference date, the temporal expression and the context phrase in which the temporal expression appears, generate the value according to the TIMEX3 scheme. In the context phrase the temporal expression appears, sometimes it will be neccesary to pay attention to the surrounding context in order to resolve the expression.

###Context phrase: {sentence}

###Temporal expression type: {type}

###Temporal expression: {expression}

###Reference date {dct}

###Value:

The model will then return a sequence with the predicted value. It can be extracted through a simple pattern match.

Evaluation

We compare our proposed models against current multilingual solutions as shown in the {paper}

Results

We managed to get an outstanding performance on over 7 languages and a promising zero-shot performance over non-trained languages.

Summary

Citation [optional]

To be defined

BibTeX:

To be defined

APA:

To be defined

Acklowledge:

• Funded by: funded by the following projects DOTT-HEALTH (MCI/AEI/FEDER, UE with identification PID2019-106942RB-C32), OBSER-MENH (MCIN/AEI/10.13039/501100011033 and NextGenerationEU”/PRTR with identification TED2021-130398B-C21), SICAMESP (with identification 2023-VICE-0029) and by the project EDHER-MED (with identification PID2022-136522OB-C21)"

Model Sources [optional]

• Paper: To be defined

Model Card Contact

[email protected]

Framework versions

• PEFT 0.10.1.dev0