An AI using Construction Grammar: Automatic Acquisition of

Knowledge about Words

Denis Kiselev

HIKIMA.NET, Sapporo, Japan

Keywords:

Natural Language Processing, Knowledge Representation and Reasoning, Cognitive Systems, Explainable

AI, Construction Grammar, Winograd Schema.

Abstract:

This paper deals with an AI implementation that uses knowledge in an original Construction Grammar (CG)

format for deep understanding of text. CG is a means of processing knowledge pieces—aka constructions—

that describe the form and meaning of text parts. Understanding consists in automatically ﬁnding in the text

the knowledge that constructions contain and in creating knowledge networks that reﬂect the text information

structure. Deeper understanding is achieved by propagating knowledge within the network, i.e. some con-

structions can share with others information about syntax, semantics, pragmatics and other text properties.

A shortcoming of this information-rich method is limited coverage: only text for which a CG database is

available can be understood; that database due to its complexity most often needs to be made up manually.

The author attempts to increase the coverage by implementing automatic acquisition of word knowledge from

sources such as an external (non-CG) knowledge base and formatting the knowledge as CG constructions. The

resulting CG database has been used in evaluation experiments to understand the Winograd Schema (WS)—a

test for AI. A 28% increase in accurate coverage and opportunities for further improvement are observed.

1 INTRODUCTION

This section introduces the issue the proposed knowl-

edge acquisition method tackles: the limited coverage

of text by deeper understanding, such as in the case of

the computational CG. In the context of contemporary

Natural Language Processing (NLP) the above issue

can be presented as follows.

An NLP history survey (Brock, 2018) shows

two major research areas: those featuring “recog-

nition” and “reasoning”. Among the former Brock

(2018) mentions “neural networks coupled to large

data stores”, e.g. deep learning. For such

approaches—aka statistical NLP—the major criterion

for “making judgement” is the count of character

combinations (words etc.) recognized in a large data

store (e.g. in a collection of user input). The latter

(reasoning) approaches—aka Natural Language Un-

derstanding (NLU)—aim for deep understanding of

human cognition (Micelli et al., 2009) and attempt

to infer by processing knowledge such as syntax, se-

mantics and common sense. Computational CG is

among such NLU approaches; Kiselev (2017) intro-

duces CG as a formalism and describes its practical

applications.

Both deep NLU and statistical NLP have strengths

accompanied by weaknesses as shown below.

Modern statistical NLP methods are usually

“trained” on large data (e.g., millions of word oc-

currences) so they can process (i.e. their knowledge

database is enough to cover) considerably large in-

puts (Zhang et al., 2018). One problem with such

methods is limited accuracy when it comes to dif-

ﬁcult understanding tasks involving deep semantics

or common sense reasoning (Mitkov, 2014; Richard-

Bollans et al., 2018). To illustrate this problem: a

number of statistical NLP methods (Peng et al., 2015;

Sharma et al., 2015; Liu et al., 2017; Emami et al.,

2018) can process considerably large quantities of

WS sentences, however with limited accuracy; Bai-

ley et al. (2015) describe WS as a test for AI. Another

problem with statistical NLP, e.g. Machine Learning

(ML), is limited transparency: it is often impossible

or impractical to ﬁnd out what exact “thoughts” led

an AI of this kind to a certain decision (Brock, 2018).

To ﬁnd that out one would need to look at millions

of nodes (words, etc) and their numeric weights (re-

sulting from mathematical manipulations of word oc-

currence counts, etc) that activated certain networks.

This limited explainability is a reason to call statistical

Kiselev, D.

An AI using Construction Grammar: Automatic Acquisition of Knowledge about Words.

DOI: 10.5220/0008865902890296

In Proceedings of the 12th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2020) - Volume 2, pages 289-296

ISBN: 978-989-758-395-7; ISSN: 2184-433X

289

NLP AI with neural networks “a black box” (Brock,

2018).

