Thread by @hadyelsahar

Tutorial 2: Story telling from structured data ans knowledge graphs #ACL2019nlp
Anirban laha @anirbanlaha Parag jain
#data2text #nlproc #NLG

@anirbanlaha Motivations for #Data2Text:
* Answer display in Question Answering systems
* kB summarization
* Question Generation

@anirbanlaha 4D perspective for #Data2Text
Paradigms, domain, tasks, Facets are 4 aspects of change.

@anirbanlaha Tutorial Part1:
Introduction to NLG
Traditional, statistical and neural methods for NLG
NLG evaluation

@anirbanlaha Traditional conventional pipeline for Data2Text #NLG
1) Content selection, 2)Document planning, 3)surface realization

@anirbanlaha Example of a rule-based NLG system "SimpleNLG"

Examples of #Data2text datasets & systems

2nd part Statistical and Neural Methods for Data2Text NLG

Statistical methods can be combined with Rule based/template methods. Through a statistical surface realizer that 1) converts meaning representations to world lattice 2) Statistical ranker is a Language Model that generates the best sentence from the lattice.

Example of an end-to-end probabilistic approach.

Using PCFGs for a generation where the grammar rules are used to generate text from a specific table structure.

* The defined grammar can be represented as a hypergraph.

* Generation is done using finding the most probable path

#ACL2019nlp

Probs. of nodes in this graph can be ranked using Inside-outside.

Finding the best overall best path can be done using Dynamic programming (e.g. Veterbi)

Parag jain, is back to explain neural models for #data2text generation

One way to adapt seq2seq models is to treat each of the records as a sequence
+ adding attention + Prior attention (fancy name for weighting matrix)

Capturing hierarchical information using RNN over the signal from attribute encoder

Copy actions:

1) Word level copy actions (shared softmax, copy/gen switch (Gulcehre et al. 2016) pointer gen. model (See et al. 2017)
arxiv.org/abs/1704.04368

Explanation of copy mechanism in Lebret et al. 2016
arxiv.org/pdf/1603.07771…

Explaining "never look back" for Nema et al. 2018 which is useful for e.g. to "once all the occupations have
been listed in the output the generator will never visit the occupation field again because there is nothing left to
say about it"
arxiv.org/pdf/1804.07789…

Content selection Gate from for #Data2Text #NLG (Puduppully et al. 2019) trained by distant supervision using information extraction
arxiv.org/pdf/1809.00582…

#ACL2019NLP

Neural Template for text generation (Sam Wiseman et al. 2019) "proposes a neural generation system using a hidden semi Markov model decoder, which learns
latent, discrete templates jointly with learning
to generate."
aclweb.org/anthology/D18-…
#ACL2019nlp

Now moving from Table2text --> Knowledge graphs to text

Where the common practice is flattening the graph into the list of triples.

Triple encoder from (Trisdey et al 2018)

Break! ☕️☕️☕️☕️☕️☕️☕️

2nd part: Storytelling Tutorial #ACL2019nlp

Evaluation methods for NLG
* word overlap metrics
* intrinsic Evaluation
* Human Evaluation

BLEU score = modified word n-gram precision x Brevity penalty

More info: en.wikipedia.org/wiki/BLEU
original paper: aclweb.org/anthology/P02-… (Papineni et al.)

Rouge: recall oriented word overlap matching
and its variants (ROUGE-N, ROUGE-L ..)

read more: en.wikipedia.org/wiki/ROUGE_(me…

Other variants: METEOR (includes paraphrases), TER, Word error rate (Levenshtein distance), NIST

Optimizing rouge using Reinforcement Learning might yield gains in rouge scores but not correlating with human evaluation (Paulus et al. 2017)
arxiv.org/abs/1705.04304

#acl2019nlp

Intrinsic evaluation using semantic similarity i.e. cosine similarity between representations of the candidate and the reference.

FYI: This btw is used in (Serban et al. 2016) for QG and referred to as Embeddings Greedy
aclweb.org/anthology/P16-…

Fluency, Adequacy, Coherence & Catchiness
Are aspects that are used for Human evaluation.

Part2: Hybrid methods for Data2text generation

Explaining placeholders for named entities / morphological analysis for NLG

two step approach for handling incoherence in end to end models for #data2text (moryossef et al. 2019)

beyond simple generation.

Controllable text generation
can we set controls while generation such as humor level, sentiment...etc

controlled natural language generation needs unsupervised methods as the combination of control variables is huge and hard to cover all of them using annotated data

unsupervised text formalization (Jain et al 2018).

Argument generations and insights about @IBM project debater

conclusions : #data2text generation should think like data scientist, artist and psychologist.

@threadreaderapp unroll