Summer Reflections

Over the summer, I have had the opportunity to greatly expand my linguistics knowledge. One area that has piqued my interest is linguistic ambiguity, in particular syntactic and semantic ambiguity.

An example of this is garden path sentences, which refers to sentences that are usually parsed wrongly at first. This mainly occurs when a sentence contains a word or group of words which can be interpreted in more than one way. Let’s take a very common example in psycholinguistic research: “the old man the boat”. Upon hearing the first part of this sentence, someone is most likely to interpret “old” as being in its adjective form, especially when they hear “man” right after. So, the initial parsing someone would be following is: determiner - adjective - noun. Yet, when listeners hear that the word “the” comes after “man”, they must re-analyse the sentence. They then realise that, here, “old” is actually a noun, and “man” is a verb. So, the sentence’s true meaning is: the old (people) operate the boat. This is one of many examples of garden path sentences that can be explored.

Another type of ambiguity comes with the use of prepositions. Take “I saw the woman with the binoculars” as an example. The sentence could mean: “I saw the woman by using the binoculars” or “I saw the woman who was using binoculars”. Without context, it is almost impossible to figure out which interpretation is the correct one. Ambiguity with prepositions extends beyond this, and thinking about this topic made me realise that, while a lot of interpretations may come naturally for us humans, computers would more likely have difficulty. So, I began to research. Let’s take the sentences: “I ate ice cream with chocolate sauce”, and “I ate ice cream with a spoon”. Humans can quite easily conclude that in the former, “with” implies that the ice cream has chocolate sauce on it to be eaten, but in the latter, “with” implies that a spoon was used to eat the ice cream - not that the spoon was being eaten too! Computers do not really have this natural common sense, so the question I had was: would they still be able to properly interpret the meanings and if so, how?

That is when I came across articles discussing the game “Madly Ambiguous”, developed by the linguist Michael White and his colleagues at Ohio State University. Its aim is precisely to explore a computer’s ability to handle ambiguity, and the demo went live in Summer 2017. In the game, users are asked to fill in the blank for the sentence “John ate spaghetti with ___”, and the system must guess the ‘role’ of the words filling in the blank. There are even two modes: basic and advanced. Basic mode uses part-of-speech tagging and lemmatisation to find the main word, before looking it up in the WordNet database. Advanced mode uses word embeddings, where words with similar meanings are clustered together, allowing the system to find the proper interpretation.

Unfortunately, I was not able to test out the game myself because the link did not work. While I was really looking forward to seeing whether I could stump the computer, I still found learning about this game interesting - and the research findings are even more interesting! A year after the game launched, the Madly Ambiguous team found that the computer’s interpreted meaning matched up with the user’s 64% of the time in basic mode, and 70% of the time in advanced mode. The system performed much better than I expected considering the technology is from over 5 years ago!

All in all, we should not overestimate a computer’s ability to handle linguistic ambiguity. However, we should not underestimate it either. While technology is far from perfect, it is making rapid progress in various ways, including when it comes to overcoming the challenges of ambiguous language, as highlighted by Madly Ambiguous and more recently, ChatGPT. I am excited to continue to follow the work of Natural Language Processing researchers and see just how close computers can get to having a “natural instinct for language”.

Credit: Pallavi Padav (Medium)

Works Cited

“Garden-Path Sentence.” Wikipedia, 26 Apr. 2024, en.wikipedia.org/wiki/Garden-path_sentence. Accessed 25 Aug. 2024.

Gokcen, Ajda, et al. Madly Ambiguous: A Game for Learning about Structural Ambiguity and Why It’s Hard for Computers.

“Parsing.” Wikipedia, 19 Aug. 2024, en.wikipedia.org/wiki/Parsing#. Accessed 25 Aug. 2024.

Schwartz, Samantha. “For Computers Discerning Linguistic Ambiguities, the Struggle Is Real.” CIO Dive, 29 June 2018, www.ciodive.com/news/for-computers-discerning-linguistic-ambiguities-the-struggle-is-real/526845/. Accessed 25 Aug. 2024.

Zimmer, Ben. “How Computers Parse the Ambiguity of Everyday Language.” The Atlantic, 27 June 2018, www.theatlantic.com/technology/archive/2018/06/how-computers-parse-the-ambiguity-of-everyday-language/563828/. Accessed 25 Aug. 2024.