Differently from statistical NLP, deep NLU ap-

proaches, e.g. CG ones (Kiselev, 2017; Raghuram

et al., 2017), show high accuracy on the WS but can

process (i.e. their knowledge database is enough to

cover) limited numbers of WS sentences. This lim-

ited coverage results from more difﬁculty—than in

the case of statistical NLP—to automatically create a

database because the database has to reﬂect a richer

information structure of text. Deep NLU requires

more human effort to (manually) create the knowl-

edge base. As for the explainability, the CG-based AI

“thoughts” are usually clear from the data structure

and/or reasoning rules used, as for instance in the case

of the Kiselev (2017) and the AI method the present

paper proposes. Implementations of this type are re-

ferred to as explainable AI (aka XAI).

To sum it up, statistical NLP (such as ML) demon-

strates large coverage but limited accuracy and lim-

ited explainability, deep NLU (such as computational

CG) demonstrates limited coverage but high accuracy

and sufﬁcient explainability.

2 CURRENT RESEARCH STAGE

AND TARGETS

This section identiﬁes the AI implementation the pro-

posed knowledge acquisition method has built upon,

and outlines future knowledge acquisition research

prospects. The section ends with the question to be

empirically answered by the present paper.

The author tackles the NLU coverage issue, ex-

plained in 1, by further improving an existing AI im-

plementation. That AI (Kiselev, 2017) has shown

high accuracy in answering questions from a WS

dataset, but its CG database is limited to that dataset

only and no automatic knowledge acquisition is im-

plemented. That database is manually populated with

information structures called constructions, each of

which contains knowledge about the form and mean-

ing of a word, a phrase or a sentence. To put it differ-

ently, the database has three types of constructions:

word, phrase and sentence ones. It is important to

note that the same constructions can be reused to un-

derstand text pieces with similar form and meaning.

This paper describes the current research stage

at which the author looks into automatic acquisition

of word constructions only. They are added to the

database of the above AI and used for WS understand-

ing. Automatic acquisition of phrase and sentence

constructions are future stages of the research.

The paper is to answer the following question.

Does automatically adding word constructions to the

existing database result in any increase of the accu-

rate coverage for unknown sentences? This ques-

tion is important because the word constructions are

meant to become components of phrase and sentence

constructions by satisfying syntactic, semantic and

pragmatic requirements (listed “inside” phrase and

sentence constructions). The answer to the ques-

tion contributes to the practical value of the proposed

method applied to understanding the WS, an espe-

cially hard (Winograd, 1980) task for AI.

3 TEXT UNDERSTANDING

Before explaining how the knowledge is acquired, it

is important to explain how the present AI implemen-

tation understands text.

3.1 Knowledge Networks

First, word constructions from the database are used

to ﬁnd known words. A construction is, simply put, a

collection of feature => value pairs. This is a tem-

plate so instead of the feature and value any mean-

ingful information can be used. That information is to

describe form and/or meaning of text pieces of three

types (recall the construction types mentioned in 2).

For instance, a familiar word “phone” was found be-

cause the value (i.e. the word spelling) in the pair

string => phone matched “phone” in the text. The

boldface words within boxes in Fig. 1 represent con-

structions for the known words: the implementation

now knows e.g. that “he” is a personal pronoun. It

knows because the construction that matched “he” has

the pair type => personal and the construction also

says that “he” is a pronoun.

Next, the AI checks if the found known words

(or, technically, the word constructions that have

matched the spelling) ﬁt into known phrase structures.

Each phrase construction in the database represents a

known phrase type, e.g. the verb phrase (VP) for “call

George” (Fig. 1). These two words ﬁt into the VP be-

cause the AI has found matching features and values.

For instance, the values action in the word construc-

tion for “call” and entity in that for “George” have

matched the same values in the VP construction; for

space considerations the ﬁgure does not show every

feature and value.

