understand metaphor

We all use metaphors like "went off the rails" and "run aground" in our daily lives and understand them easily, but will computers also be able to understand them in the future?

Israel Benjamin Galileo Magazine

"The project is about to run aground."
"The leadership has gone off the rails."
"He lost the north."

What do all the sentences above have in common? First, they are all examples of the use of metaphor in everyday language: metaphors are not only a literary device that is only required of poets, novelists and literary scholars, but are an essential part of the way we speak and communicate.

Searching for metaphors is difficult because of the difference between them and images: the metaphor does not explicitly indicate that a comparison is being made, nor does it point to the concepts it compares

These sentences also have another common denominator, because the metaphors in them all refer to a similar semantic field, in that they use the link of the project, leadership or life to concepts related to the journey. The journey can be of a ship (which may run aground), of a train (which must stay on the tracks), etc., and in any case on the journey we must know in which direction we are moving (and therefore we need, at least metaphorically, a compass). This is a very common type of metaphor, and it is easy to recall additional examples ("The horse has run out" which is now giving way to "He has run out of gas", "The new decision is a U-turn", "The relationship between them is at the beginning of its journey" and more). The choice of the type of journey may dictate the continuation of the statement - for example: "The project is about to run aground, so we must turn the wheel".

Like the hero of Moliere's play "Also in the nobles", who spoke prose for forty years without knowing what it was, we all use metaphors of this type and understand them easily - in fact, we have no problem doing so, but we will encounter a serious problem if we are required to completely avoid using such linguistic means . The things are said, of course, not in deep literary metaphors, but rather in those worn-out everyday metaphors (some of which have even become "dead metaphors", i.e. those that have become a linguistic unit and lost the context of the original meaning - the phrase "prime minister", for example, is no longer used by most of its users and hear the associations to the meaning of the word "head" as an organ in the body).

go figure

In the story "Little Lost Robot", which appears among the other stories in the file "I, Robot") written by Isaac Asimov in 1947, one of the engineers tells the robot he is working with "Get lost!" (In the English original, "Go lose yourself!"). The robot understands the instruction verbatim and hides its identity among dozens of robots that are identical to it on the outside even though it differs from them in its programming, and the tension in the story is created around the efforts to discover that robot.

One of the critics of Asimov's works pointed out the illogicality of the described situation: can we imagine that in the joint work of humans and robots the robots will not often be required to understand such statements in the way that the person intended, instead of a literal understanding? For example, referring to the example we opened with, what would happen if the robot heard that "we have to turn the steering wheel" while driving a car, when the speaker was referring metaphorically to the need to change some decisions? The same critic, Joseph Patrouch, claimed that Asimov's inability to sense this problem was due to the fact that Asimov himself tended to think in a literal way, which was also expressed in his direct and "clean" writing style.

I prefer to see Asimov's story as a success in predicting one of the biggest challenges in computer understanding of human languages ​​- the difficulty of understanding statements whose meaning does not derive from their dictionary definition. You can of course enter the "dictionary" that the computer uses with a variety of phrases with their true meaning: only in very rare cases will the phrase "get lost" mean to order someone to disappear so that it will be impossible to find them.

Unfortunately, this is not a solution in most cases, as phrases like "turn the wheel" appear in both metaphorical and literal contexts. This is not the only problem: the number of expressions we will need in that "dictionary" is not limited - in our example, for example, it would also be possible to say that you need to press the brakes, turn a U-turn, steer carefully, etc. Therefore, the "dictionary of expressions" method is not enough, and there is no escaping the need to equip the computer with some ability to understand metaphors.
Mapping between buildings

Already in the early XNUMXs, artificial intelligence developers recognized the need for understanding metaphors. In accordance with the approach that was prevalent at that time (an approach that still has supporters today), the understanding of natural language was seen as a translation of the text into a logical-mathematical world. In this world, statements are expressed as links between well-defined symbols, so that one word may have many symbols attached to it, depending on its possible meanings, and vice versa: one symbol may correspond to many expressions ("George Washington" or "the first president of the USA" or "The man they sculpted is the leftmost one on Mount Rushmore" - they all refer to the same symbol).

For such a translation, a lot of knowledge about the world is required, which is also expressed in the same symbolic language. In this approach, it is natural to see the metaphor as an expression of a mapping between structures: if in the world of projects we treat a certain project as progressing according to plan or as being at risk of delay or failure, and in the world of sailing we recognize the danger of running aground, it is natural to create the analogy between the structure "project - danger - delay" and "sailing - danger - run aground". The element of danger is what unites the two structures, and it allows us to understand that the speaker expresses fear for the success of the project. (Note: the word "advanced" is also related to the metaphor of a project as a journey: as mentioned, it is very difficult to avoid such metaphors when dealing with abstract concepts. This is of course not the only example in this article.)

