Questions about [utils_opinion.py] and [prepare_dataset_A.py]

[YGJindutiao]

There's 2 questions about the codes mentioned above:

(1) In Line121 in [utils_opinion.py] , should "adj_matrices.append(adj_matrix)" be indented under “ for relation in self.relations:” ？Can this function generate an adjacency matrix for each dependency relation, i.e. eventually this function can generate dependency matrices as many as all the dependency relations‘ number for a single sentence ( is the size of the A in a batch is batchlen * relation_nums * maxlen * maxlen )?

(2) In Line53 and Line 54 in [prepare_dataset_A.py]，the path forms the following format：.../data_preprocessed/lt/train/lt_train.txt，which is different from the data_preprocessed file in IMN's code.Also, function add_split in Line 53 use read_conll in Line 242 in [utils_opinion.py],which is not suitable for the dataset in IMN(for example: Line 261-263 (line_split[0],[1],[2])).Is the dataset name different from IMN?

I am really looking forwrad to your precious answer. Thank you so much!

[XL2248]

Hi, thanks for your attention.

(1). No, we aim to generate one independent adj_matrix for each relation. See Line 77 of model.py and this may help you understand it.
A_in = [Input(shape=(overall_maxlen, overall_maxlen), name='A_in%s'%i) for i in range(num_relations)]

(2). For this, we may preprocess the dataset of IMN to the conll format (it's too long time to remember this). If so, you can change it to the conll style.

Hope it helps!