As all form and meaning requirements of the VP

to its potential components were similarly met by the

two word constructions (i.e. certain features and val-

ues were found matching), the constructions became

ICAART 2020 - 12th International Conference on Agents and Artiﬁcial Intelligence

290

Figure 1: The knowledge network generated to ﬁnd what “he” refers to in the WS sentence “Paul tried to call George on the

phone but he wasn’t available”; a concise illustration. Solid line boxes represent CG constructions.

components

of the VP. This VP construction has

been (manually) designed to know its semantic struc-

ture patient => noun (Fig. 1): the noun denotes the

patient, i.e. the receiver of the action. This semantic

knowledge has been propagated from the VP to the

“George” word construction, i.e. the AI has copied

knowledge from the above feature-value pair. In this

way the implementation has understood phrases and

now knows that e.g. “call George” is a VP, “call” is

an action and “George” is an entity as well as the

patient (recipient) for “call”. In other words, knowl-

edge networks for phrases have been formed. In Fig. 1

the network components are connected by lines with

circles on ends, dashed arrows show how knowledge

propagated (i.e.was copied by the AI).

Sentences are understood the same way as

phrases. Constructions for the understood phrases ﬁt

into the known sentence structure if features and val-

ues match. Knowledge listed as feature-value pairs

is copied from some constructions to others. Techni-

cal details of what conditions feature-value pairs must

meet so they match or so they are copied the way

the construction designer desires are given by Kiselev

(2017) . However, for the present AI some feature-

value pairs have been discarded to make phrase and

sentence constructions lighter, and regular expres-

Becoming components means getting marked by the AI

as components and put in a specially allocated memory

container that is traceable, retrievable and printable.

sions (regex)

are incorporated into constructions for

more matching versatility. For instance, the same

construction can be designed to match multiple word

forms or parts of speech (POS).

Another feature of the present AI is recursive gen-

eration of alternative parses. Some cases for doing so

are as follows. In English there is a number of words

that, depending on the context, can be different POS,

although the spelling is the same. If such word is used

in a sentence and matches constructions for different

POS, alternative sentence parses (each containing the

construction for the different POS) are generated. If

the same group of words in a sentence matches multi-

ple constructions for different types of phrases, alter-

native sentence parses are also generated. At present

the AI is set to output parses more extensively cov-

ered by the CG database knowledge, or simply parses

with the larger number of matching constructions.

It is not true that the tree-like structure represent-

ing the knowledge network for the whole sentence in

Fig. 1 has to be a syntactic tree generated by means of

statistical NLP—athough, that is also implementable.

In other words, what types of phrases the sentence is

to consist of does not have to be determined by how

many phrases of some type are usually found in large

collections of data. By creating feature-value pairs the

construction designer is free to have any construction

Programming tools used for manipulating character

strings, e.g. ﬁnding words or changing morphemes.

An AI using Construction Grammar: Automatic Acquisition of Knowledge about Words

291

convey practically any meaning, match another con-

struction and form a network that can have practically

any syntactic structure. This kind of freedom is re-

ferred to by van Trijp (2017) as “chopping down the

syntax tree”.

By forming the sentence network exempliﬁed in

Fig. 1 the AI has understood, for instance, the fol-

lowing. The sentence (Fig. 1) has a clause, this

clause has a VP, and this VP implies negated pres-

ence. This implication has been understood because

the pairs polarity => negative and implies

=> presence from the “wasn’t” and “available”

constructions respectively were copied to the VP

construction—copied because they met certain condi-

tions as mentioned above, and met the conditions be-

cause the constructions had been manually designed

to do so. In terms of human reasoning, this knowledge

propagation can be described like so. As “wasn’t”

is negative and “available” implies presence and the

two words form a VP (the way shown in Fig. 1), that

VP implies negated presence. This is one example of

the way CG knowledge propagation can in itself be

looked upon as a chain of reasoning. The knowledge

propagation implemented by the author draws upon

the CG remarkable quality of feature and value prop-

agation referred to by Steels (2017) as “percolation”

and “merging”.

The following CG concept should also be noted.

It is not true that each single construction can de-

scribe only one unique piece of text. Constructions

are meant to be reusable for (i.e. can describe) text

pieces with similar forms and meanings. For instance

a single VP construction can be “ﬁlled with” a va-

riety of actual verbs and nouns from the text. The

same concept applies to sentence constructions. This

is what makes CG different from annotated corpora

where a particular tag is attached to one word or one

group of them.

