Semantic Role Labelling For Hindi

Github Repo Link

https://github.com/rahcode7/SRL

Semantic Role Labelling

Semantic Role Labelling, in natural language processing, is a process that assigns labels to different words in a sentence that indicate their semantic role in the sentence. This helps in finding the meaning of the sentence, and more importantly, the role of a particlar word in creating that meaning of the sentence. The task essentially boils down to identifying the various arguments associated with the predicate or the main verb of the sentence and assigning them specific roles.

Example

---

The above example has 3 distinct labels that can be seen - Agent, Theme, and the Location. It also has the predicate labelled. Using these labels we are then able to answer the question "Who did what to whom where?"

Some of the more common labels are -

• Agent
• Experiencer
• Theme
• Result
• Location

More labels (not exhaustive) can be found in these slides

Problem Classification

---

The problem can the be further decomposed into the following -

Predicate Detection: Findint the predicate in a given sentence.

Predicate Sense Disambiguation: Disambiguating the sense of the predicate found.

Argument Identification: Identifying the arguments for the given predicate for the given sense.

Argument Classification: Assigning the labels to the arguments found.

Hindi

---

For Semantic Role Labelling in Hindi, we will be labelling the words into the following roles:

Label	Description
ARG0	Agent, Experiencer, or doer
ARG1	Patient or Theme
ARG2	Beneficiary
ARG3	Instrument
ARG2-ATR	Attribute or Quality
ARG2-LOC	Physical Location
ARG2-GOL	Goal
ARG2-SOU	Source
ARGM-PRX	Noun-Verb Construction
ARGM-ADV	Adverb
ARGM-DIR	Direction
ARGM-EXT	Extent or Comparision
ARGM-MNR	Manner
ARGM-PRP	Purpose
ARGM-DIS	Discourse
ARGM-LOC	Abstract Location
ARGM-MNS	Means
ARGM-NEG	Negation
ARGM-TMP	Time
ARGM-CAU	Cause or Reason

The following labels have been taken from this paper.

Code Walkthrough

Requirements

torch==1.8.1. numpy==1.19.5. matplotlib==3.3.4. seaborn==0.11.1. pandas==1.1.5. scikit_learn==0.24.2.

Train a model

Download dataset

1.Download the dataset from the link - https://drive.google.com/drive/folders/1JLrZ0HgvuXKZL7PhB5Y89V_4BzIp7ucV?usp=sharing
2.Unzip the file and place it in data/processed folder

Run following commands
cd src

1. For logistic regression model

python classifier_base.py

2. For Bidirectional LSTM model

python SRL_NN_train.py --EMBEDDING_DIM=300 --NUM_HIDDEN_NODES=100 --epochs=50 --batchsize=64 --learning_rate=0.001

Hyperparameters

The following table lists optimization/training hyperparameters for the neural LSTM based SRL model

Name	Type	Description	Default value
learning_rate	float	Initial learning rate.	0.001
EMBEDDING_DIM	int	Dimensionality of the word vectors	300
NUM_HIDDEN_NODES	int	Number of hidden nodes of the neural model	100
epochs	int	Number of times datasets needs to be iterated for a model	50
batchsize	int	Size of a batch of training examples sampled from a dataset	64

Model Checkpoints/Trained models

Our models are stored at models/srl_hindi_bilstm_50e.pth

Results

Final Model - Bidirectional LSTM