nielsr HF staff commited on
Commit
59f65cd
1 Parent(s): 593b55a

First draft of model card

Browse files
Files changed (1) hide show
  1. README.md +105 -0
README.md ADDED
@@ -0,0 +1,105 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ language: en
3
+ tags:
4
+ - tapas
5
+ - question-answering
6
+ license: apache-2.0
7
+ datasets:
8
+ - msr_sqa
9
+ ---
10
+
11
+ # TAPAS mini model fine-tuned on Sequential Question Answering (SQA)
12
+
13
+ This model has 2 versions which can be used. The default version corresponds to the `tapas_sqa_inter_masklm_mini_reset` checkpoint of the [original Github repository](https://github.com/google-research/tapas).
14
+ This model was pre-trained on MLM and an additional step which the authors call intermediate pre-training, and then fine-tuned on [SQA](https://www.microsoft.com/en-us/download/details.aspx?id=54253). It uses relative position embeddings (i.e. resetting the position index at every cell of the table).
15
+
16
+ The other (non-default) version which can be used is:
17
+ - `no_reset`, which corresponds to `tapas_sqa_inter_masklm_mini` (intermediate pre-training, absolute position embeddings).
18
+
19
+ Disclaimer: The team releasing TAPAS did not write a model card for this model so this model card has been written by
20
+ the Hugging Face team and contributors.
21
+
22
+ ## Model description
23
+
24
+ TAPAS is a BERT-like transformers model pretrained on a large corpus of English data from Wikipedia in a self-supervised fashion.
25
+ This means it was pretrained on the raw tables and associated texts only, with no humans labelling them in any way (which is why it
26
+ can use lots of publicly available data) with an automatic process to generate inputs and labels from those texts. More precisely, it
27
+ was pretrained with two objectives:
28
+
29
+ - Masked language modeling (MLM): taking a (flattened) table and associated context, the model randomly masks 15% of the words in
30
+ the input, then runs the entire (partially masked) sequence through the model. The model then has to predict the masked words.
31
+ This is different from traditional recurrent neural networks (RNNs) that usually see the words one after the other,
32
+ or from autoregressive models like GPT which internally mask the future tokens. It allows the model to learn a bidirectional
33
+ representation of a table and associated text.
34
+ - Intermediate pre-training: to encourage numerical reasoning on tables, the authors additionally pre-trained the model by creating
35
+ a balanced dataset of millions of syntactically created training examples. Here, the model must predict (classify) whether a sentence
36
+ is supported or refuted by the contents of a table. The training examples are created based on synthetic as well as counterfactual statements.
37
+
38
+ This way, the model learns an inner representation of the English language used in tables and associated texts, which can then be used
39
+ to extract features useful for downstream tasks such as answering questions about a table, or determining whether a sentence is entailed
40
+ or refuted by the contents of a table. Fine-tuning is done by adding a cell selection head on top of the pre-trained model, and then jointly
41
+ train this randomly initialized classification head with the base model on SQA.
42
+
43
+
44
+ ## Intended uses & limitations
45
+
46
+ You can use this model for answering questions related to a table in a conversational set-up.
47
+
48
+ For code examples, we refer to the documentation of TAPAS on the HuggingFace website.
49
+
50
+
51
+ ## Training procedure
52
+
53
+ ### Preprocessing
54
+
55
+ The texts are lowercased and tokenized using WordPiece and a vocabulary size of 30,000. The inputs of the model are
56
+ then of the form:
57
+
58
+ ```
59
+ [CLS] Question [SEP] Flattened table [SEP]
60
+ ```
61
+
62
+ ### Fine-tuning
63
+
64
+ The model was fine-tuned on 32 Cloud TPU v3 cores for 200,000 steps with maximum sequence length 512 and batch size of 128.
65
+ In this setup, fine-tuning takes around 20 hours. The optimizer used is Adam with a learning rate of 1.25e-5, and a warmup ratio
66
+ of 0.2. An inductive bias is added such that the model only selects cells of the same column. This is reflected by the
67
+ `select_one_column` parameter of `TapasConfig`. See also table 12 of the [original paper](https://arxiv.org/abs/2004.02349).
68
+
69
+
70
+ ### BibTeX entry and citation info
71
+
72
+ ```bibtex
73
+ @misc{herzig2020tapas,
74
+ title={TAPAS: Weakly Supervised Table Parsing via Pre-training},
75
+ author={Jonathan Herzig and Paweł Krzysztof Nowak and Thomas Müller and Francesco Piccinno and Julian Martin Eisenschlos},
76
+ year={2020},
77
+ eprint={2004.02349},
78
+ archivePrefix={arXiv},
79
+ primaryClass={cs.IR}
80
+ }
81
+ ```
82
+
83
+ ```bibtex
84
+ @misc{eisenschlos2020understanding,
85
+ title={Understanding tables with intermediate pre-training},
86
+ author={Julian Martin Eisenschlos and Syrine Krichene and Thomas Müller},
87
+ year={2020},
88
+ eprint={2010.00571},
89
+ archivePrefix={arXiv},
90
+ primaryClass={cs.CL}
91
+ }
92
+ ```
93
+
94
+ ```bibtex
95
+ @InProceedings{iyyer2017search-based,
96
+ author = {Iyyer, Mohit and Yih, Scott Wen-tau and Chang, Ming-Wei},
97
+ title = {Search-based Neural Structured Learning for Sequential Question Answering},
98
+ booktitle = {Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics},
99
+ year = {2017},
100
+ month = {July},
101
+ abstract = {Recent work in semantic parsing for question answering has focused on long and complicated questions, many of which would seem unnatural if asked in a normal conversation between two humans. In an effort to explore a conversational QA setting, we present a more realistic task: answering sequences of simple but inter-related questions. We collect a dataset of 6,066 question sequences that inquire about semi-structured tables from Wikipedia, with 17,553 question-answer pairs in total. To solve this sequential question answering task, we propose a novel dynamic neural semantic parsing framework trained using a weakly supervised reward-guided search. Our model effectively leverages the sequential context to outperform state-of-the-art QA systems that are designed to answer highly complex questions.},
102
+ publisher = {Association for Computational Linguistics},
103
+ url = {https://www.microsoft.com/en-us/research/publication/search-based-neural-structured-learning-sequential-question-answering/},
104
+ }
105
+ ```