3.2 Inference Rules

Rules are instructions for the present AI implementa-

tion to follow when more complex reasoning—such

as the anaphora resolution task exempliﬁed by the

caption for Fig. 1—is involved. A rule deﬁnes a rea-

soning pattern as explained below. Kiselev (2017) has

shown that WS anaphora resolution can be handled

without using rules. However, that results in construc-

tion data proliferation: the need to use a larger number

of long constructions each of which describes a sepa-

rate sentence as a whole. Using rules for the present

AI aims for shorter and more versatile constructions.

A rule has condition and action parts: the action

is taken if the condition is met. Below is a human

language gloss of a rule for anaphora resolution in

the Fig. 1 sentence and in similar sentences; the con-

struction data dealt with can be looked up by locating

dashed boxes and arrows. In a compound sentence

a clause of which implies negated presence, into a

part that is a nominative case personal pronoun, in-

sert refers to => pointing to a part that denotes

the patient (the receiver of the action). To gloss it fur-

ther: the sentence speaks about not being present, so

“he” refers to “George”. This is just one possible res-

olution of anaphora, a construction designer is free to

come up with others.

4 KNOWLEDGE ACQUISITION

The knowledge (i.e. word constructions) acquisition

targets unknown words. A word is unknown if to

match its spelling (recall 3.1) the AI has tried every

word construction available from the database, but all

unsuccessfully. In this case it does not even know if

some character string between spaces can be a word.

So the acquisition task can be divided into two steps.

First, identifying words as such. Next, making word

constructions for them.

4.1 Identifying Words

For lack of CG tools covering the morphology and

syntax of large textual data, the author has used

the external tool TreeTagger

to identify words and

tag them as POS. TreeTagger is a probabilistic

(Hidden Markov Model) POS tagger using decision

trees (Schmid, 1994). The decision algorithm fea-

tures analyzing morphemes of the word in ques-

tion and available POS tags for words preceding it,

which demonstrates rather high precision with a lim-

ited amount of training data (Schmid, 1999). Tree-

Tagger has been incorporated into the present AI and

has been conﬁgured to output in the Penn Treebank

format which attempts (Taylor et al., 2003) to utilize

fewer, more universal tags and observes a standard

list (Santorini, 1990) of thirty-six POS tags displayed

in Table 3 (see APPENDIX). For instance “call” (as

used in the Fig. 1 sentence) according to this list is

tagged VB i.e. a verb, base form.

4.2 Making Word Constructions

After an unknown word (or, technically, a character

string) is identiﬁed as a POS, a construction for it is

Available at http://www.cis.uni-muenchen.de/schmid/

tools/TreeTagger/ retrieved on May 3, 2019.

ICAART 2020 - 12th International Conference on Agents and Artiﬁcial Intelligence

292

CONSTRUCTION => call_VB

___Text_Form

ID => VB

lemma => call

string => calls|called

to_match => ID

___Text_Meaning

ID => A

to_match => ID

wn_senses => 28

Figure 2: A word construction example: a minimal con-

struction for the verb “call”.

generated. By tradition (De Saussure, 2011) a con-

struction describes a linguistic sign of a certain form

(e.g., a string of characters that are a word) and the

meaning corresponding to that form. Constructions

automatically generated by the present AI have two

major parts Text Form and Text Meaning (Fig. 2).

To identify the form and meaning of a word (so the

identiﬁer can match one in other constructions to form

networks), the two parts have pairs such as ID =>

VB (“verb, base form”) and ID => A (“action”), see

Fig. 2.

The notation in Fig. 2 is interpreted as follows:

the value call VB means this is a construction for the

word “call” and its POS tag is VB; the ID values VB

and A result from the TreeTagger output the way de-

scribed later; the lemma pair is also made out of the

TreeTagger output and shows the basic form of the

word; calls|called (given as mere examples) en-

able the construction to match the former or the lat-

ter verb form in a text, if needed; to match => ID

is a “dummy” pair used to tell the AI to look for

matches only in the ID pair of this word construc-

tion and of phrase or sentence constructions (when

matches in other form or meaning features are not re-

quired for forming a network); the wn senses pair

points to word senses as described later.

