ReCAM

Our repository for the code, literature review, and results on SemEval-2021 Task 4: Reading Comprehension of Abstract Meaning

Usage

1. Run the setup.py file for imports to work : python setup.py install

Tasks

Imperceptibility

• Approaches:
  • Plain GAReader (Baseline)
  • Common-sense Reasoning
  • Transformers-Based Model for Cloze-Style Question Answering
  • Incorporating The Imperceptibility Rating from the Concreteness Rating Dataset
    • The dataset is present in data/Imperceptibility/Concreteness Ratings.

Non-Specificity

• Approaches:
  • Use WordNet's hypernyms, augment the dataset (replace nouns with their hypernyms), pretrain BERT on this augmented dataset.

    • Code:

       python -m nltk.downloader 'popular'
       pip install -U pywsd
      

      and use the augment(sent) function in src/non-specificity/hypernym/augment.py.

    • Code for Finetuning BERT on the Augmented Dataset

To-Do

• Exploratory Data Analysis

  • Find Most Common Words
  • The GloVe/CoVe representation of words in passage, words in question, and words in answer (Couple of examples).
  • Word Split Based EDA (Stats)
  • Bert Tokenizer Based EDA (Stats)
  • WordClouds

• Build Basic Q/A system

  • GA Reader

• Literature Reviews

Data

Statistics

Word/Char Python Split-Based

	Task_1_train	Task_1_dev	Task_2_train	Task_2_dev
Samples	3227	837	3318	851
Duplicated Examples	0	0	0	0
Duplicated Articles	57	4	55	2
Mean Char Count (Articles)	1548.47	1582.79	2448.93	2501.41
Std Char Count (Articles)	905.301	992.333	1788.42	1809.55
Max Char Count (Articles)	10177	9563	10370	9628
Min Char Count (Articles)	162	255	169	247
Mean Word Count (Articles)	262.44	268.608	417.727	427.048
Std Word Count (Articles)	154.534	171.253	307.248	310.994
Max Word Count (Articles)	1754	1641	1700	1651
Min Word Count (Articles)	31	41	31	45
Duplicated Questions	0	0	1	0
Mean Char Count (Questions)	137.826	137.455	148.234	149.385
Std Char Count (Questions)	31.1171	30.6809	47.7494	46.9035
Max Char Count (Questions)	389	387	443	413
Min Char Count (Questions)	32	48	31	37
Mean Word Count (Questions)	24.6771	24.4695	26.8852	27.2127
Std Word Count (Questions)	6.18453	6.04612	9.1822	9.16159
Max Word Count (Questions)	73	73	83	83
Min Word Count (Questions)	6	9	7	6

BertTokenizer Based

	Task_1_train	Task_1_dev	Task_2_train	Task_2_dev
Mean Token Count (Articles)	332.603	341.452	528.58	539.303
Std Token Count (Articles)	193.729	217.673	387.756	396.286
Max Token Count (Articles)	2206	2043	2188	2272
Min Token Count (Articles)	38	52	46	52
Mean Token Count (Questions)	28.6319	28.3441	30.9629	31.1657
Std Token Count (Questions)	6.83765	6.75112	10.0225	9.98314
Max Token Count (Questions)	84	84	91	89
Min Token Count (Questions)	9	11	9	8

Note :

• Word is based on .split() on Python strings, including all the words (stopwords). We can also use torch/nltk tokenizers to get tokens and check their lengths.
• The spaces are also included in character counting.
• No preprocessing was done for either of the statistics.
• For questions, we do not remove @placeholder.

Most Frequent Words

Note :

• Tokens were made using NLTK Tokenizer, and only alphanumeric characters, after removing stop words were used.

Articles

	Task_1_train	Task_1_dev	Task_2_train	Task_2_dev
Unique Token Count	41087	20447	53285	25779
Total Count	482713	128258	775342	203877

