New State-of-the-Art in Preposition Sense Disambiguation
Supervisor:
Institutions:
Project Description
The disambiguation of words is a central part of NLP tasks. In particular, there is the ambiguity of prepositions, which has been a problem in NLP for over a decade and still is. For example the preposition 'in' can have a temporal (e.g. in 2021) or a spatial (e.g. in Frankuft) meaning. A strong motivation behind the learning of these meanings are current research attempts to transfer text to artifical scenes. A good understanding of the real meaning of prepositions is crucial in order for the machine to create matching scenes.
With the birth of the transformer models in 2017 [1], attention based models have been pushing boundries in many NLP disciplines. In particular, bert, a transformer model by google and pre-trained on more than 3,000 M words, obtained state-of-the-art results on many NLP tasks and Corpus.
The goal of this project is to use modern transformer models to tackle the problem of preposition sense disambiguation. Therefore, we trained a simple bert model on the SemEval 2007 dataset [2], a central benchmark dataset for this task. To the best of our knowledge, the best purposed model for disambiguating the meanings of prepositions on the SemEval achives an accuracy of up to 88% [3]. Neither more recent approaches surpass this frontier[4][5] . Our model achives an accuracy of 90.84%, out-performing the current state-of-the-art.
How to train
To meet our goals, we cleand the SemEval 2007 dataset to only contain the needed information. We have added it to the repository and can be found in ./data/training-data.tsv.
Train a bert model:
First, install the requirements.txt. Afterwards, you can train the bert-model by:
python3 trainer.py --batch-size 16 --learning-rate 1e-4 --epochs 4 --data-path "./data/training_data.tsv"
The chosen hyper-parameters in the above example are tuned and already set by default. After training, this will save the weights and config to a new folder ./model_save/. Feel free to omit this training-step and use our trained weights directly.
Examples
We attach an example tagger, which can be used in an interactive manner. python3 -i tagger.py
Sourrond the preposition, for which you like to know the meaning of, with <head>...</head> and feed it to the tagger:
>>> tagger.tag("I am <head>in</head> big trouble")
Predicted Meaning: Indicating a state/condition/form, often a mental/emotional one that is being experienced
>>> tagger.tag("I am speaking <head>in</head> portuguese.")
Predicted Meaning: Indicating the language, medium, or means of encoding (e.g., spoke in German)
>>> tagger.tag("He is swimming <head>with</head> his hands.")
Predicted Meaning: Indicating the means or material used to perform an action or acting as the complement of similar participle adjectives (e.g., crammed with, coated with, covered with)
>>> tagger.tag("She blinked <head>with</head> confusion.")
Predicted Meaning: Because of / due to (the physical/mental presence of) (e.g., boiling with anger, shining with dew)
References
[1] Vaswani, Ashish et al. (2017). Attention is all you need. Advances in neural information processing systems. P. 5998--6008.
[2] Litkowski, Kenneth C and Hargraves, Orin (2007). SemEval-2007 Task 06: Word-sense disambiguation of prepositions. Proceedings of the Fourth International Workshop on Semantic Evaluations (SemEval-2007). P. 24--29
[3] Litkowski, Ken. (2013). Preposition disambiguation: Still a problem. CL Research, Damascus, MD.
[4] Gonen, Hila and Goldberg, Yoav. (2016). Semi supervised preposition-sense disambiguation using multilingual data. Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers. P. 2718--2729
[5] Gong, Hongyu and Mu, Jiaqi and Bhat, Suma and Viswanath, Pramod (2018). Preposition Sense Disambiguation and Representation. Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing. P. 1510--1521
1 Nov 11, 2021
586 Dec 24, 2022
3 Jun 28, 2022
353 Dec 31, 2022
5.3k Jan 04, 2023
9 Oct 11, 2022
3.4k Dec 29, 2022
227 Jan 02, 2023
12 Feb 16, 2022
16 Nov 07, 2022
160 Dec 23, 2022
16 Dec 11, 2022
95 Jan 05, 2023
21 Dec 17, 2022
45 Jan 05, 2023
438 Dec 31, 2022
19 Dec 17, 2022
2.1k Jan 01, 2023
90 Dec 27, 2022
0 Jul 24, 2022