As mentioned above, the identiﬁers VB and A

(Fig. 2) result—i.e. are automatically made—from

the TreeTagger output. The word “call” as it is used

in the sentence (Fig. 1) is parsed by TreeTagger as

VB. The meaning identiﬁer A (Fig. 2) corresponds to

the form identiﬁer VB (as a verb denotes an action).

Meaning tags that correspond to other POS are also

generated. The form and meaning identiﬁers are used

so they can match the same ones in, for instance, a

phrase construction that may say that its prospective

parts must match those identiﬁers; in this way the

AI understands what words are parts of what phrases

etc., recall 3.1. Table 3 (see APPENDIX) lists and

describes form and meaning identiﬁers automatically

inserted into word constructions.

The wn senses => 28 pair (Fig. 2) says that there

are data for twenty-eight senses of the verb “call” in

WordNet (Miller, 1998), a publicly available lexical

database with look-up tools. WordNet has been incor-

porated into the present AI and the number of senses

is automatically inserted into the word construction.

For each of the senses WordNet lists synonyms and

other related words; the total number of senses like

28 facilitates faster (parallel) processing of the data

for each sense. The author uses these WordNet sense

data so the same inference rules (such as that exem-

pliﬁed in 3.2) can be applied to text parts with similar

meanings, however in-depth research into that kind of

rule application is left for the future.

At present minimal constructions like one shown

in Fig. 2 are automatically generated, additional

feature-value need to be entered manually.

5 EVALUATION EXPERIMENTS

5.1 Setting

The purpose of the experiments is to answer the ques-

tion (posed in 2) about the coverage increase. To fur-

ther detail the question: is there any increase in the

number of accurately formed sentence networks due

to the described automatic generation of word con-

structions? To gloss the question for more clarity:

how many sentence constructions (i.e. their ID val-

ues, etc.) are correctly matched by word and phrase

ones; do the formed sentence networks correctly re-

ﬂect the syntax and semantics? As the phrase and sen-

tence constructions come from the limited database

(introduced in 2) it is interesting how large a portion

of the whole WS collection the database knowledge

can cover—in the experiments the database is used to

understand the whole WS collection, differently from

Kiselev (2017) who applied that database to a part of

the WS collection.

The whole WS collection of 150 items or 300

statement-question sets (as of the time the experi-

ments were preformed) was utilized. That collec-

tion was obtained from a publicly available online

source

. For more clarity the original collection for-

mat that uses slashes and square brackets was changed

(by writing and executing an additional program) into

that shown in Fig. 3, the item numbering corresponds

to the original collection numbering.

The CG database mentioned above was originally

designed to cover 7 items or 14 statement-question

sets (Kiselev, 2017) from the WS collection. In the

http://www.cs.nyu.edu/faculty/davise/papers/

WinogradSchemas/WSCollection.html retrieved on

June 20, 2019.

An AI using Construction Grammar: Automatic Acquisition of Knowledge about Words

293

- 52 -

The fish ate the worm. It was tasty.

What was tasty?

The fish ate the worm. It was hungry.

What was hungry?

Figure 3: A WS as it is used in the experiments. There are

two statement-question sets: the statement part is directly

followed by the question. A WS is preceded by its number.

Table 1: Construction quantities in the database before the

automatic generation of word constructions.

Word Constructions 79

Phrase Constructions 19

Sentence Constructions 13

experiments the database is applied to all of the 150

items or 300 sets; recall that the same constructions

can be reused to understand text pieces with similar

form and meaning. Table 1 lists the quantities for con-

structions of each type in the database. Before the ex-

periments the database was formatted so the existing

constructions could form networks with the new auto-

matically generated ones. For instance, the new form

and meaning notation for ID pairs (described in 4.2)

was incorporated into constructions.

First, the AI applied the database without auto-

matically generating word constructions. Next, it ap-

plied the same database again and while doing so au-

tomatically added constructions for unknown words.

5.2 Results

In Table 2 the row TTL gives the total number of

statement-question sets in which at least one sen-

tence network was formed, i.e. at least one sentence

construction matched phrase and/or sentence ones;

the row Accurate gives the part of the above TTL

where the sentence network correctly—from the hu-

man prospective—reﬂects the syntax and semantics;

the row Accurate Coverage % shows the share of the

above Accurate count in the number of all the WS

collection statement-question sets (i.e. in 300).