	Task_1_train			Task_1_dev			Task_2_train			Task_2_dev
	Word	Count	Frequency	Word	Count	Frequency	Word	Count	Frequency	Word	Count	Frequency
0	said	7978	0.0165274	said	1940	0.0151258	said	9594	0.0123739	said	2524	0.01238
1	mr	2465	0.00510655	mr	701	0.00546555	mr	3549	0.00457733	mr	993	0.00487058
2	would	2194	0.00454514	would	593	0.00462349	would	3539	0.00456444	would	954	0.00467929
3	also	2019	0.00418261	also	506	0.00394517	people	3335	0.00430133	people	872	0.00427709
4	people	1897	0.00392987	one	486	0.00378924	one	3182	0.004104	one	814	0.0039926
5	one	1705	0.00353212	people	481	0.00375025	also	2784	0.00359067	also	675	0.00331082
6	last	1461	0.00302664	first	395	0.00307973	says	2201	0.00283875	two	562	0.00275656
7	two	1349	0.00279462	last	391	0.00304854	two	2112	0.00272396	says	547	0.00268299
8	first	1344	0.00278426	new	372	0.0029004	years	2084	0.00268785	new	545	0.00267318
9	new	1341	0.00277805	told	349	0.00272108	new	2081	0.00268398	time	532	0.00260942
10	year	1262	0.00261439	two	346	0.00269769	time	2064	0.00266205	us	531	0.00260451
11	years	1248	0.00258539	us	342	0.0026665	could	2002	0.00258209	could	517	0.00253584
12	could	1222	0.00253152	years	340	0.00265091	last	1957	0.00252405	years	516	0.00253094
13	told	1203	0.00249216	year	329	0.00256514	first	1944	0.00250728	first	509	0.0024966
14	us	1187	0.00245902	time	307	0.00239361	us	1820	0.00234735	told	493	0.00241812
15	time	1159	0.00240101	could	291	0.00226886	year	1768	0.00228028	year	462	0.00226607
16	government	1036	0.0021462	bbc	256	0.00199598	like	1727	0.0022274	last	455	0.00223174
17	made	940	0.00194733	added	246	0.00191801	told	1659	0.0021397	bbc	418	0.00205026
18	bbc	894	0.00185203	made	236	0.00184004	police	1504	0.00193979	like	408	0.00200121
19	police	892	0.00184789	get	227	0.00176987	get	1476	0.00190368	government	404	0.00198159

Questions

	Task_1_train	Task_1_dev	Task_2_train	Task_2_dev
Unique Token Count	10032	4602	11098	4966
Total Count	43683	11289	47297	12255

	Task_1_train			Task_1_dev			Task_2_train			Task_2_dev
	Word	Count	Frequency	Word	Count	Frequency	Word	Count	Frequency	Word	Count	Frequency
0	placeholder	3227	0.0738731	placeholder	837	0.074143	placeholder	3318	0.0701524	placeholder	851	0.069441
1	said	268	0.00613511	new	69	0.00611214	new	229	0.00484174	new	71	0.00579355
2	year	230	0.00526521	says	62	0.00549207	year	218	0.00460917	two	60	0.00489596
3	new	222	0.00508207	year	56	0.00496058	two	209	0.00441888	says	57	0.00465116
4	says	197	0.00450976	us	48	0.00425193	one	201	0.00424974	year	54	0.00440636
5	two	172	0.00393746	said	46	0.00407476	said	200	0.0042286	said	54	0.00440636
6	first	159	0.00363986	first	43	0.00380902	man	191	0.00403831	us	49	0.00399837
7	world	158	0.00361697	two	43	0.00380902	police	185	0.00391145	people	48	0.00391677
8	one	148	0.00338805	world	38	0.00336611	people	177	0.00374231	one	47	0.00383517
9	man	143	0.00327358	uk	35	0.00310036	first	174	0.00367888	man	46	0.00375357
10	people	141	0.0032278	one	34	0.00301178	world	168	0.00355202	first	44	0.00359037
11	us	129	0.00295309	man	32	0.00283462	years	154	0.00325602	years	37	0.00301918
12	police	127	0.00290731	people	32	0.00283462	us	147	0.00310802	police	36	0.00293758
13	uk	127	0.00290731	police	31	0.00274604	uk	133	0.00281202	uk	32	0.00261118
14	wales	118	0.00270128	former	31	0.00274604	says	130	0.00274859	world	31	0.00252958
15	years	115	0.0026326	england	31	0.00274604	three	113	0.00238916	could	31	0.00252958
16	former	112	0.00256393	wales	29	0.00256887	england	111	0.00234687	city	28	0.00228478
17	government	110	0.00251814	league	29	0.00256887	old	105	0.00222001	found	27	0.00220318
18	could	109	0.00249525	years	28	0.00248029	could	102	0.00215658	may	27	0.00220318
19	city	106	0.00242657	three	25	0.00221455	last	100	0.0021143	bbc	27	0.00220318

Note:

• BERT Tokenizer based splits
  • Process - Take passages, convert to lowercase, remove all non-alphanumeric characters, tokenize, then remove stop words.

