Building tree annotation with a spreadsheet

Alastair Butler

Hirosaki University

A spreadsheet consists of a table of cells arranged into columns and rows. The columns are normally represented by letters, “A”, “B”, “C”, etc., while the rows are normally represented by numbers, “1”, “2”, “3”, etc. A single cell can be referenced by its column and row (A5 would represent the cell containing the value jumps in table (2) below). Additionally, spreadsheets have the concept of a range, a group of cells, normally contiguous. For example, the first ten cells in the first column of (2) form the range “A1:A10”, which is filled with the words of the example sentence (1). Punctuation is treated like a word, where each consecutive word fills a single cell of a distinct consecutive row. Cell B11 contains an ID for the sentence, with there needing to be at least one completely blank row (row 12 in (2)) following the row of an ID cell.

(1): The quick brown fox jumps over the lazy dog.

(2)

	A	B
1	The
2	quick
3	brown
4	fox
5	jumps
6	over
7	the
8	lazy
9	dog
10	.
11	ID	example
12

We need to add word class tag information to the words of (2). We first insert cells to the left of A1:A10, so that the contents of A1:A10 become B1:B10, and A1:A10 become empty cells, as in (3). There is now room for adding word class tags.

(3)

	A	B
1		The
2		quick
3		brown
4		fox
5		jumps
6		over
7		the
8		lazy
9		dog
10		.
11	ID	example
12

We add word class tags to the cells of A1:A10, resulting in (4).

(4)

	A	B
1	D	The
2	ADJ	quick
3	ADJ	brown
4	N	fox
5	VBP	jumps
6	P-ROLE	over
7	D	the
8	ADJ	lazy
9	N	dog
10	PUNC	.
11	ID	example
12

Now that every word has a word class tag, we can start projecting phrase structure. In particular, let's start by projecting ADJP structures. This requires that we first insert cells to the left of the ADJ word tags, resulting in (5).

(5)

	A	B	C
1	D	The
2		ADJ	quick
3		ADJ	brown
4	N	fox
5	VBP	jumps
6	P-ROLE	over
7	D	the
8		ADJ	lazy
9	N	dog
10	PUNC	.
11	ID	example
12

We can now fill the empty cells created in (5) with the ADJP tag, resulting in (6). Note that two instances of ADJP are given indexes “;@2” and “;@3“. This is because these ADJP instances occur consecutively and yet they need to be distinct ADJP projections. Aside from making consecutive projections distinct, the chosen numbering is not important.

(6)

	A	B	C
1	D	The
2	ADJP;@2	ADJ	quick
3	ADJP;@3	ADJ	brown
4	N	fox
5	VBP	jumps
6	P-ROLE	over
7	D	the
8	ADJP	ADJ	lazy
9	N	dog
10	PUNC	.
11	ID	example
12

We next insert cell structure for a subject noun phrase, resulting in (7).

(7)

	A	B	C	D
1		D	The
2		ADJP;@2	ADJ	quick
3		ADJP;@3	ADJ	brown
4		N	fox
5	VBP	jumps
6	P-ROLE	over
7	D	the
8	ADJP	ADJ	lazy
9	N	dog
10	PUNC	.
11	ID	example
12

We next fill the subject noun phrase cell structure created in (7) with the NP-SBJ tag, resulting in (8). There is no indexing with NP-SBJ because all consecutive instances of NP-SBJ (A1:A4) are the same NP-SBJ projection.

(8)

	A	B	C	D
1	NP-SBJ	D	The
2	NP-SBJ	ADJP;@2	ADJ	quick
3	NP-SBJ	ADJP;@3	ADJ	brown
4	NP-SBJ	N	fox
5	VBP	jumps
6	P-ROLE	over
7	D	the
8	ADJP	ADJ	lazy
9	N	dog
10	PUNC	.
11	ID	example
12

We next insert cell structure for a second noun phrase, resulting in (9).