By comparing the entries in the row Accurate

Coverage % a 28% increase can be observed.

5.3 Discussion

The above results are not meant to extensively cover

the whole WS collection, they are rather a step to-

wards that—the following steps being phrase and sen-

tence knowledge acquisition as mentioned in 2.

An increase in accurate coverage of sentences

without automatically generating phrase and sen-

tence constructions can be considered encouraging

Table 2: Numbers and percentages for sentence networks

formed with and without the automatic generation of word

constructions: Proposed and Baseline respectively.

Baseline Proposed

TTL 17 170

Accurate 17 101

Accurate Coverage % 6% 34%

and speaks in favor of CG scalability to a certain ex-

tent: phrase and sentence constructions have proven

reusable on the WS.

Along with an encouragement there is room for

improvement. Namely, by comparing the ﬁgures for

TTL and Accurate under Proposed in Table 2 it is

clear that 69 sentence networks were not parsed cor-

rectly. The reason is that some constructions may

have been designed excessively versatile, i.e. allow-

ing matches of unwanted POS, word forms etc. There

is a need for a better look into using a larger number

of non-versatile constructions as opposed to a smaller

number of versatile ones.

The recursive generation of alternative parses

is rather time-consuming (about 25 seconds for a

statement-question set with an especially large num-

ber of parses) so the implementation of parallel pro-

cessing for variant parses may be needed.

6 CONCLUSIONS AND FUTURE

WORK

This paper has described a step towards automating

the hard task of understanding the WS by means of

deep NLU, featuring an original explainable AI that

utilizes CG and automatically generates word knowl-

edge. That knowledge generation augmented phrase

and sentence knowledge already existing in a database

and led to accurate understanding of the sentence

structure in a larger portion of the WS text. How-

ever, a more complete unsupervised understanding

necessitates automatic generation of not only word

knowledge but also knowledge about the phrase and

sentence form and meaning. Research into this kind

of knowledge generation is on the long-term agenda.

The short-term agenda includes research into the ap-

propriate degree of the construction versatility and

into parallel processing for alternative parses of text.

REFERENCES

Bailey, D., Harrison, A. J., Lierler, Y., Lifschitz, V., and

Michael, J. (2015). The winograd schema challenge

ICAART 2020 - 12th International Conference on Agents and Artiﬁcial Intelligence

294

and reasoning about correlation. In 2015 AAAI Spring

Symposium Series.

Brock, D. C. (2018). Learning from artiﬁcial intelligence’s

previous awakenings: The history of expert systems.

AI Magazine, 39(3):3–15.

De Saussure, F. (2011). Course in general linguistics.

Columbia University Press.

Emami, A., Trischler, A., Suleman, K., and Cheung, J. C. K.

(2018). A generalized knowledge hunting framework

for the winograd schema challenge. In Proceedings

of the 2018 Conference of the North American Chap-

ter of the Association for Computational Linguistics:

Student Research Workshop, pages 25–31.

Kiselev, D. (2017). A construction-grammar approach to

computationally solving the winograd schema: Imple-

mentation and evaluation. In 2017 AAAI Spring Sym-

posium Series, Stanford University, Palo Alto, USA.

Liu, Q., Jiang, H., Evdokimov, A., Ling, Z.-H., Zhu, X.,

Wei, S., and Hu, Y. (2017). Cause-effect knowledge

acquisition and neural association model for solving

a set of winograd schema problems. In The Twenty-

Sixth International Joint Conference on Artiﬁcial In-

telligence (IJCAI-17).

Micelli, V., van Trijp, R., and De Beule, J. (2009). Framing

ﬂuid construction grammar. In the 31th annual con-

ference of the cognitive science society, pages 3023–

3027.

Miller, G. (1998). WordNet: An electronic lexical database.

MIT press.

Mitkov, R. (2014). Anaphora resolution. Routledge.

Peng, H., Khashabi, D., and Roth, D. (2015). Solving hard

coreference problems. In Proceedings of the 2015

Conference of the North American Chapter of the As-

sociation for Computational Linguistics: Human Lan-

guage Technologies, pages 809–819.