ATT-Meta project

A project that represents this approach is the ATT-Meta project led by Professor John Barnden from the University of Birmingham in England. The project is dedicated to the field of metaphors dealing with representations of mental states. Among other things, a database was built as part of the project that includes more than a thousand textual examples, and their link to categories of metaphors such as "the mind as a physical space" ("I keep it in my head", "I opened a door into a hidden room in my mind"), "the mind as a living being" ( "Let your thoughts move in all directions", "The mind chooses its path carefully in a field of thorns"), "Thoughts as inner speech" ("I said to myself", "A part of me that I didn't know before raised its head in protest") and more.

As you can see from these examples, the categories include "literary" statements alongside common and banal expressions, and one example may use several metaphors at the same time. Although the intent of the text is easy to understand, the symbolic representation required is complex as a result of the great variety of forms of expression and the different structure each such form can assume. The project uses these symbolic representations to identify such metaphorical references, process them and understand them.

One of the members of the project team found reinforcement for the need to understand metaphors in the challenges of "textual drag" - a text analysis task proposed by Dr. Ado Dagan from Bar-Ilan University and which has received a lot of attention in recent years.

Dagan classified each subtask as requiring metaphorical understanding or as not needing it. It was found that the programs that achieved the highest achievements in these challenges were much weaker for those subtasks in which a metaphor was involved, even though it is possible that most human readers would not even sense the existence of a metaphor in these tasks (for example, "the stock market recovered"). This finding demonstrates the significant weakness of artificial intelligence in this area, even compared to its (so far limited) achievements in other areas of text understanding.
Identifying metaphors as a literary research tool

Projects like ATT-Meta show the potential of the understanding of metaphor developed in linguistics and the study of literature to contribute to advances in artificial intelligence. The project "The mind is a metaphor" tries to harness artificial intelligence for the study of literature. The project was started by Brad Pasanek (now a professor of 18th century literature at the University of Virginia) at Stanford University in the USA. While doing his doctoral work, he kept a handwritten list of metaphors related to the mind and thinking in the Bible and in the writings of Milton and Shakespeare. When the list grew, Pesnak enlisted the help of his friends from the computer science faculty in creating an electronic database into which he entered his findings.

Although the move to a digital repository made storage easier, Pesnak realized that as long as searching and identifying the metaphors required human labor, progress would be limited.

How can you teach a computer to recognize a metaphor in a literary work? Today, searching for the appearance of a certain word is easy and simple. Searching for images is also not particularly difficult - you have to look for sentences like "thought is like" and collect all those expressions that appear there (this is of course an oversimplification, and a more general method is needed, but the idea remains similar). Searching for metaphors is difficult because of the difference between them and images: the metaphor does not explicitly indicate that a comparison is being made, nor does it point to the concepts it compares.

Pesanek told his childhood friend about this problem when they met by chance at a wedding. This friend, who is now a computer scientist, suggested that he use computer learning methods: if we provide the computer with examples of the forms in which a certain metaphor appears, the computer will be able to identify statistical and probabilistic characteristics of the texts provided as an example and search for additional texts that match that example. This method is used, among other things, to filter out "junk mail" by computerized study of letters that have been classified as junk mail and using the learned information to classify new letters.

Was Aristotle wrong?

Today, the database contains about 8,000 texts containing metaphors, about 6,000 of which were identified using a computerized search using this method and other automated text analysis methods. The texts are classified according to the type of metaphor as well as according to the date, gender and profession of the author, the genre of the text and more. For example, one can find that images from the field of physics almost did not appear before the 17th century and became very common in the 18th century, at the beginning of which the philosopher Bishop Berkeley compared the mind to a sphere in motion. Images from the computer world, of course, appeared only recently ("I need to do a reset").

This is not the only example of technological effects on the images of the mind: 300 years ago, when the modern novel began to be created, the images of paper and books appeared, while the first comparison to an engine found in this database appears in a Shakespeare play from 1594.

Pesnak uses the many possibilities of classification and search that this database gives him to follow the changes in the metaphors of the soul throughout the history of the works that he studies. For example, the young soul was compared in the fourth century BC to a "flat board" (tabula rasa). In the 17th century, the philosopher Locke used a similar image from a different writing technology - "white paper, without any written letter". In the 18th century, images of a spinning skewer appear, as if the soul is being cooked in the fire of learning.

Aristotle wrote that metaphors are the superior part of writing, in that the author creates a link between seemingly unrelated things. Prof. Pesanek cites Aristotle's claim that mastery of metaphors is the only thing the poet cannot learn from others. He intends to show that this claim is not true for the identification of metaphors, and to use the computer to learn - while the computer teaches itself.

The writer works at ClickSoftware developing advanced optimization methods.

More on the subject on the science website

• The scientist robot
• Singularity - the end of the human era
• By 2050, robots will be able to play soccer against humans
• How to teach a computer to read body language