Articles

	Task_1_train	Task_1_dev	Task_2_train	Task_2_dev
Unique Token Count	21952	15858	23842	18157
Total Count	559481	148611	898160	235177

	Task_1_train			Task_1_dev			Task_2_train			Task_2_dev
	Word	Count	Frequency	Word	Count	Frequency	Word	Count	Frequency	Word	Count	Frequency
0	##s	8094	0.014467	##s	2393	0.0161024	##s	13511	0.015043	##s	3477	0.0147846
1	said	7994	0.0142882	said	1941	0.0130609	said	9597	0.0106852	said	2525	0.0107366
2	mr	2466	0.00440766	mr	702	0.00472374	##t	4282	0.00476752	##t	1068	0.00454126
3	##t	2266	0.00405018	##t	674	0.00453533	mr	3550	0.00395253	mr	993	0.00422235
4	would	2204	0.00393937	would	595	0.00400374	would	3549	0.00395141	would	964	0.00409904
5	also	2020	0.00361049	one	513	0.00345197	one	3348	0.00372762	people	873	0.0037121
6	people	1897	0.00339064	also	506	0.00340486	people	3335	0.00371315	one	855	0.00363556
7	one	1822	0.00325659	people	481	0.00323664	also	2786	0.0031019	also	675	0.00287018
8	two	1479	0.00264352	first	408	0.00274542	two	2277	0.00253518	two	611	0.00259804
9	last	1477	0.00263995	last	399	0.00268486	says	2201	0.00245057	new	557	0.00236843
10	first	1417	0.0025327	new	380	0.00255701	new	2129	0.0023704	says	547	0.00232591
11	new	1372	0.00245227	two	379	0.00255028	years	2084	0.0023203	us	546	0.00232166
12	year	1281	0.00228962	told	349	0.00234841	time	2076	0.00231139	time	540	0.00229614
13	years	1248	0.00223064	us	345	0.0023215	first	2056	0.00228912	first	525	0.00223236
14	could	1224	0.00218774	years	340	0.00228785	could	2007	0.00223457	could	518	0.0022026
15	##m	1219	0.0021788	##e	339	0.00228112	last	1975	0.00219894	years	516	0.00219409
16	us	1206	0.00215557	year	330	0.00222056	us	1844	0.00205309	told	493	0.00209629
17	told	1203	0.00215021	bbc	328	0.0022071	year	1793	0.0019963	bbc	475	0.00201976
18	time	1173	0.00209659	time	311	0.00209271	##ing	1785	0.0019874	##e	472	0.002007
19	bbc	1112	0.00198756	could	291	0.00195813	##e	1761	0.00196068	year	468	0.00198999

Questions

	Task_1_train	Task_1_dev	Task_2_train	Task_2_dev
Unique Token Count	9771	4876	10702	5251
Total Count	51629	13284	55858	14331

	Task_1_train			Task_1_dev			Task_2_train			Task_2_dev
	Word	Count	Frequency	Word	Count	Frequency	Word	Count	Frequency	Word	Count	Frequency
0	place	3256	0.0630653	place	844	0.0635351	place	3349	0.0599556	place	859	0.05994
1	##holder	3227	0.0625036	##holder	837	0.0630081	##holder	3318	0.0594006	##holder	851	0.0593818
2	said	272	0.00526836	new	69	0.00519422	##s	333	0.00596155	new	72	0.00502407
3	year	230	0.00445486	##s	66	0.00496838	new	230	0.00411758	##s	72	0.00502407
4	##s	223	0.00431928	says	62	0.00466727	year	218	0.00390275	two	60	0.00418673
5	new	222	0.00429991	year	56	0.0042156	two	209	0.00374163	says	57	0.00397739
6	says	197	0.00381568	us	48	0.00361337	one	201	0.00359841	year	54	0.00376806
7	two	172	0.00333146	said	46	0.00346281	said	200	0.00358051	said	54	0.00376806
8	first	160	0.00309903	first	43	0.00323698	man	196	0.0035089	man	50	0.00348894
9	world	158	0.0030603	two	43	0.00323698	police	185	0.00331197	us	49	0.00341916
10	man	150	0.00290534	world	38	0.00286058	people	177	0.00316875	people	48	0.00334938
11	one	148	0.00286661	uk	36	0.00271003	first	174	0.00311504	one	47	0.0032796
12	people	141	0.00273102	man	35	0.00263475	world	168	0.00300763	first	44	0.00307027
13	uk	135	0.00261481	one	34	0.00255947	years	154	0.00275699	years	37	0.00258182
14	us	130	0.00251796	people	32	0.00240891	us	147	0.00263167	police	36	0.00251204
15	police	129	0.0024986	police	31	0.00233363	uk	142	0.00254216	uk	33	0.0023027
16	wales	118	0.00228554	former	31	0.00233363	says	130	0.00232733	world	32	0.00223292
17	years	115	0.00222743	england	31	0.00233363	three	113	0.00202299	could	31	0.00216314
18	former	112	0.00216932	wales	29	0.00218308	england	111	0.00198718	city	28	0.00195381
19	government	110	0.00213059	league	29	0.00218308	old	105	0.00187977	found	27	0.00188403

References

Imperceptibility

• Parameters of Abstraction, Meaningfulness, and Pronunciability for 329 Nouns
• Gated-Attention Readers for Text Comprehension
• Concreteness Ratings for 40 Thousand Generally Known English Word Lemmas

Non-Specificity

• WordNet Interface (Hypernyms)