Raghuram, V., Trott, S., Shen, K., Goldberg, E., and Oder-

berg, S. (2017). Semantically-driven coreference res-

olution with embodied construction grammar. In 2017

AAAI Spring Symposium Series, Stanford University,

Palo Alto, USA.

Richard-Bollans, A., Gomez Alvarez, L., and Cohn, A. G.

(2018). The role of pragmatics in solving the wino-

grad schema challenge. In Proceedings of the Thir-

teenth International Symposium on Commonsense

Reasoning (Commonsense 2017). CEUR Workshop

Proceedings.

Santorini, B. (1990). Part-of-speech tagging guidelines for

the Penn Treebank Project. University of Pennsylva-

nia, School of Engineering and Applied Science.

Schmid, H. (1994). Probabilistic part-of-speech tagging us-

ing decision trees, intl. In Conference on New Meth-

ods in Language Processing. Manchester, UK.

Schmid, H. (1999). Improvements in part-of-speech tag-

ging with an application to german. In Natural lan-

guage processing using very large corpora, pages 13–

25. Springer.

Sharma, A., Vo, N. H., Gaur, S., and Baral, C. (2015). An

approach to solve winograd schema challenge using

automatically extracted commonsense knowledge. In

2015 AAAI Spring Symposium Series.

Steels, L. (2017). Basics of ﬂuid construction grammar.

Constructions and frames, 9(2):178–225.

Taylor, A., Marcus, M., and Santorini, B. (2003). The penn

treebank: an overview. In Treebanks, pages 5–22.

Springer.

van Trijp, R. (2017). A computational construction gram-

mar for english. In 2017 AAAI Spring Symposium Se-

ries, Stanford University, Palo Alto, USA.

Winograd, T. (1980). What does it mean to understand lan-

guage? Cognitive science, 4(3):209–241.

Zhang, L., Wang, S., and Liu, B. (2018). Deep learning

for sentiment analysis: A survey. Wiley Interdisci-

plinary Reviews: Data Mining and Knowledge Dis-

covery, 8(4):e1253.

An AI using Construction Grammar: Automatic Acquisition of Knowledge about Words

295

APPENDIX

Table 3: The columns No., Form ID and POS Explanation respectively give the item number, the part-of-speech (POS)

identiﬁer (ID) tag and the explanation as they are in the Penn Treebank Project (Santorini, 1990). The column Meaning ID

lists IDs showing a meaning for each POS (e.g., the meaning “action” for a verb). The Meaning IDs: A = action, E = entity,

P = property, U = currently unspeciﬁed. This simplistic meaning identiﬁcation meets the current research needs but may be

reconsidered in the future.

No. Form ID Meaning ID POS Explanation

1. CC U Coordinating conjunction

2. CD U Cardinal number

3. DT U Determiner

4. EX U Existential there

5. FW U Foreign word

6. IN U Preposition or subordinating conjunction

7. JJ P Adjective

8. JJR P Adjective, comparative

9. JJS P Adjective, superlative

10. LS U List item marker

11. MD U Modal

12. NN E Noun, singular or mass

13. NNS E Noun, plural

14. NP E Proper noun, singular

15. NPS E Proper noun, plural

16. PDT U Predeterminer

17. POS U Possessive ending

18. PP U Personal pronoun

19. PP$ U Possessive pronoun

20. RB P Adverb

21. RBR P Adverb, comparative

22. RBS P Adverb, superlative

23. RP U Particle

24. SYM U Symbol

25. TO U to

26. UH U Interjection

27. VB A Verb, base form

28. VBD A Verb, past tense

29. VBG U Verb, gerund or present participle

30. VBN U Verb, past participle

31. VBP A Verb, non-3rd person singular present

32. VBZ A Verb, 3rd person singular present

33. WDT U Wh-determiner

34. WP U Wh-pronoun

35. WP$ U Possessive wh-pronoun

36. WRB U Wh-adverb

ICAART 2020 - 12th International Conference on Agents and Artiﬁcial Intelligence

296