(9)

	A	B	C	D
1	NP-SBJ	D	The
2	NP-SBJ	ADJP;@2	ADJ	quick
3	NP-SBJ	ADJP;@3	ADJ	brown
4	NP-SBJ	N	fox
5	VBP	jumps
6	P-ROLE	over
7		D	the
8		ADJP	ADJ	lazy
9		N	dog
10	PUNC	.
11	ID	example
12

We next fill the noun phrase cell structure created in (9) with the NP tag, resulting in (10).

(10)

	A	B	C	D
1	NP-SBJ	D	The
2	NP-SBJ	ADJP;@2	ADJ	quick
3	NP-SBJ	ADJP;@3	ADJ	brown
4	NP-SBJ	N	fox
5	VBP	jumps
6	P-ROLE	over
7	NP	D	the
8	NP	ADJP	ADJ	lazy
9	NP	N	dog
10	PUNC	.
11	ID	example
12

We next insert cell structure for a preposition phrase, resulting in (11).

(11)

	A	B	C	D	E
1	NP-SBJ	D	The
2	NP-SBJ	ADJP;@2	ADJ	quick
3	NP-SBJ	ADJP;@3	ADJ	brown
4	NP-SBJ	N	fox
5	VBP	jumps
6		P-ROLE	over
7		NP	D	the
8		NP	ADJP	ADJ	lazy
9		NP	N	dog
10	PUNC	.
11	ID	example
12

We next fill the preposition phrase cell structure created in (11) with the PP-DIR tag, resulting in (12).

(12)

	A	B	C	D	E
1	NP-SBJ	D	The
2	NP-SBJ	ADJP;@2	ADJ	quick
3	NP-SBJ	ADJP;@3	ADJ	brown
4	NP-SBJ	N	fox
5	VBP	jumps
6	PP-DIR	P-ROLE	over
7	PP-DIR	NP	D	the
8	PP-DIR	NP	ADJP	ADJ	lazy
9	PP-DIR	NP	N	dog
10	PUNC	.
11	ID	example
12

We next insert cell structure for a matrix clause, resulting in (13).

(13)

	A	B	C	D	E	F
1		NP-SBJ	D	The
2		NP-SBJ	ADJP;@2	ADJ	quick
3		NP-SBJ	ADJP;@3	ADJ	brown
4		NP-SBJ	N	fox
5		VBP	jumps
6		PP-DIR	P-ROLE	over
7		PP-DIR	NP	D	the
8		PP-DIR	NP	ADJP	ADJ	lazy
9		PP-DIR	NP	N	dog
10		PUNC	.
11	ID	example
12

We next fill the matrix clause cell structure created in (13) with the IP-MAT tag, resulting in the completed tree information of (14).

(14)

	A	B	C	D	E	F
1	IP-MAT	NP-SBJ	D	The
2	IP-MAT	NP-SBJ	ADJP;@2	ADJ	quick
3	IP-MAT	NP-SBJ	ADJP;@3	ADJ	brown
4	IP-MAT	NP-SBJ	N	fox
5	IP-MAT	VBP	jumps
6	IP-MAT	PP-DIR	P-ROLE	over
7	IP-MAT	PP-DIR	NP	D	the
8	IP-MAT	PP-DIR	NP	ADJP	ADJ	lazy
9	IP-MAT	PP-DIR	NP	N	dog
10	IP-MAT	PUNC	.
11	ID	example
12

The completed tree information of (14) can be represented in traditional bracketed (Penn/CorpusSearch) format as (15). Note that the bracketing keeps consecutive constituents with the same node tag distinct, so that the node indexing introduced in (6) is no longer required.

(15)

( (IP-MAT (NP-SBJ (D The)
                  (ADJP (ADJ quick))
                  (ADJP (ADJ brown))
                  (N fox))
          (VBP jumps)
          (PP-DIR (P-ROLE over)
                  (NP (D the)
                      (ADJP (ADJ lazy))
                      (N dog)))
          (PUNC .))
  (ID example))

The completed tree information of (14) can be represented as the graphical tree (16).

(16)