We Used Three Approches To Find Adequate Solution For Given Problem As Follows

Approch 1 : Abstractive

Neural sequence-to-sequence models have provided a viable new approach for abstractive text summarization (meaning they are not restricted to simply selecting and rearranging passages from the original text). However, these models have two shortcomings: they are liable to reproduce factual details inaccurately, and they tend to repeat themselves. See et. al. [5] propose a novel architecture of hybrid pointer-generator network that can copy words from the source text via pointing, which aids accurate reproduction of information, while retaining the ability to produce novel words through the generator. The network can be viewed as a balance between extractive and abstractive approaches. We use coverage to keep track of what has been summarized, which discourages repetition. The ability to produce OOV(Out-of-vocabulary) words is one of the primary advantages of pointer-generator models.A processed data will be fed to the model for training. Model is Encoder-Decoder model with attention mechanism. For the encoder LSTM cell with stack_bidirectional_dynamic_rnn is used. For the decoder step LSTM with BasicDecoder for training, and BeamSearchDecoder for inference has been used. We use the code available at [3] and [2]. Approach 3: SVO based approach In this approach, we experiment with linguistic information of a news article. We take inspiration from LexRank algorithm that ranks the sentences in terms of their importance using graph based information. We first resolve the coreferences from the article and extract tuples. We then rank these tuples using their importance score. We are experimenting with different scoring and ranking techniques including TextRank and LexRank. A tweet can be generated by using the high importance tuples and corresponding highly ranked sentence having that tuple

Approach 2 : SVO based approach

In this approach, we experiment with linguistic information of a news article. We take inspiration from LexRank algorithm that ranks the sentences in terms of their importance using graph based information. We first resolve the coreferences from the article and extract tuples. We then rank these tuples using their importance score. We are experimenting with different scoring and ranking techniques including TextRank and LexRank. A tweet can be generated by using the high importance tuples and corresponding highly ranked sentence having that tuple

Approch 3 : Topic sentence for tweets generation

In this approach we follow the work of Putra et.al, where it compares the results of neural headline generation with topic sentence and first sentence as input to a encoder-decoder model. We modified the work to generate the tweets as there is a striking resemblance between the two. Idea revolves around to reduce the input from full article text to only important sentences to avoid large number of words which could cause vanishing gradient problem, We have used Spacy, a powerful python module, along with regular expression to extract the topic sentence from each news article. Also for encoder-decoder, we have used vanilla implementation of it in OpenNMT with default parameter configuration . We are also experimenting with different types of inputs for this approach such as using one or multiple summaries in place of first sentence or topic sentence